From 53c6d2f6e99c87acf0f5cec1089ddd930bc1b20d Mon Sep 17 00:00:00 2001
From: Michael Ryaboy <michael@context.dev>
Date: Thu, 24 Sep 2026 02:23:58 -0700
Subject: [PATCH 1/2] Measure gzip and Zstd text classification

---
 eval/README.md                |  101 ++
 eval/compression-results.json | 2655 +++++++++++++++++++++++++++++++++
 eval/compression.py           |  371 +++++
 3 files changed, 3127 insertions(+)
 create mode 100644 eval/compression-results.json
 create mode 100644 eval/compression.py

diff --git a/eval/README.md b/eval/README.md
index 2009d1f..24d5ea8 100644
--- a/eval/README.md
+++ b/eval/README.md
@@ -119,3 +119,104 @@ profile, not on a 0.004 difference in F1.
 In rough order of value: add cases from a real taxonomy with labels assigned
 before any model output is seen; split into tune and report halves; raise n past
 about thirty so confidence intervals mean something.
+
+## Compression classifiers
+
+`compression.py` tests whether gzip/DEFLATE or Zstd can replace a learned text
+classifier. It compares gzip NCD nearest neighbors, per-class DEFLATE and Zstd
+dictionaries, Zstd dictionary mixtures, and a word/character TF-IDF logistic
+regression control. It does not change the Worker or call a paid model API.
+
+```sh
+git clone https://github.com/fstandhartinger/jevbench.git /tmp/compression-jevbench
+git -C /tmp/compression-jevbench checkout 2fa63fa3226cb369795525ed011800f57dcbd894
+OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 uv run --python 3.12 eval/compression.py \
+  --out eval/data/compression-reproduction --jevbench /tmp/compression-jevbench
+```
+
+Run from the repository root. The output directory must be new. `uv` installs
+the script's pinned dependencies; Hugging Face downloads are cached. Omit
+`--jevbench` to run only the supervised datasets. `--datasets`, `--families`,
+`--seed`, `--train-cap`, `--validation-n`, and `--n` bound additional experiments.
+
+The default uses seed 0 and 500 evaluation rows, matching the sampling protocol
+of [dhruvmehra/jevbench](https://github.com/dhruvmehra/jevbench). Each dataset has
+up to 10,000 training rows and 600 separate validation rows. Training is shuffled
+and deduplicated after excluding exact normalized matches anywhere in the
+held-out split. Classifier settings and probability temperature are chosen only
+on validation rows, then frozen for evaluation. This excludes exact duplicates,
+not semantic paraphrases or SST-2's related review fragments. Test rows remain
+in their original benchmark sample, including any duplicates within that split.
+
+Gzip nearest neighbors deliberately retain at most 1,000 balanced examples; its
+accuracy is a bounded-compute baseline, not an estimate of full-data gzip kNN.
+DEFLATE uses raw zlib preset dictionaries (the same compression algorithm as
+gzip, without gzip framing). Zstd tries trained dictionaries and raw class
+text, with minimum or mean length over random shards. Compression length is
+only a score: softmax temperature is fitted on validation labels. Dictionaries
+store information from labeled examples; this is supervised learning, not a
+zero-shot or memory-free model. Oversized dictionary training failures are
+recorded in the validation sweep, never silently replaced by another method.
+
+Every run produces `summary.json`, per-dataset validation sweeps and selected
+settings, per-item predictions/probabilities, split hashes, dataset fingerprints,
+and software versions. Timing covers serial local normalization, scoring,
+probabilities and tie-breaking; it excludes training, loading, serving and
+network overhead. Empty, Unicode and long inputs are exercised for every
+selected supervised model. The output manifest identifies the script by hash.
+
+The optional Benchmark Heaven diagnostic uses symmetric gzip NCD between the
+state/instructions and each label/criterion. The prediction function accepts
+only those inference fields; expected answers and author rationales cannot
+enter it. This has no labeled support set and returns labels without invented
+calibrated probabilities. Its three public cohorts total 231 decisions. These
+are public diagnostic accuracies, not an official JevBench score: private,
+sealed, and imported tasks are absent. A supervised banking/news classifier
+cannot be submitted as if it solves arbitrary typed decisions.
+
+The compression idea comes from [Nathan Barry's gzip language-model experiment](https://nathan.rs/posts/gzip-lm/)
+and [FTCC's class-dictionary approach](https://github.com/cyrilou242/ftcc).
+The TF-IDF control matters: [Gzip versus bag-of-words for text classification](https://arxiv.org/abs/2307.15002)
+examines whether compression's gains survive comparison with conventional text
+features.
+
+### Measured results (Apple M5 Max)
+
+Generated from [compression-results.json](compression-results.json). All rows
+use 500 held-out items per dataset; percentages are accuracy. Each family selects
+its settings on validation data. These are exploratory measurements on one seed,
+not evidence that small differences are statistically significant.
+
+| Up to 10k training rows | Gzip kNN (≤1k memory) | DEFLATE | Zstd mixture | TF-IDF + logistic |
+|---|---:|---:|---:|---:|
+| agnews | 61.8% | 76.6% | 84.8% | 88.8% |
+| banking77 | 54.8% | 82.0% | 80.6% | 90.8% |
+| sst2 | 54.8% | 65.6% | 71.8% | 79.6% |
+| emotion | 29.8% | 40.6% | 56.2% | 85.2% |
+
+Larger-training follow-up (same evaluation items, separately held-back validation):
+
+| Dataset | Training rows | Zstd mixture | p50 | TF-IDF + logistic | p50 |
+|---|---:|---:|---:|---:|---:|
+| agnews | 119,239 | 90.0% | 0.246 ms | 91.2% | 0.593 ms |
+| sst2 | 66,378 | 70.0% | 0.034 ms | 85.0% | 0.432 ms |
+
+The larger news mixture retains 28.25 MB of class text across 20 dictionaries;
+this is not a tiny parameter-free model. It is competitive here, but the lexical
+control is more accurate on every measured dataset. More data does not resolve
+the sentiment weakness.
+
+Reproduce the larger run by adding `--datasets agnews sst2 --families zstd
+zstd-mixture tfidf-logistic --train-cap 120000` and choosing a new output directory.
+
+On Benchmark Heaven’s public cohorts, zero-shot gzip scores **45.8% easy**,
+**36.1% original**, and **42.3% hard**. All 231 labels were checked against the
+upstream scorer and were invariant to reversing the option order. No official
+score or leaderboard submission is claimed. The audit fixed ordinal gold labels
+being compared as integers against strings, and removed class-order bias when
+gzip neighbor distances tie.
+
+The checked-in JSON combines the two generated `summary.json` files and the
+audit record. Raw item predictions stay under `eval/data/`. Across the default
+and larger runs, 13,000 probability vectors and accuracies were independently
+checked; the default run’s 8,000 dictionary/lexical predictions reproduced exactly.
diff --git a/eval/compression-results.json b/eval/compression-results.json
new file mode 100644
index 0000000..3f9563e
--- /dev/null
+++ b/eval/compression-results.json
@@ -0,0 +1,2655 @@
+{
+  "default_run": {
+    "manifest": {
+      "args": {
+        "datasets": [
+          "agnews",
+          "banking77",
+          "sst2",
+          "emotion"
+        ],
+        "families": null,
+        "seed": 0,
+        "n": 500,
+        "train_cap": 10000,
+        "validation_n": 600,
+        "out": "eval/data/compression-verified",
+        "jevbench": "/tmp/classifier-gzip-jevbench"
+      },
+      "platform": "macOS-26.5.1-arm64-arm-64bit",
+      "processor": "arm",
+      "python": "3.12.14",
+      "zlib": "1.2.12",
+      "thread_environment": {
+        "OMP_NUM_THREADS": "1",
+        "OPENBLAS_NUM_THREADS": "1"
+      },
+      "timing": "Serial local warm inference: normalization, compression/features, probabilities, tie-breaking. Excludes training, network and data loading.",
+      "versions": {
+        "datasets": "5.0.1",
+        "numpy": "2.5.3",
+        "scipy": "1.18.1",
+        "scikit-learn": "1.9.1",
+        "zstandard": "0.25.0"
+      },
+      "script_sha256": "5342902063b3071509c854905aa18c2410c8202e74c1fcad2ab76ae89491a2f4"
+    },
+    "datasets": {
+      "agnews": {
+        "dataset": {
+          "repo": "fancyzhx/ag_news",
+          "hf_fingerprints": {
+            "train": "2bc1e2be15d7f2ee",
+            "test": "0ebcc57b6aea2fc9"
+          },
+          "train_n": 10000,
+          "validation_n": 600,
+          "test_n": 500,
+          "removed_duplicates_or_eval_overlaps": 161,
+          "hashes": {
+            "train": "04944cfb62e7665ceda0faf53c118d18743495868c573e315806570d7fa2c616",
+            "validation": "b8edd80e646b60c5ca1a76e2744cd41f336090666c93886b305374e69ad44ef6",
+            "test": "76c1f71538ffa2609b9dbb9bbf842e60d19b79470d1f4d0cc554d88d95ae47ed"
+          },
+          "labels": [
+            "World",
+            "Sports",
+            "Business",
+            "Sci/Tech"
+          ]
+        },
+        "validation_sweep": [
+          {
+            "family": "deflate",
+            "config": {
+              "codec": "deflate",
+              "size": 4096
+            },
+            "validation_accuracy": 0.5933333333333334,
+            "fit_s": 0.004283166999812238,
+            "dictionary_or_sample_bytes": 16384
+          },
+          {
+            "family": "deflate",
+            "config": {
+              "codec": "deflate",
+              "size": 32768
+            },
+            "validation_accuracy": 0.785,
+            "fit_s": 0.004138415999477729,
+            "dictionary_or_sample_bytes": 131072
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "trained": true
+            },
+            "validation_accuracy": 0.62,
+            "fit_s": 0.13306408299831674,
+            "dictionary_or_sample_bytes": 4096
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "trained": false
+            },
+            "validation_accuracy": 0.395,
+            "fit_s": 0.003096458996878937,
+            "dictionary_or_sample_bytes": 4096
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "trained": true
+            },
+            "validation_accuracy": 0.7116666666666667,
+            "fit_s": 0.148906624999654,
+            "dictionary_or_sample_bytes": 16384
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "trained": false
+            },
+            "validation_accuracy": 0.5316666666666666,
+            "fit_s": 0.0029460840014507994,
+            "dictionary_or_sample_bytes": 16384
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16384,
+              "trained": true
+            },
+            "validation_accuracy": 0.7816666666666666,
+            "fit_s": 0.12514475000352832,
+            "dictionary_or_sample_bytes": 65536
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16384,
+              "trained": false
+            },
+            "validation_accuracy": 0.6383333333333333,
+            "fit_s": 0.0035843750010826625,
+            "dictionary_or_sample_bytes": 65536
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 65536,
+              "level": 9
+            },
+            "validation_accuracy": 0.755,
+            "fit_s": 0.14723604200116824,
+            "dictionary_or_sample_bytes": 262144
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 262144,
+              "level": 9
+            },
+            "validation_accuracy": 0.7933333333333333,
+            "fit_s": 0.2483245409966912,
+            "dictionary_or_sample_bytes": 1048576
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "level": 9,
+              "trained": false
+            },
+            "validation_accuracy": 0.8466666666666667,
+            "fit_s": 0.003961707996495534,
+            "dictionary_or_sample_bytes": 2366771
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "shards": 4
+            },
+            "validation_accuracy": 0.7016666666666667,
+            "fit_s": 0.11391575000016019,
+            "dictionary_or_sample_bytes": 16384
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "shards": 8
+            },
+            "validation_accuracy": 0.7266666666666667,
+            "fit_s": 0.12054720900050597,
+            "dictionary_or_sample_bytes": 32768
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "shards": 4
+            },
+            "validation_accuracy": 0.7783333333333333,
+            "fit_s": 0.17654266599856783,
+            "dictionary_or_sample_bytes": 65536
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "shards": 8
+            },
+            "validation_accuracy": 0.7983333333333333,
+            "fit_s": 0.174204084003577,
+            "dictionary_or_sample_bytes": 131072
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "min"
+            },
+            "validation_accuracy": 0.8633333333333333,
+            "fit_s": 0.0033049999983632006,
+            "dictionary_or_sample_bytes": 2366763
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.895,
+            "fit_s": 0.0030992079991847277,
+            "dictionary_or_sample_bytes": 2366763
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "min"
+            },
+            "validation_accuracy": 0.8766666666666667,
+            "fit_s": 0.003387500000826549,
+            "dictionary_or_sample_bytes": 2366755
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.8833333333333333,
+            "fit_s": 0.003104374998656567,
+            "dictionary_or_sample_bytes": 2366755
+          },
+          {
+            "family": "gzip-knn",
+            "config": {
+              "k": 1
+            },
+            "validation_accuracy": 0.6116666666666667,
+            "fit_s": 0.010663459004717879,
+            "dictionary_or_sample_bytes": 232003
+          },
+          {
+            "family": "gzip-knn",
+            "config": {
+              "k": 5
+            },
+            "validation_accuracy": 0.6383333333333333,
+            "fit_s": 0.012955791004060302,
+            "dictionary_or_sample_bytes": 232003
+          },
+          {
+            "family": "tfidf-logistic",
+            "config": {},
+            "validation_accuracy": 0.9033333333333333,
+            "fit_s": 5.491404250002233,
+            "dictionary_or_sample_bytes": null
+          }
+        ],
+        "selected": {
+          "deflate": {
+            "family": "deflate",
+            "config": {
+              "codec": "deflate",
+              "size": 32768
+            },
+            "validation_accuracy": 0.785,
+            "fit_s": 0.004138415999477729,
+            "dictionary_or_sample_bytes": 131072,
+            "temperature": 4.29444227551508,
+            "accuracy": 0.766,
+            "macro_f1": 0.771227501914879,
+            "ece": 0.06008734299381312,
+            "nll": 0.6778785936410218,
+            "brier": 0.349022246408619,
+            "tie_rate": 0.03,
+            "p50_ms": 0.0734795,
+            "p95_ms": 0.09922049999999999,
+            "serial_items_per_s": 13096.827675922877,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.7269477983375824,
+              0.8009959046039771
+            ],
+            "confidence_at_least_90pct": {
+              "n": 180,
+              "coverage": 0.36,
+              "accuracy": 0.9222222222222223
+            }
+          },
+          "zstd": {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "level": 9,
+              "trained": false
+            },
+            "validation_accuracy": 0.8466666666666667,
+            "fit_s": 0.003961707996495534,
+            "dictionary_or_sample_bytes": 2366771,
+            "temperature": 4.637143191628681,
+            "accuracy": 0.808,
+            "macro_f1": 0.8072402391715388,
+            "ece": 0.057153539576628706,
+            "nll": 0.6757015422951654,
+            "brier": 0.2950965013185466,
+            "tie_rate": 0.022,
+            "p50_ms": 0.044521,
+            "p95_ms": 0.0784274,
+            "serial_items_per_s": 20026.02101066041,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.7711789076945584,
+              0.8401243272904052
+            ],
+            "confidence_at_least_90pct": {
+              "n": 263,
+              "coverage": 0.526,
+              "accuracy": 0.9125475285171103
+            }
+          },
+          "zstd-mixture": {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.895,
+            "fit_s": 0.0030992079991847277,
+            "dictionary_or_sample_bytes": 2366763,
+            "temperature": 3.410951603860984,
+            "accuracy": 0.848,
+            "macro_f1": 0.8480560127251228,
+            "ece": 0.04504076255490379,
+            "nll": 0.5399901438401714,
+            "brier": 0.22715867018883865,
+            "tie_rate": 0.006,
+            "p50_ms": 0.12854149999999998,
+            "p95_ms": 0.19065174999999993,
+            "serial_items_per_s": 7451.103215557868,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.8138851771600609,
+              0.8768080883424302
+            ],
+            "confidence_at_least_90pct": {
+              "n": 302,
+              "coverage": 0.604,
+              "accuracy": 0.9370860927152318
+            }
+          },
+          "gzip-knn": {
+            "family": "gzip-knn",
+            "config": {
+              "k": 5
+            },
+            "validation_accuracy": 0.6383333333333333,
+            "fit_s": 0.012955791004060302,
+            "dictionary_or_sample_bytes": 232003,
+            "temperature": 2.1518556436402267,
+            "accuracy": 0.618,
+            "macro_f1": 0.6199502440461163,
+            "ece": 0.10602590208369551,
+            "nll": 0.9974734324691097,
+            "brier": 0.5320566795008369,
+            "tie_rate": 0.2,
+            "p50_ms": 9.704708,
+            "p95_ms": 11.13455415,
+            "serial_items_per_s": 102.52232695558666,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.5746644739023564,
+              0.6595361161243504
+            ],
+            "confidence_at_least_90pct": {
+              "n": 0,
+              "coverage": 0.0,
+              "accuracy": null
+            }
+          },
+          "tfidf-logistic": {
+            "family": "tfidf-logistic",
+            "config": {},
+            "validation_accuracy": 0.9033333333333333,
+            "fit_s": 5.491404250002233,
+            "dictionary_or_sample_bytes": null,
+            "temperature": 0.7345152395906567,
+            "accuracy": 0.888,
+            "macro_f1": 0.8879638858343402,
+            "ece": 0.06050312885515615,
+            "nll": 0.44172785940001424,
+            "brier": 0.19518180715294506,
+            "tie_rate": 0.0,
+            "p50_ms": 0.5366455,
+            "p95_ms": 0.70514435,
+            "serial_items_per_s": 1814.9127050219765,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.8573456933034787,
+              0.9127376027165404
+            ],
+            "confidence_at_least_90pct": {
+              "n": 368,
+              "coverage": 0.736,
+              "accuracy": 0.9266304347826086
+            }
+          }
+        }
+      },
+      "banking77": {
+        "dataset": {
+          "repo": "legacy-datasets/banking77",
+          "hf_fingerprints": {
+            "train": "8360b37e52e4eb2a",
+            "test": "1d7c14b4c23350b9"
+          },
+          "train_n": 9392,
+          "validation_n": 600,
+          "test_n": 500,
+          "removed_duplicates_or_eval_overlaps": 11,
+          "hashes": {
+            "train": "3d092fadae6b94818c39469faa6c8997e81f07219855d5dd0a510fb78fd79fc6",
+            "validation": "4e12f6da49e8667d0a3464d75bd7c82ec2338a890264835660468ccad307fddd",
+            "test": "709b7b5509738c4bce13a37aa816b4934b835be78e91cf0c79792ba5b1716dae"
+          },
+          "labels": [
+            "activate_my_card",
+            "age_limit",
+            "apple_pay_or_google_pay",
+            "atm_support",
+            "automatic_top_up",
+            "balance_not_updated_after_bank_transfer",
+            "balance_not_updated_after_cheque_or_cash_deposit",
+            "beneficiary_not_allowed",
+            "cancel_transfer",
+            "card_about_to_expire",
+            "card_acceptance",
+            "card_arrival",
+            "card_delivery_estimate",
+            "card_linking",
+            "card_not_working",
+            "card_payment_fee_charged",
+            "card_payment_not_recognised",
+            "card_payment_wrong_exchange_rate",
+            "card_swallowed",
+            "cash_withdrawal_charge",
+            "cash_withdrawal_not_recognised",
+            "change_pin",
+            "compromised_card",
+            "contactless_not_working",
+            "country_support",
+            "declined_card_payment",
+            "declined_cash_withdrawal",
+            "declined_transfer",
+            "direct_debit_payment_not_recognised",
+            "disposable_card_limits",
+            "edit_personal_details",
+            "exchange_charge",
+            "exchange_rate",
+            "exchange_via_app",
+            "extra_charge_on_statement",
+            "failed_transfer",
+            "fiat_currency_support",
+            "get_disposable_virtual_card",
+            "get_physical_card",
+            "getting_spare_card",
+            "getting_virtual_card",
+            "lost_or_stolen_card",
+            "lost_or_stolen_phone",
+            "order_physical_card",
+            "passcode_forgotten",
+            "pending_card_payment",
+            "pending_cash_withdrawal",
+            "pending_top_up",
+            "pending_transfer",
+            "pin_blocked",
+            "receiving_money",
+            "Refund_not_showing_up",
+            "request_refund",
+            "reverted_card_payment?",
+            "supported_cards_and_currencies",
+            "terminate_account",
+            "top_up_by_bank_transfer_charge",
+            "top_up_by_card_charge",
+            "top_up_by_cash_or_cheque",
+            "top_up_failed",
+            "top_up_limits",
+            "top_up_reverted",
+            "topping_up_by_card",
+            "transaction_charged_twice",
+            "transfer_fee_charged",
+            "transfer_into_account",
+            "transfer_not_received_by_recipient",
+            "transfer_timing",
+            "unable_to_verify_identity",
+            "verify_my_identity",
+            "verify_source_of_funds",
+            "verify_top_up",
+            "virtual_card_not_working",
+            "visa_or_mastercard",
+            "why_verify_identity",
+            "wrong_amount_of_cash_received",
+            "wrong_exchange_rate_for_cash_withdrawal"
+          ]
+        },
+        "validation_sweep": [
+          {
+            "family": "deflate",
+            "config": {
+              "codec": "deflate",
+              "size": 4096
+            },
+            "validation_accuracy": 0.805,
+            "fit_s": 0.0074147080013062805,
+            "dictionary_or_sample_bytes": 304038
+          },
+          {
+            "family": "deflate",
+            "config": {
+              "codec": "deflate",
+              "size": 32768
+            },
+            "validation_accuracy": 0.7983333333333333,
+            "fit_s": 0.008205541002098471,
+            "dictionary_or_sample_bytes": 567780
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "trained": true
+            },
+            "validation_accuracy": 0.6466666666666666,
+            "fit_s": 0.05034095799783245,
+            "dictionary_or_sample_bytes": 78848
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "trained": false
+            },
+            "validation_accuracy": 0.6416666666666667,
+            "fit_s": 0.007061042000714224,
+            "dictionary_or_sample_bytes": 78848
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "trained": true
+            },
+            "validation_accuracy": 0.6933333333333334,
+            "fit_s": 0.08712499999819556,
+            "dictionary_or_sample_bytes": 286121
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "trained": false
+            },
+            "validation_accuracy": 0.7333333333333333,
+            "fit_s": 0.007319832999201026,
+            "dictionary_or_sample_bytes": 304038
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16384,
+              "trained": true
+            },
+            "validation_accuracy": 0.7133333333333334,
+            "fit_s": 0.12062429200159386,
+            "dictionary_or_sample_bytes": 411372
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16384,
+              "trained": false
+            },
+            "validation_accuracy": 0.7083333333333334,
+            "fit_s": 0.007289250002941117,
+            "dictionary_or_sample_bytes": 567780
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 65536,
+              "level": 9
+            },
+            "validation_accuracy": 0.7483333333333333,
+            "fit_s": 0.11870433299918659,
+            "dictionary_or_sample_bytes": 411372
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 262144,
+              "level": 9
+            },
+            "validation_accuracy": 0.7483333333333333,
+            "fit_s": 0.12351499999931548,
+            "dictionary_or_sample_bytes": 411372
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "level": 9,
+              "trained": false
+            },
+            "validation_accuracy": 0.7816666666666666,
+            "fit_s": 0.007592082998598926,
+            "dictionary_or_sample_bytes": 567780
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "shards": 4
+            },
+            "validation_accuracy": 0.72,
+            "fit_s": 0.12457012500090059,
+            "dictionary_or_sample_bytes": 281325
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "shards": 8
+            },
+            "training_error": "cannot train dict: Src size is incorrect"
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "shards": 4
+            },
+            "validation_accuracy": 0.7366666666666667,
+            "fit_s": 0.20517695900343824,
+            "dictionary_or_sample_bytes": 454504
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "shards": 8
+            },
+            "training_error": "cannot train dict: Src size is incorrect"
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "min"
+            },
+            "validation_accuracy": 0.8166666666666667,
+            "fit_s": 0.007201125001301989,
+            "dictionary_or_sample_bytes": 567626
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.835,
+            "fit_s": 0.007468750001862645,
+            "dictionary_or_sample_bytes": 567626
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "min"
+            },
+            "validation_accuracy": 0.8183333333333334,
+            "fit_s": 0.0075589999978546984,
+            "dictionary_or_sample_bytes": 567472
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.825,
+            "fit_s": 0.00782808299845783,
+            "dictionary_or_sample_bytes": 567472
+          },
+          {
+            "family": "gzip-knn",
+            "config": {
+              "k": 1
+            },
+            "validation_accuracy": 0.49833333333333335,
+            "fit_s": 0.011530167001183145,
+            "dictionary_or_sample_bytes": 52739
+          },
+          {
+            "family": "gzip-knn",
+            "config": {
+              "k": 5
+            },
+            "validation_accuracy": 0.45166666666666666,
+            "fit_s": 0.011287084002105985,
+            "dictionary_or_sample_bytes": 52739
+          },
+          {
+            "family": "tfidf-logistic",
+            "config": {},
+            "validation_accuracy": 0.9016666666666666,
+            "fit_s": 6.198137958999723,
+            "dictionary_or_sample_bytes": null
+          }
+        ],
+        "selected": {
+          "deflate": {
+            "family": "deflate",
+            "config": {
+              "codec": "deflate",
+              "size": 4096
+            },
+            "validation_accuracy": 0.805,
+            "fit_s": 0.0074147080013062805,
+            "dictionary_or_sample_bytes": 304038,
+            "temperature": 1.35753398137484,
+            "accuracy": 0.82,
+            "macro_f1": 0.7992486209415565,
+            "ece": 0.06290305365662752,
+            "nll": 0.748491242227106,
+            "brier": 0.28054942628044477,
+            "tie_rate": 0.094,
+            "p50_ms": 0.5353749999999999,
+            "p95_ms": 0.71514165,
+            "serial_items_per_s": 1798.0028453251189,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.7839246244600844,
+              0.8511956196801373
+            ],
+            "confidence_at_least_90pct": {
+              "n": 244,
+              "coverage": 0.488,
+              "accuracy": 0.9590163934426229
+            }
+          },
+          "zstd": {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "level": 9,
+              "trained": false
+            },
+            "validation_accuracy": 0.7816666666666666,
+            "fit_s": 0.007592082998598926,
+            "dictionary_or_sample_bytes": 567780,
+            "temperature": 1.5828442546700814,
+            "accuracy": 0.778,
+            "macro_f1": 0.7634533737731454,
+            "ece": 0.03987457002093967,
+            "nll": 0.8437418486929151,
+            "brier": 0.31676958909518704,
+            "tie_rate": 0.1,
+            "p50_ms": 0.14937499999999998,
+            "p95_ms": 0.3112707999999996,
+            "serial_items_per_s": 5918.135240090225,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.7395294756647782,
+              0.8122312364320395
+            ],
+            "confidence_at_least_90pct": {
+              "n": 234,
+              "coverage": 0.468,
+              "accuracy": 0.9700854700854701
+            }
+          },
+          "zstd-mixture": {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.835,
+            "fit_s": 0.007468750001862645,
+            "dictionary_or_sample_bytes": 567626,
+            "temperature": 1.35753398137484,
+            "accuracy": 0.806,
+            "macro_f1": 0.7948050126018442,
+            "ece": 0.027330381293454824,
+            "nll": 0.7245106932888512,
+            "brier": 0.27127663578583444,
+            "tie_rate": 0.026,
+            "p50_ms": 0.5831459999999999,
+            "p95_ms": 1.1346207999999984,
+            "serial_items_per_s": 1549.8975985756417,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.7690596512561212,
+              0.838274082202966
+            ],
+            "confidence_at_least_90pct": {
+              "n": 273,
+              "coverage": 0.546,
+              "accuracy": 0.9743589743589743
+            }
+          },
+          "gzip-knn": {
+            "family": "gzip-knn",
+            "config": {
+              "k": 1
+            },
+            "validation_accuracy": 0.49833333333333335,
+            "fit_s": 0.011530167001183145,
+            "dictionary_or_sample_bytes": 52739,
+            "temperature": 1.0782500762595328,
+            "accuracy": 0.548,
+            "macro_f1": 0.5242677003523842,
+            "ece": 0.06063191239609045,
+            "nll": 2.6533837260533817,
+            "brier": 0.7007323897060809,
+            "tie_rate": 0.0,
+            "p50_ms": 4.4909795,
+            "p95_ms": 5.387016699999999,
+            "serial_items_per_s": 217.7996094879138,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.5041745952330335,
+              0.5910934413879999
+            ],
+            "confidence_at_least_90pct": {
+              "n": 0,
+              "coverage": 0.0,
+              "accuracy": null
+            }
+          },
+          "tfidf-logistic": {
+            "family": "tfidf-logistic",
+            "config": {},
+            "validation_accuracy": 0.9016666666666666,
+            "fit_s": 6.198137958999723,
+            "dictionary_or_sample_bytes": null,
+            "temperature": 0.5402886714960652,
+            "accuracy": 0.908,
+            "macro_f1": 0.8941185048731366,
+            "ece": 0.02614868859476964,
+            "nll": 0.3168123349805611,
+            "brier": 0.1376886364442629,
+            "tie_rate": 0.0,
+            "p50_ms": 0.383479,
+            "p95_ms": 0.52263335,
+            "serial_items_per_s": 2489.7499360395686,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.8794606777802376,
+              0.9303176334985452
+            ],
+            "confidence_at_least_90pct": {
+              "n": 392,
+              "coverage": 0.784,
+              "accuracy": 0.9897959183673469
+            }
+          }
+        }
+      },
+      "sst2": {
+        "dataset": {
+          "repo": "stanfordnlp/sst2",
+          "hf_fingerprints": {
+            "train": "e7e6610fdbb3c277",
+            "validation": "c1ddc6497ec97f98",
+            "test": "14a60218fd61ec38"
+          },
+          "train_n": 10000,
+          "validation_n": 600,
+          "test_n": 500,
+          "removed_duplicates_or_eval_overlaps": 371,
+          "hashes": {
+            "train": "bb18ecae29bb9ea84b25405f6a0428096951a9a69efac09bb100c9facea1869d",
+            "validation": "fafee5f02c9c933efd0edf3c5c0b1621b8522669dfaca255450c809c6a815590",
+            "test": "380747f6d5a4f05ce8516c891e34d9dd29500beace1c0937b524634411bb287e"
+          },
+          "labels": [
+            "negative",
+            "positive"
+          ]
+        },
+        "validation_sweep": [
+          {
+            "family": "deflate",
+            "config": {
+              "codec": "deflate",
+              "size": 4096
+            },
+            "validation_accuracy": 0.5966666666666667,
+            "fit_s": 0.0017086660009226762,
+            "dictionary_or_sample_bytes": 8192
+          },
+          {
+            "family": "deflate",
+            "config": {
+              "codec": "deflate",
+              "size": 32768
+            },
+            "validation_accuracy": 0.6766666666666666,
+            "fit_s": 0.0017910410024342127,
+            "dictionary_or_sample_bytes": 65536
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "trained": true
+            },
+            "validation_accuracy": 0.5166666666666667,
+            "fit_s": 0.027700541002559476,
+            "dictionary_or_sample_bytes": 2048
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "trained": false
+            },
+            "validation_accuracy": 0.5066666666666667,
+            "fit_s": 0.0017736669979058206,
+            "dictionary_or_sample_bytes": 2048
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "trained": true
+            },
+            "validation_accuracy": 0.5833333333333334,
+            "fit_s": 0.04100783300236799,
+            "dictionary_or_sample_bytes": 8192
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "trained": false
+            },
+            "validation_accuracy": 0.5483333333333333,
+            "fit_s": 0.002048333000857383,
+            "dictionary_or_sample_bytes": 8192
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16384,
+              "trained": true
+            },
+            "validation_accuracy": 0.6283333333333333,
+            "fit_s": 0.038298833002045285,
+            "dictionary_or_sample_bytes": 32768
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16384,
+              "trained": false
+            },
+            "validation_accuracy": 0.5933333333333334,
+            "fit_s": 0.0018218339973827824,
+            "dictionary_or_sample_bytes": 32768
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 65536,
+              "level": 9
+            },
+            "validation_accuracy": 0.6433333333333333,
+            "fit_s": 0.046571332997700665,
+            "dictionary_or_sample_bytes": 131072
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 262144,
+              "level": 9
+            },
+            "validation_accuracy": 0.7066666666666667,
+            "fit_s": 0.08784708299936028,
+            "dictionary_or_sample_bytes": 437677
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "level": 9,
+              "trained": false
+            },
+            "validation_accuracy": 0.7283333333333334,
+            "fit_s": 0.0018446670001139864,
+            "dictionary_or_sample_bytes": 536488
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "shards": 4
+            },
+            "validation_accuracy": 0.5583333333333333,
+            "fit_s": 0.029137124998669606,
+            "dictionary_or_sample_bytes": 8192
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "shards": 8
+            },
+            "validation_accuracy": 0.58,
+            "fit_s": 0.031216542003676295,
+            "dictionary_or_sample_bytes": 16384
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "shards": 4
+            },
+            "validation_accuracy": 0.6216666666666667,
+            "fit_s": 0.0438428749985178,
+            "dictionary_or_sample_bytes": 32768
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "shards": 8
+            },
+            "validation_accuracy": 0.6416666666666667,
+            "fit_s": 0.04907000000093831,
+            "dictionary_or_sample_bytes": 65536
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "min"
+            },
+            "validation_accuracy": 0.725,
+            "fit_s": 0.00196666600095341,
+            "dictionary_or_sample_bytes": 536484
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.7533333333333333,
+            "fit_s": 0.001909541002532933,
+            "dictionary_or_sample_bytes": 536484
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "min"
+            },
+            "validation_accuracy": 0.7366666666666667,
+            "fit_s": 0.0017700829994282685,
+            "dictionary_or_sample_bytes": 536480
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.7616666666666667,
+            "fit_s": 0.001830375003919471,
+            "dictionary_or_sample_bytes": 536480
+          },
+          {
+            "family": "gzip-knn",
+            "config": {
+              "k": 1
+            },
+            "validation_accuracy": 0.5733333333333334,
+            "fit_s": 0.005904542005737312,
+            "dictionary_or_sample_bytes": 52471
+          },
+          {
+            "family": "gzip-knn",
+            "config": {
+              "k": 5
+            },
+            "validation_accuracy": 0.595,
+            "fit_s": 0.006380707993230317,
+            "dictionary_or_sample_bytes": 52471
+          },
+          {
+            "family": "tfidf-logistic",
+            "config": {},
+            "validation_accuracy": 0.84,
+            "fit_s": 0.5434237500012387,
+            "dictionary_or_sample_bytes": null
+          }
+        ],
+        "selected": {
+          "deflate": {
+            "family": "deflate",
+            "config": {
+              "codec": "deflate",
+              "size": 32768
+            },
+            "validation_accuracy": 0.6766666666666666,
+            "fit_s": 0.0017910410024342127,
+            "dictionary_or_sample_bytes": 65536,
+            "temperature": 6.304134293587772,
+            "accuracy": 0.656,
+            "macro_f1": 0.655332724153962,
+            "ece": 0.02910858412013078,
+            "nll": 0.6184185625087876,
+            "brier": 0.43000688230669276,
+            "tie_rate": 0.074,
+            "p50_ms": 0.0276875,
+            "p95_ms": 0.036502099999999996,
+            "serial_items_per_s": 35514.841972093716,
+            "n": 500,
+            "auroc": 0.7180215804303278,
+            "accuracy_wilson_95": [
+              0.6133133706304251,
+              0.6963077483879331
+            ],
+            "confidence_at_least_90pct": {
+              "n": 11,
+              "coverage": 0.022,
+              "accuracy": 0.9090909090909091
+            }
+          },
+          "zstd": {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "level": 9,
+              "trained": false
+            },
+            "validation_accuracy": 0.7283333333333334,
+            "fit_s": 0.0018446670001139864,
+            "dictionary_or_sample_bytes": 536488,
+            "temperature": 10.790252743818396,
+            "accuracy": 0.65,
+            "macro_f1": 0.6499985999944,
+            "ece": 0.03593030036559715,
+            "nll": 0.618717418902004,
+            "brier": 0.43029277027337204,
+            "tie_rate": 0.086,
+            "p50_ms": 0.013625,
+            "p95_ms": 0.022675299999999992,
+            "serial_items_per_s": 69523.71113120501,
+            "n": 500,
+            "auroc": 0.7180696080942622,
+            "accuracy_wilson_95": [
+              0.6071920689703412,
+              0.6905205454703878
+            ],
+            "confidence_at_least_90pct": {
+              "n": 2,
+              "coverage": 0.004,
+              "accuracy": 1.0
+            }
+          },
+          "zstd-mixture": {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.7616666666666667,
+            "fit_s": 0.001830375003919471,
+            "dictionary_or_sample_bytes": 536480,
+            "temperature": 2.925419034377105,
+            "accuracy": 0.718,
+            "macro_f1": 0.7175922031413361,
+            "ece": 0.04554804951123206,
+            "nll": 0.5547712132787098,
+            "brier": 0.37369763190577826,
+            "tie_rate": 0.028,
+            "p50_ms": 0.041292,
+            "p95_ms": 0.06892435,
+            "serial_items_per_s": 22944.298676765586,
+            "n": 500,
+            "auroc": 0.7985319544057375,
+            "accuracy_wilson_95": [
+              0.6770114417638189,
+              0.7556642245567071
+            ],
+            "confidence_at_least_90pct": {
+              "n": 72,
+              "coverage": 0.144,
+              "accuracy": 0.9166666666666666
+            }
+          },
+          "gzip-knn": {
+            "family": "gzip-knn",
+            "config": {
+              "k": 5
+            },
+            "validation_accuracy": 0.595,
+            "fit_s": 0.006380707993230317,
+            "dictionary_or_sample_bytes": 52471,
+            "temperature": 5.007191993770399,
+            "accuracy": 0.548,
+            "macro_f1": 0.5474134478283856,
+            "ece": 0.013168762839443831,
+            "nll": 0.6913485232561188,
+            "brier": 0.4975115063432846,
+            "tie_rate": 0.0,
+            "p50_ms": 5.481375,
+            "p95_ms": 6.29362325,
+            "serial_items_per_s": 182.7036415967676,
+            "n": 500,
+            "auroc": 0.5551037397540983,
+            "accuracy_wilson_95": [
+              0.5041745952330335,
+              0.5910934413879999
+            ],
+            "confidence_at_least_90pct": {
+              "n": 0,
+              "coverage": 0.0,
+              "accuracy": null
+            }
+          },
+          "tfidf-logistic": {
+            "family": "tfidf-logistic",
+            "config": {},
+            "validation_accuracy": 0.84,
+            "fit_s": 0.5434237500012387,
+            "dictionary_or_sample_bytes": null,
+            "temperature": 0.7345152395906567,
+            "accuracy": 0.796,
+            "macro_f1": 0.7959706197692469,
+            "ece": 0.03953913908774128,
+            "nll": 0.41708729362573255,
+            "brier": 0.26258361413544473,
+            "tie_rate": 0.0,
+            "p50_ms": 0.412729,
+            "p95_ms": 0.5067076999999999,
+            "serial_items_per_s": 2391.932066718238,
+            "n": 500,
+            "auroc": 0.8961321721311475,
+            "accuracy_wilson_95": [
+              0.7584839354593685,
+              0.8290022903703368
+            ],
+            "confidence_at_least_90pct": {
+              "n": 243,
+              "coverage": 0.486,
+              "accuracy": 0.9382716049382716
+            }
+          }
+        }
+      },
+      "emotion": {
+        "dataset": {
+          "repo": "dair-ai/emotion",
+          "hf_fingerprints": {
+            "train": "27c5e4f1e47ffa42",
+            "validation": "5cae643edfd5aaf3",
+            "test": "bd45cf0a5e0b9690"
+          },
+          "train_n": 10000,
+          "validation_n": 600,
+          "test_n": 500,
+          "removed_duplicates_or_eval_overlaps": 42,
+          "hashes": {
+            "train": "5bcc8e86df660889cb24bfd3a261f5528d8fb3b142f4693c73fd739338b9cb86",
+            "validation": "bda5d620b7ea488ba509f07b3591f772feb319b8a396d7812b26f89421143e7a",
+            "test": "62fe848e5c2808aed7e0077efd032834d274b6eb6e7d323e76be32ef97b1feea"
+          },
+          "labels": [
+            "sadness",
+            "joy",
+            "love",
+            "anger",
+            "fear",
+            "surprise"
+          ]
+        },
+        "validation_sweep": [
+          {
+            "family": "deflate",
+            "config": {
+              "codec": "deflate",
+              "size": 4096
+            },
+            "validation_accuracy": 0.32666666666666666,
+            "fit_s": 0.002004999994824175,
+            "dictionary_or_sample_bytes": 24576
+          },
+          {
+            "family": "deflate",
+            "config": {
+              "codec": "deflate",
+              "size": 32768
+            },
+            "validation_accuracy": 0.43833333333333335,
+            "fit_s": 0.00225687499914784,
+            "dictionary_or_sample_bytes": 196608
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "trained": true
+            },
+            "validation_accuracy": 0.33,
+            "fit_s": 0.04342029100371292,
+            "dictionary_or_sample_bytes": 6144
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "trained": false
+            },
+            "validation_accuracy": 0.235,
+            "fit_s": 0.0020858330026385374,
+            "dictionary_or_sample_bytes": 6144
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "trained": true
+            },
+            "validation_accuracy": 0.445,
+            "fit_s": 0.05890566699963529,
+            "dictionary_or_sample_bytes": 24576
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "trained": false
+            },
+            "validation_accuracy": 0.275,
+            "fit_s": 0.002086583997879643,
+            "dictionary_or_sample_bytes": 24576
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16384,
+              "trained": true
+            },
+            "validation_accuracy": 0.44666666666666666,
+            "fit_s": 0.058133709004323464,
+            "dictionary_or_sample_bytes": 98304
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16384,
+              "trained": false
+            },
+            "validation_accuracy": 0.37666666666666665,
+            "fit_s": 0.0023508340018452145,
+            "dictionary_or_sample_bytes": 98304
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 65536,
+              "level": 9
+            },
+            "validation_accuracy": 0.4033333333333333,
+            "fit_s": 0.11329087500052992,
+            "dictionary_or_sample_bytes": 349551
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 262144,
+              "level": 9
+            },
+            "validation_accuracy": 0.43666666666666665,
+            "fit_s": 0.15255850000539795,
+            "dictionary_or_sample_bytes": 856643
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "level": 9,
+              "trained": false
+            },
+            "validation_accuracy": 0.44166666666666665,
+            "fit_s": 0.0022252499984460883,
+            "dictionary_or_sample_bytes": 979245
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "shards": 4
+            },
+            "validation_accuracy": 0.3616666666666667,
+            "fit_s": 0.0483227920049103,
+            "dictionary_or_sample_bytes": 24576
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "shards": 8
+            },
+            "validation_accuracy": 0.3883333333333333,
+            "fit_s": 0.0537706250033807,
+            "dictionary_or_sample_bytes": 49152
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "shards": 4
+            },
+            "validation_accuracy": 0.515,
+            "fit_s": 0.07201379199977964,
+            "dictionary_or_sample_bytes": 98304
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "shards": 8
+            },
+            "validation_accuracy": 0.48,
+            "fit_s": 0.08771720799995819,
+            "dictionary_or_sample_bytes": 195367
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "min"
+            },
+            "validation_accuracy": 0.5066666666666667,
+            "fit_s": 0.002359584002988413,
+            "dictionary_or_sample_bytes": 979233
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.535,
+            "fit_s": 0.0026667080019251443,
+            "dictionary_or_sample_bytes": 979233
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "min"
+            },
+            "validation_accuracy": 0.5033333333333333,
+            "fit_s": 0.002307417002157308,
+            "dictionary_or_sample_bytes": 979221
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.5533333333333333,
+            "fit_s": 0.00223658300092211,
+            "dictionary_or_sample_bytes": 979221
+          },
+          {
+            "family": "gzip-knn",
+            "config": {
+              "k": 1
+            },
+            "validation_accuracy": 0.25833333333333336,
+            "fit_s": 0.007107000004907604,
+            "dictionary_or_sample_bytes": 98188
+          },
+          {
+            "family": "gzip-knn",
+            "config": {
+              "k": 5
+            },
+            "validation_accuracy": 0.24166666666666667,
+            "fit_s": 0.006843499999376945,
+            "dictionary_or_sample_bytes": 98188
+          },
+          {
+            "family": "tfidf-logistic",
+            "config": {},
+            "validation_accuracy": 0.8416666666666667,
+            "fit_s": 3.909456957997463,
+            "dictionary_or_sample_bytes": null
+          }
+        ],
+        "selected": {
+          "deflate": {
+            "family": "deflate",
+            "config": {
+              "codec": "deflate",
+              "size": 32768
+            },
+            "validation_accuracy": 0.43833333333333335,
+            "fit_s": 0.00225687499914784,
+            "dictionary_or_sample_bytes": 196608,
+            "temperature": 3.410951603860984,
+            "accuracy": 0.406,
+            "macro_f1": 0.3684907264569877,
+            "ece": 0.03192085479609739,
+            "nll": 1.486202601279427,
+            "brier": 0.7179354684289194,
+            "tie_rate": 0.168,
+            "p50_ms": 0.0647915,
+            "p95_ms": 0.10161844999999997,
+            "serial_items_per_s": 14590.575182943197,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.3638296860034575,
+              0.44960374228035244
+            ],
+            "confidence_at_least_90pct": {
+              "n": 1,
+              "coverage": 0.002,
+              "accuracy": 1.0
+            }
+          },
+          "zstd": {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16384,
+              "trained": true
+            },
+            "validation_accuracy": 0.44666666666666666,
+            "fit_s": 0.058133709004323464,
+            "dictionary_or_sample_bytes": 98304,
+            "temperature": 2.709220452261488,
+            "accuracy": 0.454,
+            "macro_f1": 0.4018783545119495,
+            "ece": 0.05846382416841619,
+            "nll": 1.463362818214318,
+            "brier": 0.698117185244948,
+            "tie_rate": 0.214,
+            "p50_ms": 0.0115,
+            "p95_ms": 0.017131249999999997,
+            "serial_items_per_s": 81165.0627909159,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.4108749466266584,
+              0.49782651827818475
+            ],
+            "confidence_at_least_90pct": {
+              "n": 2,
+              "coverage": 0.004,
+              "accuracy": 0.5
+            }
+          },
+          "zstd-mixture": {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.5533333333333333,
+            "fit_s": 0.00223658300092211,
+            "dictionary_or_sample_bytes": 979221,
+            "temperature": 2.925419034377105,
+            "accuracy": 0.562,
+            "macro_f1": 0.46353991429511865,
+            "ece": 0.08457185931239918,
+            "nll": 1.2537522096435618,
+            "brier": 0.597552891264581,
+            "tie_rate": 0.038,
+            "p50_ms": 0.103604,
+            "p95_ms": 0.2117881499999997,
+            "serial_items_per_s": 8727.203975946568,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.5182021184950548,
+              0.6048524288071132
+            ],
+            "confidence_at_least_90pct": {
+              "n": 22,
+              "coverage": 0.044,
+              "accuracy": 0.8181818181818182
+            }
+          },
+          "gzip-knn": {
+            "family": "gzip-knn",
+            "config": {
+              "k": 1
+            },
+            "validation_accuracy": 0.25833333333333336,
+            "fit_s": 0.007107000004907604,
+            "dictionary_or_sample_bytes": 98188,
+            "temperature": 8.570386453309192,
+            "accuracy": 0.298,
+            "macro_f1": 0.2781904296547765,
+            "ece": 0.0427759626994248,
+            "nll": 1.743637872594331,
+            "brier": 0.814830939581835,
+            "tie_rate": 0.0,
+            "p50_ms": 5.6842915000000005,
+            "p95_ms": 6.971593449999999,
+            "serial_items_per_s": 171.43746654479273,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.25957253815100145,
+              0.3395078077354835
+            ],
+            "confidence_at_least_90pct": {
+              "n": 0,
+              "coverage": 0.0,
+              "accuracy": null
+            }
+          },
+          "tfidf-logistic": {
+            "family": "tfidf-logistic",
+            "config": {},
+            "validation_accuracy": 0.8416666666666667,
+            "fit_s": 3.909456957997463,
+            "dictionary_or_sample_bytes": null,
+            "temperature": 0.5834042638846706,
+            "accuracy": 0.852,
+            "macro_f1": 0.7837710838613873,
+            "ece": 0.030185522477487146,
+            "nll": 0.4129957238398832,
+            "brier": 0.22078743602615492,
+            "tie_rate": 0.0,
+            "p50_ms": 0.426187,
+            "p95_ms": 0.5887381499999998,
+            "serial_items_per_s": 2240.786993074041,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.818193200072725,
+              0.8804390684815191
+            ],
+            "confidence_at_least_90pct": {
+              "n": 322,
+              "coverage": 0.644,
+              "accuracy": 0.9472049689440993
+            }
+          }
+        }
+      }
+    },
+    "jevbench": {
+      "easy": {
+        "n": 48,
+        "sha256": "231df3c2c8e88a1a8c137ebe85de96ba70fabd330849098ac7b3c52c70b7172b",
+        "accuracy": 0.4583333333333333,
+        "uniform_chance": 0.284375,
+        "p50_ms": 0.074271,
+        "families": {
+          "extraction": {
+            "n": 12,
+            "accuracy": 0.4166666666666667
+          },
+          "fact": {
+            "n": 12,
+            "accuracy": 0.5833333333333334
+          },
+          "intent": {
+            "n": 12,
+            "accuracy": 0.4166666666666667
+          },
+          "tool_selection": {
+            "n": 12,
+            "accuracy": 0.4166666666666667
+          }
+        }
+      },
+      "original": {
+        "n": 72,
+        "sha256": "5c2414edb3006b8bfcb70fda433f0f9ca015759433849f8d3104328a1f7c4180",
+        "accuracy": 0.3611111111111111,
+        "uniform_chance": 0.3111111111111111,
+        "p50_ms": 0.0727085,
+        "families": {
+          "adequacy": {
+            "n": 12,
+            "accuracy": 0.4166666666666667
+          },
+          "extraction": {
+            "n": 12,
+            "accuracy": 0.3333333333333333
+          },
+          "intent": {
+            "n": 12,
+            "accuracy": 0.16666666666666666
+          },
+          "ordinal": {
+            "n": 12,
+            "accuracy": 0.4166666666666667
+          },
+          "policy": {
+            "n": 12,
+            "accuracy": 0.5
+          },
+          "routing": {
+            "n": 12,
+            "accuracy": 0.3333333333333333
+          }
+        }
+      },
+      "hard": {
+        "n": 111,
+        "sha256": "89e9e6becb33ed88c1de7d42dcc87531b2fb64cfaef4e1986faf7c37b3f80ebb",
+        "accuracy": 0.42342342342342343,
+        "uniform_chance": 0.3361861861861862,
+        "p50_ms": 0.183583,
+        "families": {
+          "adversarial": {
+            "n": 6,
+            "accuracy": 0.6666666666666666
+          },
+          "ambiguous": {
+            "n": 7,
+            "accuracy": 0.42857142857142855
+          },
+          "judge_hard": {
+            "n": 17,
+            "accuracy": 0.4117647058823529
+          },
+          "long_policy": {
+            "n": 19,
+            "accuracy": 0.21052631578947367
+          },
+          "multi_hop": {
+            "n": 18,
+            "accuracy": 0.2222222222222222
+          },
+          "probability": {
+            "n": 10,
+            "accuracy": 0.5
+          },
+          "routing_hard": {
+            "n": 5,
+            "accuracy": 1.0
+          },
+          "temporal_numeric": {
+            "n": 15,
+            "accuracy": 0.3333333333333333
+          },
+          "tradeoff": {
+            "n": 6,
+            "accuracy": 1.0
+          },
+          "trap": {
+            "n": 8,
+            "accuracy": 0.5
+          }
+        }
+      },
+      "source_commit": "2fa63fa3226cb369795525ed011800f57dcbd894",
+      "method": "Zero-shot symmetric gzip NCD against label and rubric; no training or calibration. Public diagnostic, not an official score."
+    }
+  },
+  "large_training_run": {
+    "manifest": {
+      "args": {
+        "datasets": [
+          "agnews",
+          "sst2"
+        ],
+        "families": [
+          "zstd",
+          "zstd-mixture",
+          "tfidf-logistic"
+        ],
+        "seed": 0,
+        "n": 500,
+        "train_cap": 120000,
+        "validation_n": 600,
+        "out": "eval/data/compression-large",
+        "jevbench": null
+      },
+      "platform": "macOS-26.5.1-arm64-arm-64bit",
+      "processor": "arm",
+      "python": "3.12.14",
+      "zlib": "1.2.12",
+      "thread_environment": {
+        "OMP_NUM_THREADS": "1",
+        "OPENBLAS_NUM_THREADS": "1"
+      },
+      "timing": "Serial local warm inference: normalization, compression/features, probabilities, tie-breaking. Excludes training, network and data loading.",
+      "versions": {
+        "datasets": "5.0.1",
+        "numpy": "2.5.3",
+        "scipy": "1.18.1",
+        "scikit-learn": "1.9.1",
+        "zstandard": "0.25.0"
+      },
+      "script_sha256": "5342902063b3071509c854905aa18c2410c8202e74c1fcad2ab76ae89491a2f4"
+    },
+    "datasets": {
+      "agnews": {
+        "dataset": {
+          "repo": "fancyzhx/ag_news",
+          "hf_fingerprints": {
+            "train": "2bc1e2be15d7f2ee",
+            "test": "0ebcc57b6aea2fc9"
+          },
+          "train_n": 119239,
+          "validation_n": 600,
+          "test_n": 500,
+          "removed_duplicates_or_eval_overlaps": 161,
+          "hashes": {
+            "train": "d1fb37dc340dc5ed27d4162ef9a98b4e83a80b5dc958e3e4e97101aa0f9e62ed",
+            "validation": "89cea8648589dcdb7aacc5e04f4efcd1b6058e5946f346d672c492fe8b03d403",
+            "test": "76c1f71538ffa2609b9dbb9bbf842e60d19b79470d1f4d0cc554d88d95ae47ed"
+          },
+          "labels": [
+            "World",
+            "Sports",
+            "Business",
+            "Sci/Tech"
+          ]
+        },
+        "validation_sweep": [
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "trained": true
+            },
+            "validation_accuracy": 0.655,
+            "fit_s": 1.095241957998951,
+            "dictionary_or_sample_bytes": 4096
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "trained": false
+            },
+            "validation_accuracy": 0.38333333333333336,
+            "fit_s": 0.05557016700186068,
+            "dictionary_or_sample_bytes": 4096
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "trained": true
+            },
+            "validation_accuracy": 0.7733333333333333,
+            "fit_s": 1.6625714999972843,
+            "dictionary_or_sample_bytes": 16384
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "trained": false
+            },
+            "validation_accuracy": 0.5033333333333333,
+            "fit_s": 0.0551289579962031,
+            "dictionary_or_sample_bytes": 16384
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16384,
+              "trained": true
+            },
+            "validation_accuracy": 0.805,
+            "fit_s": 1.376491791997978,
+            "dictionary_or_sample_bytes": 65536
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16384,
+              "trained": false
+            },
+            "validation_accuracy": 0.6483333333333333,
+            "fit_s": 0.05820737500471296,
+            "dictionary_or_sample_bytes": 65536
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 65536,
+              "level": 9
+            },
+            "validation_accuracy": 0.795,
+            "fit_s": 1.1389680420033983,
+            "dictionary_or_sample_bytes": 262144
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 262144,
+              "level": 9
+            },
+            "validation_accuracy": 0.7983333333333333,
+            "fit_s": 1.0690593339968473,
+            "dictionary_or_sample_bytes": 1048576
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "level": 9,
+              "trained": false
+            },
+            "validation_accuracy": 0.855,
+            "fit_s": 0.05941937500028871,
+            "dictionary_or_sample_bytes": 28247565
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "shards": 4
+            },
+            "validation_accuracy": 0.7283333333333334,
+            "fit_s": 1.1204033750036615,
+            "dictionary_or_sample_bytes": 16384
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "shards": 8
+            },
+            "validation_accuracy": 0.73,
+            "fit_s": 1.1286926250031684,
+            "dictionary_or_sample_bytes": 32768
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "shards": 4
+            },
+            "validation_accuracy": 0.8066666666666666,
+            "fit_s": 1.700393332997919,
+            "dictionary_or_sample_bytes": 65536
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "shards": 8
+            },
+            "validation_accuracy": 0.8166666666666667,
+            "fit_s": 1.7476613329999964,
+            "dictionary_or_sample_bytes": 131072
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "min"
+            },
+            "validation_accuracy": 0.8766666666666667,
+            "fit_s": 0.0598740830027964,
+            "dictionary_or_sample_bytes": 28247557
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.8983333333333333,
+            "fit_s": 0.057473291002679616,
+            "dictionary_or_sample_bytes": 28247557
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "min"
+            },
+            "validation_accuracy": 0.8966666666666666,
+            "fit_s": 0.058699958004581276,
+            "dictionary_or_sample_bytes": 28247549
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.9116666666666666,
+            "fit_s": 0.060205792004126124,
+            "dictionary_or_sample_bytes": 28247549
+          },
+          {
+            "family": "tfidf-logistic",
+            "config": {},
+            "validation_accuracy": 0.9366666666666666,
+            "fit_s": 72.26334416699683,
+            "dictionary_or_sample_bytes": null
+          }
+        ],
+        "selected": {
+          "zstd": {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "level": 9,
+              "trained": false
+            },
+            "validation_accuracy": 0.855,
+            "fit_s": 0.05941937500028871,
+            "dictionary_or_sample_bytes": 28247565,
+            "temperature": 5.838236970939662,
+            "accuracy": 0.856,
+            "macro_f1": 0.8578350680660315,
+            "ece": 0.07148579683576123,
+            "nll": 0.573052244855724,
+            "brier": 0.2274947566483206,
+            "tie_rate": 0.02,
+            "p50_ms": 0.0481665,
+            "p95_ms": 0.09512054999999997,
+            "serial_items_per_s": 16460.801843188412,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.822508879125479,
+              0.8840623924805179
+            ],
+            "confidence_at_least_90pct": {
+              "n": 246,
+              "coverage": 0.492,
+              "accuracy": 0.9471544715447154
+            }
+          },
+          "zstd-mixture": {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.9116666666666666,
+            "fit_s": 0.060205792004126124,
+            "dictionary_or_sample_bytes": 28247549,
+            "temperature": 3.683149054535093,
+            "accuracy": 0.9,
+            "macro_f1": 0.8997427715689791,
+            "ece": 0.05852817346189611,
+            "nll": 0.4590178473955788,
+            "brier": 0.17604696824424826,
+            "tie_rate": 0.004,
+            "p50_ms": 0.24593700000000002,
+            "p95_ms": 0.4126940499999994,
+            "serial_items_per_s": 3567.4229415418463,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.8705774917999974,
+              0.9233228133752799
+            ],
+            "confidence_at_least_90pct": {
+              "n": 360,
+              "coverage": 0.72,
+              "accuracy": 0.9472222222222222
+            }
+          },
+          "tfidf-logistic": {
+            "family": "tfidf-logistic",
+            "config": {},
+            "validation_accuracy": 0.9366666666666666,
+            "fit_s": 72.26334416699683,
+            "dictionary_or_sample_bytes": null,
+            "temperature": 0.9247663594340342,
+            "accuracy": 0.912,
+            "macro_f1": 0.9127427169277376,
+            "ece": 0.02759829499776148,
+            "nll": 0.28636336613317387,
+            "brier": 0.14006859532256022,
+            "tie_rate": 0.0,
+            "p50_ms": 0.593167,
+            "p95_ms": 0.7779229499999999,
+            "serial_items_per_s": 1645.5464238809047,
+            "n": 500,
+            "accuracy_wilson_95": [
+              0.8839229514479334,
+              0.9337943628826022
+            ],
+            "confidence_at_least_90pct": {
+              "n": 397,
+              "coverage": 0.794,
+              "accuracy": 0.9697732997481109
+            }
+          }
+        }
+      },
+      "sst2": {
+        "dataset": {
+          "repo": "stanfordnlp/sst2",
+          "hf_fingerprints": {
+            "train": "e7e6610fdbb3c277",
+            "validation": "c1ddc6497ec97f98",
+            "test": "14a60218fd61ec38"
+          },
+          "train_n": 66378,
+          "validation_n": 600,
+          "test_n": 500,
+          "removed_duplicates_or_eval_overlaps": 371,
+          "hashes": {
+            "train": "488b7c60c5677461e842fc5ecd32774c1ddbdd5878c8e6c793c7430eeb5720e1",
+            "validation": "745798b437dae118d5c5f3498399b935f25a45040fb07d6423a8d18964f91fd7",
+            "test": "380747f6d5a4f05ce8516c891e34d9dd29500beace1c0937b524634411bb287e"
+          },
+          "labels": [
+            "negative",
+            "positive"
+          ]
+        },
+        "validation_sweep": [
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "trained": true
+            },
+            "validation_accuracy": 0.57,
+            "fit_s": 0.1833836249934393,
+            "dictionary_or_sample_bytes": 2048
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "trained": false
+            },
+            "validation_accuracy": 0.54,
+            "fit_s": 0.01816158300061943,
+            "dictionary_or_sample_bytes": 2048
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "trained": true
+            },
+            "validation_accuracy": 0.61,
+            "fit_s": 0.2721342500008177,
+            "dictionary_or_sample_bytes": 8192
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "trained": false
+            },
+            "validation_accuracy": 0.56,
+            "fit_s": 0.019221125003241468,
+            "dictionary_or_sample_bytes": 8192
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16384,
+              "trained": true
+            },
+            "validation_accuracy": 0.6816666666666666,
+            "fit_s": 0.2559416250005597,
+            "dictionary_or_sample_bytes": 32768
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16384,
+              "trained": false
+            },
+            "validation_accuracy": 0.6333333333333333,
+            "fit_s": 0.018674541999644134,
+            "dictionary_or_sample_bytes": 32768
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 65536,
+              "level": 9
+            },
+            "validation_accuracy": 0.6616666666666666,
+            "fit_s": 0.2445404169993708,
+            "dictionary_or_sample_bytes": 131072
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 262144,
+              "level": 9
+            },
+            "validation_accuracy": 0.7466666666666667,
+            "fit_s": 0.2638120420015184,
+            "dictionary_or_sample_bytes": 524288
+          },
+          {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "level": 9,
+              "trained": false
+            },
+            "validation_accuracy": 0.7816666666666666,
+            "fit_s": 0.018079166002280544,
+            "dictionary_or_sample_bytes": 3567328
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "shards": 4
+            },
+            "validation_accuracy": 0.6133333333333333,
+            "fit_s": 0.1946094590020948,
+            "dictionary_or_sample_bytes": 8192
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 1024,
+              "shards": 8
+            },
+            "validation_accuracy": 0.5866666666666667,
+            "fit_s": 0.19280641600198578,
+            "dictionary_or_sample_bytes": 16384
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "shards": 4
+            },
+            "validation_accuracy": 0.6583333333333333,
+            "fit_s": 0.2784915840020403,
+            "dictionary_or_sample_bytes": 32768
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 4096,
+              "shards": 8
+            },
+            "validation_accuracy": 0.66,
+            "fit_s": 0.28512387500086334,
+            "dictionary_or_sample_bytes": 65536
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "min"
+            },
+            "validation_accuracy": 0.7916666666666666,
+            "fit_s": 0.01775899999483954,
+            "dictionary_or_sample_bytes": 3567324
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.79,
+            "fit_s": 0.0159552499972051,
+            "dictionary_or_sample_bytes": 3567324
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "min"
+            },
+            "validation_accuracy": 0.7883333333333333,
+            "fit_s": 0.017087957996409386,
+            "dictionary_or_sample_bytes": 3567320
+          },
+          {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 5,
+              "trained": false,
+              "level": 9,
+              "reduction": "mean"
+            },
+            "validation_accuracy": 0.79,
+            "fit_s": 0.016865915997186676,
+            "dictionary_or_sample_bytes": 3567320
+          },
+          {
+            "family": "tfidf-logistic",
+            "config": {},
+            "validation_accuracy": 0.915,
+            "fit_s": 3.577201125001011,
+            "dictionary_or_sample_bytes": null
+          }
+        ],
+        "selected": {
+          "zstd": {
+            "family": "zstd",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "level": 9,
+              "trained": false
+            },
+            "validation_accuracy": 0.7816666666666666,
+            "fit_s": 0.018079166002280544,
+            "dictionary_or_sample_bytes": 3567328,
+            "temperature": 8.570386453309192,
+            "accuracy": 0.662,
+            "macro_f1": 0.6608591301137025,
+            "ece": 0.04250272619027978,
+            "nll": 0.6126490882855867,
+            "brier": 0.4222055784678397,
+            "tie_rate": 0.09,
+            "p50_ms": 0.015396,
+            "p95_ms": 0.025002049999999998,
+            "serial_items_per_s": 62596.52383976092,
+            "n": 500,
+            "auroc": 0.7419633709016394,
+            "accuracy_wilson_95": [
+              0.6194419387868491,
+              0.7020876848091384
+            ],
+            "confidence_at_least_90pct": {
+              "n": 6,
+              "coverage": 0.012,
+              "accuracy": 0.6666666666666666
+            }
+          },
+          "zstd-mixture": {
+            "family": "zstd-mixture",
+            "config": {
+              "codec": "zstd",
+              "size": 16777216,
+              "shards": 3,
+              "trained": false,
+              "level": 9,
+              "reduction": "min"
+            },
+            "validation_accuracy": 0.7916666666666666,
+            "fit_s": 0.01775899999483954,
+            "dictionary_or_sample_bytes": 3567324,
+            "temperature": 7.3504331266672,
+            "accuracy": 0.7,
+            "macro_f1": 0.6996106954613179,
+            "ece": 0.059913157093607276,
+            "nll": 0.6007650774839849,
+            "brier": 0.4103150298882698,
+            "tie_rate": 0.084,
+            "p50_ms": 0.034,
+            "p95_ms": 0.06059134999999999,
+            "serial_items_per_s": 28267.394171399006,
+            "n": 500,
+            "auroc": 0.7595174820696722,
+            "accuracy_wilson_95": [
+              0.6584314090816449,
+              0.7385187435059938
+            ],
+            "confidence_at_least_90pct": {
+              "n": 13,
+              "coverage": 0.026,
+              "accuracy": 0.8461538461538461
+            }
+          },
+          "tfidf-logistic": {
+            "family": "tfidf-logistic",
+            "config": {},
+            "validation_accuracy": 0.915,
+            "fit_s": 3.577201125001011,
+            "dictionary_or_sample_bytes": null,
+            "temperature": 0.7345152395906567,
+            "accuracy": 0.85,
+            "macro_f1": 0.8499513842484965,
+            "ece": 0.038211722573768335,
+            "nll": 0.3873007503417493,
+            "brier": 0.22026970652824607,
+            "tie_rate": 0.0,
+            "p50_ms": 0.43174999999999997,
+            "p95_ms": 0.5471291999999999,
+            "serial_items_per_s": 2263.669087689429,
+            "n": 500,
+            "auroc": 0.9229956454918032,
+            "accuracy_wilson_95": [
+              0.8160382472597502,
+              0.8786245197686174
+            ],
+            "confidence_at_least_90pct": {
+              "n": 320,
+              "coverage": 0.64,
+              "accuracy": 0.940625
+            }
+          }
+        }
+      }
+    }
+  },
+  "verification": {
+    "predictions_validated": 10000,
+    "dictionary_and_lexical_predictions_reproduce_exactly": true,
+    "script_hash_matches": true,
+    "jevbench": {
+      "official_label_scoring_matches": true,
+      "option_reversal_invariant": true,
+      "results": [
+        {
+          "tier": "easy",
+          "n": 48,
+          "correct": 22,
+          "accuracy": 0.4583333333333333
+        },
+        {
+          "tier": "hard",
+          "n": 111,
+          "correct": 47,
+          "accuracy": 0.42342342342342343
+        },
+        {
+          "tier": "original",
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      ]
+    },
+    "large_training_predictions_validated": 3000,
+    "large_training_uses_identical_test_sample": true
+  }
+}
diff --git a/eval/compression.py b/eval/compression.py
new file mode 100644
index 0000000..35cad0c
--- /dev/null
+++ b/eval/compression.py
@@ -0,0 +1,371 @@
+#!/usr/bin/env python3
+# /// script
+# requires-python = ">=3.12"
+# dependencies = ["datasets==5.0.1", "numpy==2.5.3", "scipy==1.18.1", "scikit-learn==1.9.1", "zstandard==0.25.0"]
+# ///
+"""Compression classification experiment. Run with: uv run eval/compression.py.
+
+All tuning uses held-back training rows. Each family gets one final test run.
+Raw predictions, validation sweeps, split hashes and environment are saved.
+"""
+import argparse
+import gzip
+import hashlib
+import importlib.metadata
+import json
+import os
+import platform
+import random
+import subprocess
+import time
+import zlib
+from pathlib import Path
+
+import numpy as np
+import zstandard as zstd
+from datasets import load_dataset
+from scipy.special import softmax
+from sklearn.feature_extraction.text import TfidfVectorizer
+from sklearn.pipeline import FeatureUnion
+from sklearn.linear_model import LogisticRegression
+from sklearn.metrics import accuracy_score, f1_score, log_loss, roc_auc_score
+from sklearn.model_selection import train_test_split
+
+DATASETS = {
+    "agnews": ("fancyzhx/ag_news", None, "text", "test"),
+    "banking77": ("legacy-datasets/banking77", None, "text", "test"),
+    "sst2": ("stanfordnlp/sst2", None, "sentence", "validation"),
+    "emotion": ("dair-ai/emotion", "split", "text", "test"),
+}
+
+
+def normalize(text):
+    return " ".join(text.lower().split())
+
+
+def digest(value):
+    return hashlib.sha256(json.dumps(value, ensure_ascii=False, sort_keys=True).encode()).hexdigest()
+
+
+def load_split(name, seed, n, train_cap, val_n):
+    repo, config, field, test = DATASETS[name]
+    ds = load_dataset(repo, config)
+    labels = ds["train"].features["label"].names
+    heldout = [(normalize(r[field]), int(r["label"])) for r in ds[test]]
+    rng = random.Random(seed)
+    rng.shuffle(heldout)
+    evaluation = heldout[:n]
+    seen = {x for x, _ in heldout}
+    pool = []
+    duplicates = 0
+    for r in ds["train"]:
+        x, y = normalize(r[field]), int(r["label"])
+        if x in seen:
+            duplicates += 1
+            continue
+        seen.add(x)
+        pool.append((x, y))
+    rng.shuffle(pool)
+    pool = pool[:train_cap + val_n]
+    train, val = train_test_split(pool, test_size=val_n, random_state=seed,
+                                  stratify=[r[1] for r in pool])
+    assert not ({r[0] for r in train} & {r[0] for r in val})
+    assert not ({r[0] for r in train + val} & {r[0] for r in heldout})
+    assert set(y for _, y in train) == set(range(len(labels)))
+    meta = {"repo": repo, "hf_fingerprints": {k: v._fingerprint for k, v in ds.items()},
+            "train_n": len(train), "validation_n": len(val), "test_n": len(evaluation),
+            "removed_duplicates_or_eval_overlaps": duplicates,
+            "hashes": {k: digest(v) for k, v in [("train", train), ("validation", val), ("test", evaluation)]},
+            "labels": labels}
+    return train, val, evaluation, meta
+
+
+class Dictionaries:
+    def __init__(self, rows, classes, codec, size, shards=1, seed=42, trained=True, level=3, reduction="min"):
+        self.codec, self.models, self.bytes = codec, [], 0
+        self.reduction = reduction
+        rng = random.Random(seed)
+        for label in range(classes):
+            texts = [x.encode() for x, y in rows if y == label]
+            rng.shuffle(texts)
+            models = []
+            for i in range(shards):
+                samples = texts[i::shards]
+                if not samples:
+                    raise ValueError("empty class shard")
+                if codec == "zstd":
+                    if trained:
+                        # Do not silently change the algorithm if training fails.
+                        dictionary = zstd.train_dictionary(size, samples, threads=0)
+                    else:
+                        dictionary = zstd.ZstdCompressionDict(b"\n".join(samples)[-size:], dict_type=zstd.DICT_TYPE_RAWCONTENT)
+                    self.bytes += len(dictionary.as_bytes())
+                    models.append(zstd.ZstdCompressor(level=level, dict_data=dictionary))
+                else:
+                    dictionary = b"\n".join(samples)[-size:]
+                    self.bytes += len(dictionary)
+                    models.append(zlib.compressobj(level=6, wbits=-15, zdict=dictionary))
+            self.models.append(models)
+
+    def score(self, text):
+        x = text.encode()
+        scores = []
+        for models in self.models:
+            lengths = []
+            for model in models:
+                if self.codec == "zstd":
+                    lengths.append(len(model.compress(x)))
+                else:
+                    c = model.copy()
+                    lengths.append(len(c.compress(x) + c.flush()))
+            scores.append(-float(np.mean(lengths)) if self.reduction == "mean" else -min(lengths))
+        return np.asarray(scores, dtype=float)
+
+
+class GzipKNN:
+    def __init__(self, rows, classes, k=5, cap=1000, seed=42):
+        rng = random.Random(seed)
+        rows = list(rows)
+        rng.shuffle(rows)
+        # Balanced memory avoids a majority-class prior from subsampling.
+        self.samples = []
+        for label in range(classes):
+            self.samples += [(x.encode(), y) for x, y in rows if y == label][:max(1, cap // classes)]
+        self.lengths = [len(gzip.compress(x, mtime=0)) for x, _ in self.samples]
+        self.tie_keys = [hashlib.sha256(x).hexdigest() for x, _ in self.samples]
+        self.classes, self.k = classes, k
+        self.bytes = sum(len(x) for x, _ in self.samples)
+
+    def score(self, text):
+        x = text.encode()
+        cx = len(gzip.compress(x, mtime=0))
+        distances = [(len(gzip.compress(z + b" " + x, mtime=0)) - min(cx, cz)) / max(cx, cz)
+                     for (z, _), cz in zip(self.samples, self.lengths)]
+        nearest = np.lexsort((self.tie_keys, distances))[:self.k]
+        votes = np.full(self.classes, 0.01)
+        for i in nearest:
+            votes[self.samples[i][1]] += 1
+        return np.log(votes)
+
+
+class Lexical:
+    def __init__(self, rows, classes):
+        self.vectorizer = FeatureUnion([
+            ("word", TfidfVectorizer(ngram_range=(1, 2), sublinear_tf=True, min_df=2)),
+            ("char", TfidfVectorizer(analyzer="char", ngram_range=(3, 5), sublinear_tf=True, min_df=2, max_features=100000)),
+        ])
+        x = self.vectorizer.fit_transform([x for x, _ in rows])
+        self.classifier = LogisticRegression(C=4, max_iter=500).fit(x, [y for _, y in rows])
+        self.bytes = None
+
+    def score(self, text):
+        return self.classifier.predict_log_proba(self.vectorizer.transform([text]))[0]
+
+
+def scores_for(model, rows, temp=1):
+    scores, times = [], []
+    for text, _ in rows:
+        start = time.perf_counter_ns()
+        score = model.score(normalize(text))
+        softmax(score / temp)
+        predictions([score], [text])
+        scores.append(score)
+        times.append((time.perf_counter_ns() - start) / 1e6)
+    return np.asarray(scores), times
+
+
+def predictions(scores, texts):
+    # Exact ties use a reproducible text hash instead of always selecting class 0.
+    out = []
+    for score, text in zip(scores, texts):
+        ties = np.flatnonzero(score == score.max())
+        index = int(hashlib.sha256(text.encode()).hexdigest(), 16) % len(ties)
+        out.append(int(ties[index]))
+    return np.asarray(out)
+
+
+def temperature(scores, gold):
+    choices = np.geomspace(.05, 100, 100)
+    losses = [log_loss(gold, softmax(scores / t, axis=1), labels=np.arange(scores.shape[1])) for t in choices]
+    return float(choices[np.argmin(losses)])
+
+
+def metrics(scores, rows, times, temp):
+    gold = np.asarray([y for _, y in rows])
+    pred = predictions(scores, [x for x, _ in rows])
+    prob = softmax(scores / temp, axis=1)
+    conf = prob.max(axis=1)
+    correct = pred == gold
+    ece = 0.
+    for i in range(10):
+        mask = (conf >= i / 10) & ((conf < (i + 1) / 10) if i < 9 else (conf <= 1))
+        if mask.any():
+            ece += mask.mean() * abs(correct[mask].mean() - conf[mask].mean())
+    result = {"accuracy": float(correct.mean()), "macro_f1": float(f1_score(gold, pred, labels=np.arange(scores.shape[1]), average="macro", zero_division=0)),
+              "ece": float(ece), "nll": float(log_loss(gold, prob, labels=np.arange(scores.shape[1]))),
+              "brier": float(np.mean(np.sum((prob - np.eye(scores.shape[1])[gold]) ** 2, axis=1))),
+              "tie_rate": float(np.mean((scores == scores.max(axis=1, keepdims=True)).sum(axis=1) > 1)),
+              "p50_ms": float(np.median(times)), "p95_ms": float(np.percentile(times, 95)),
+              "serial_items_per_s": float(1000 / np.mean(times)), "n": len(rows)}
+    if scores.shape[1] == 2:
+        result["auroc"] = float(roc_auc_score(gold, prob[:, 1]))
+    n, p, z = len(rows), float(correct.mean()), 1.96
+    center = (p + z * z / (2 * n)) / (1 + z * z / n)
+    radius = z * np.sqrt(p * (1 - p) / n + z * z / (4 * n * n)) / (1 + z * z / n)
+    result["accuracy_wilson_95"] = [float(center - radius), float(center + radius)]
+    mask = conf >= .9
+    result["confidence_at_least_90pct"] = {
+        "n": int(mask.sum()), "coverage": float(mask.mean()),
+        "accuracy": float(correct[mask].mean()) if mask.any() else None,
+    }
+    return result, [{"text_sha256": hashlib.sha256(x.encode()).hexdigest(), "gold": int(y),
+                      "prediction": int(p), "probabilities": ps.tolist(), "ms": ms}
+                     for (x, y), p, ps, ms in zip(rows, pred, prob, times)]
+
+
+def run_dataset(name, args, outdir):
+    train, val, test, meta = load_split(name, args.seed, args.n, args.train_cap, args.validation_n)
+    classes = len(meta["labels"])
+    configs = [("deflate", {"codec": "deflate", "size": s}) for s in [4096, 32768]]
+    configs += [("zstd", {"codec": "zstd", "size": s, "trained": trained})
+                for s in [1024, 4096, 16384] for trained in [True, False]]
+    configs += [("zstd", {"codec": "zstd", "size": s, "level": 9})
+                for s in [65536, 262144]]
+    configs += [("zstd", {"codec": "zstd", "size": 16777216, "level": 9, "trained": False})]
+    configs += [("zstd-mixture", {"codec": "zstd", "size": s, "shards": shards})
+                for s in [1024, 4096] for shards in [4, 8]]
+    configs += [("zstd-mixture", {"codec": "zstd", "size": 16777216, "shards": shards,
+                                  "trained": False, "level": 9, "reduction": reduction})
+                for shards in [3, 5] for reduction in ["min", "mean"]]
+    configs += [("gzip-knn", {"k": k}) for k in [1, 5]]
+    configs += [("tfidf-logistic", {})]
+    if args.families:
+        configs = [(family, config) for family, config in configs if family in args.families]
+    best, sweep = {}, []
+    for family, config in configs:
+        start = time.perf_counter()
+        try:
+            if family == "gzip-knn":
+                model = GzipKNN(train, classes, seed=args.seed, **config)
+            elif family == "tfidf-logistic":
+                model = Lexical(train, classes)
+            else:
+                model = Dictionaries(train, classes, seed=args.seed, **config)
+        except zstd.ZstdError as exc:
+            sweep.append({"family": family, "config": config, "training_error": str(exc)})
+            continue
+        fit_s = time.perf_counter() - start
+        scores, times = scores_for(model, val)
+        pred = predictions(scores, [x for x, _ in val])
+        acc = float(accuracy_score([y for _, y in val], pred))
+        entry = {"family": family, "config": config, "validation_accuracy": acc,
+                 "fit_s": fit_s, "dictionary_or_sample_bytes": model.bytes}
+        sweep.append(entry)
+        print(name, family, config, f"validation={acc:.3f}", flush=True)
+        if family not in best or acc > best[family][0]["validation_accuracy"]:
+            best[family] = entry, model, scores
+    report = {"dataset": meta, "validation_sweep": sweep, "selected": {}}
+    for family, (entry, model, scores) in best.items():
+        temp = temperature(scores, [y for _, y in val])
+        test_scores, times = scores_for(model, test, temp)
+        summary, rows = metrics(test_scores, test, times, temp)
+        # OOD-like inputs must yield a finite score for every allowed class.
+        for text in ["", "🙂 café 你好", "x" * 50000]:
+            edge = model.score(text)
+            assert edge.shape == (classes,) and np.isfinite(edge).all()
+        report["selected"][family] = {**entry, "temperature": temp, **summary}
+        (outdir / f"{name}-{family}-predictions.json").write_text(json.dumps(rows))
+        print(name, family, "TEST", json.dumps(summary), flush=True)
+    (outdir / f"{name}.json").write_text(json.dumps(report, indent=2))
+    return report
+
+
+def predict_jevbench(state, question, labels):
+    if not isinstance(state, str):
+        state = json.dumps(state, ensure_ascii=False, sort_keys=True)
+    x = normalize(state + "\n" + question["instructions"]).encode()
+    cx = len(gzip.compress(x, mtime=0))
+    criteria = question.get("criteria") or {}
+    if isinstance(criteria, list):
+        criteria = {str(i): description for i, description in enumerate(criteria)}
+    scores = []
+    for label in labels:
+        key = {"yes": "true", "no": "false"}.get(label, label) if question["type"] == "noul" else label
+        description = criteria.get(key)
+        if description is None:
+            description = ""
+        z = normalize(label.replace("_", " ") + ": " + str(description)).encode()
+        cz = len(gzip.compress(z, mtime=0))
+        # Symmetric NCD: both concat directions, normalized for label length.
+        joint = min(len(gzip.compress(x + b" " + z, mtime=0)), len(gzip.compress(z + b" " + x, mtime=0)))
+        scores.append(-(joint - min(cx, cz)) / max(cx, cz))
+    scores = np.asarray(scores)
+    ties = np.flatnonzero(scores == scores.max())
+    # Choose by label name hash so option permutation does not change ties.
+    index = min(ties, key=lambda i: hashlib.sha256(x + labels[i].encode()).hexdigest())
+    return labels[index], len(ties) > 1
+
+
+def jevbench(root, outdir):
+    results = {}
+    for tier in ["easy", "original", "hard"]:
+        path = root / "datasets" / "public" / f"{tier}.jsonl"
+        rows = [json.loads(line) for line in path.read_text().splitlines() if line.strip()]
+        predictions_out = []
+        if any(r["expected"] is None or r.get("provenance", {}).get("exclude_reason") for r in rows):
+            raise ValueError("This diagnostic requires fully scored public cohorts")
+        for row in rows:
+            start = time.perf_counter_ns()
+            prediction, tie = predict_jevbench(row["state"], row["question"], row["labels"])
+            predictions_out.append({"id": row["id"], "group": row.get("group"), "family": row["family"],
+                                    "gold": str(row["expected"]), "prediction": prediction,
+                                    "tie": tie, "chance": 1 / len(row["labels"]),
+                                    "ms": (time.perf_counter_ns() - start) / 1e6})
+        family_scores = {}
+        for family in sorted({r["family"] for r in predictions_out}):
+            group = [r for r in predictions_out if r["family"] == family]
+            family_scores[family] = {"n": len(group), "accuracy": np.mean([r["gold"] == r["prediction"] for r in group])}
+        results[tier] = {"n": len(rows), "sha256": hashlib.sha256(path.read_bytes()).hexdigest(),
+                         "accuracy": float(np.mean([r["gold"] == r["prediction"] for r in predictions_out])),
+                         "uniform_chance": float(np.mean([r["chance"] for r in predictions_out])),
+                         "p50_ms": float(np.median([r["ms"] for r in predictions_out])), "families": family_scores}
+        (outdir / f"jevbench-{tier}-predictions.json").write_text(json.dumps(predictions_out, indent=2))
+    results["source_commit"] = subprocess.check_output(["git", "-C", str(root), "rev-parse", "HEAD"], text=True).strip()
+    results["method"] = "Zero-shot symmetric gzip NCD against label and rubric; no training or calibration. Public diagnostic, not an official score."
+    (outdir / "jevbench.json").write_text(json.dumps(results, indent=2))
+    print("JEVBENCH", json.dumps(results), flush=True)
+    return results
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--datasets", nargs="+", choices=DATASETS, default=list(DATASETS))
+    ap.add_argument("--families", nargs="+", choices=["deflate", "zstd", "zstd-mixture", "gzip-knn", "tfidf-logistic"])
+    ap.add_argument("--seed", type=int, default=0)
+    ap.add_argument("--n", type=int, default=500)
+    ap.add_argument("--train-cap", type=int, default=10000)
+    ap.add_argument("--validation-n", type=int, default=600)
+    ap.add_argument("--out", type=Path, default=Path(__file__).parent / "data" / "compression")
+    ap.add_argument("--jevbench", type=Path)
+    args = ap.parse_args()
+    if min(args.n, args.train_cap, args.validation_n) <= 0:
+        ap.error("sample counts must be positive")
+    args.out.mkdir(parents=True, exist_ok=False)
+    metadata = {"args": {k: str(v) if isinstance(v, Path) else v for k, v in vars(args).items()},
+                "platform": platform.platform(), "processor": platform.processor(),
+                "python": platform.python_version(), "zlib": zlib.ZLIB_RUNTIME_VERSION,
+                "thread_environment": {k: os.environ.get(k) for k in ["OMP_NUM_THREADS", "OPENBLAS_NUM_THREADS"]},
+                "timing": "Serial local warm inference: normalization, compression/features, probabilities, tie-breaking. Excludes training, network and data loading.",
+                "versions": {k: importlib.metadata.version(k) for k in ["datasets", "numpy", "scipy", "scikit-learn", "zstandard"]},
+                "script_sha256": hashlib.sha256(Path(__file__).read_bytes()).hexdigest()}
+    (args.out / "manifest.json").write_text(json.dumps(metadata, indent=2))
+    summary = {"manifest": metadata, "datasets": {}}
+    for name in args.datasets:
+        summary["datasets"][name] = run_dataset(name, args, args.out)
+    if args.jevbench:
+        summary["jevbench"] = jevbench(args.jevbench, args.out)
+    (args.out / "summary.json").write_text(json.dumps(summary, indent=2))
+
+
+if __name__ == "__main__":
+    main()

From 9005d6e48a2baa87317b115be37ea12a3bd39f09 Mon Sep 17 00:00:00 2001
From: Michael Ryaboy <michael@context.dev>
Date: Thu, 24 Sep 2026 02:38:37 -0700
Subject: [PATCH 2/2] Focus compression study on zero-shot JevBench performance

---
 eval/README.md                       |  149 +-
 eval/compression-results.json        | 2655 --------------------------
 eval/compression.py                  |  371 ----
 eval/jevbench-zero-shot-results.json | 2093 ++++++++++++++++++++
 eval/jevbench_hybrid.py              |  198 ++
 5 files changed, 2348 insertions(+), 3118 deletions(-)
 delete mode 100644 eval/compression-results.json
 delete mode 100644 eval/compression.py
 create mode 100644 eval/jevbench-zero-shot-results.json
 create mode 100644 eval/jevbench_hybrid.py

diff --git a/eval/README.md b/eval/README.md
index 24d5ea8..aa72838 100644
--- a/eval/README.md
+++ b/eval/README.md
@@ -120,103 +120,68 @@ In rough order of value: add cases from a real taxonomy with labels assigned
 before any model output is seen; split into tune and report halves; raise n past
 about thirty so confidence intervals mean something.
 
-## Compression classifiers
+## Zero-shot TF-IDF and compression on JevBench
 
-`compression.py` tests whether gzip/DEFLATE or Zstd can replace a learned text
-classifier. It compares gzip NCD nearest neighbors, per-class DEFLATE and Zstd
-dictionaries, Zstd dictionary mixtures, and a word/character TF-IDF logistic
-regression control. It does not change the Worker or call a paid model API.
+`jevbench_hybrid.py` measures arbitrary typed decisions using only the current
+state, instructions, labels, and option descriptions. There is no labeled
+training, support set, learned combination weight, or probability calibration.
+TF-IDF vocabulary and IDF are computed from the current request alone; compressor
+dictionaries are seeded with that request’s text at inference.
 
 ```sh
 git clone https://github.com/fstandhartinger/jevbench.git /tmp/compression-jevbench
 git -C /tmp/compression-jevbench checkout 2fa63fa3226cb369795525ed011800f57dcbd894
-OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 uv run --python 3.12 eval/compression.py \
-  --out eval/data/compression-reproduction --jevbench /tmp/compression-jevbench
+OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 uv run --python 3.12 eval/jevbench_hybrid.py \
+  --jevbench /tmp/compression-jevbench --out eval/data/jevbench-zero-shot-reproduction
 ```
 
-Run from the repository root. The output directory must be new. `uv` installs
-the script's pinned dependencies; Hugging Face downloads are cached. Omit
-`--jevbench` to run only the supervised datasets. `--datasets`, `--families`,
-`--seed`, `--train-cap`, `--validation-n`, and `--n` bound additional experiments.
-
-The default uses seed 0 and 500 evaluation rows, matching the sampling protocol
-of [dhruvmehra/jevbench](https://github.com/dhruvmehra/jevbench). Each dataset has
-up to 10,000 training rows and 600 separate validation rows. Training is shuffled
-and deduplicated after excluding exact normalized matches anywhere in the
-held-out split. Classifier settings and probability temperature are chosen only
-on validation rows, then frozen for evaluation. This excludes exact duplicates,
-not semantic paraphrases or SST-2's related review fragments. Test rows remain
-in their original benchmark sample, including any duplicates within that split.
-
-Gzip nearest neighbors deliberately retain at most 1,000 balanced examples; its
-accuracy is a bounded-compute baseline, not an estimate of full-data gzip kNN.
-DEFLATE uses raw zlib preset dictionaries (the same compression algorithm as
-gzip, without gzip framing). Zstd tries trained dictionaries and raw class
-text, with minimum or mean length over random shards. Compression length is
-only a score: softmax temperature is fitted on validation labels. Dictionaries
-store information from labeled examples; this is supervised learning, not a
-zero-shot or memory-free model. Oversized dictionary training failures are
-recorded in the validation sweep, never silently replaced by another method.
-
-Every run produces `summary.json`, per-dataset validation sweeps and selected
-settings, per-item predictions/probabilities, split hashes, dataset fingerprints,
-and software versions. Timing covers serial local normalization, scoring,
-probabilities and tie-breaking; it excludes training, loading, serving and
-network overhead. Empty, Unicode and long inputs are exercised for every
-selected supervised model. The output manifest identifies the script by hash.
-
-The optional Benchmark Heaven diagnostic uses symmetric gzip NCD between the
-state/instructions and each label/criterion. The prediction function accepts
-only those inference fields; expected answers and author rationales cannot
-enter it. This has no labeled support set and returns labels without invented
-calibrated probabilities. Its three public cohorts total 231 decisions. These
-are public diagnostic accuracies, not an official JevBench score: private,
-sealed, and imported tasks are absent. A supervised banking/news classifier
-cannot be submitted as if it solves arbitrary typed decisions.
-
-The compression idea comes from [Nathan Barry's gzip language-model experiment](https://nathan.rs/posts/gzip-lm/)
-and [FTCC's class-dictionary approach](https://github.com/cyrilou242/ftcc).
-The TF-IDF control matters: [Gzip versus bag-of-words for text classification](https://arxiv.org/abs/2307.15002)
-examines whether compression's gains survive comparison with conventional text
-features.
-
-### Measured results (Apple M5 Max)
-
-Generated from [compression-results.json](compression-results.json). All rows
-use 500 held-out items per dataset; percentages are accuracy. Each family selects
-its settings on validation data. These are exploratory measurements on one seed,
-not evidence that small differences are statistically significant.
-
-| Up to 10k training rows | Gzip kNN (≤1k memory) | DEFLATE | Zstd mixture | TF-IDF + logistic |
+Run from the repository root. The output directory must be new. Dependencies
+are pinned in the script; no API keys, model downloads, or paid calls are needed.
+The benchmark repository supplies the 231 public questions and published Jev
+outcomes for comparison. Private, sealed, and imported questions are absent.
+
+The fixed sweep contains 54 methods: five TF-IDF scores (word, window, character,
+label-only, and a window/character combination), three compression scores
+(gzip normalized compression distance, DEFLATE conditional gain, and Zstd
+conditional gain), all 45 pairwise blends at 25/50/75% TF-IDF weight, and one
+equal four-way blend. Scores are standardized within each question before
+combining them. Negligible numerical variation is collapsed and final scores
+rounded before ties are resolved by label name, matching JevBench’s scorer.
+
+### Public results
+
+Generated measurements are in [jevbench-zero-shot-results.json](jevbench-zero-shot-results.json).
+No training examples were used for any row. Jev is a published reference on
+the exact same item IDs, not a fresh API run.
+
+| Method | Overall (231) | Easy (48) | Original (72) | Hard (111) |
 |---|---:|---:|---:|---:|
-| agnews | 61.8% | 76.6% | 84.8% | 88.8% |
-| banking77 | 54.8% | 82.0% | 80.6% | 90.8% |
-| sst2 | 54.8% | 65.6% | 71.8% | 79.6% |
-| emotion | 29.8% | 40.6% | 56.2% | 85.2% |
-
-Larger-training follow-up (same evaluation items, separately held-back validation):
-
-| Dataset | Training rows | Zstd mixture | p50 | TF-IDF + logistic | p50 |
-|---|---:|---:|---:|---:|---:|
-| agnews | 119,239 | 90.0% | 0.246 ms | 91.2% | 0.593 ms |
-| sst2 | 66,378 | 70.0% | 0.034 ms | 85.0% | 0.432 ms |
-
-The larger news mixture retains 28.25 MB of class text across 20 dictionaries;
-this is not a tiny parameter-free model. It is competitive here, but the lexical
-control is more accurate on every measured dataset. More data does not resolve
-the sentiment weakness.
-
-Reproduce the larger run by adding `--datasets agnews sst2 --families zstd
-zstd-mixture tfidf-logistic --train-cap 120000` and choosing a new output directory.
-
-On Benchmark Heaven’s public cohorts, zero-shot gzip scores **45.8% easy**,
-**36.1% original**, and **42.3% hard**. All 231 labels were checked against the
-upstream scorer and were invariant to reversing the option order. No official
-score or leaderboard submission is claimed. The audit fixed ordinal gold labels
-being compared as integers against strings, and removed class-order bias when
-gzip neighbor distances tie.
-
-The checked-in JSON combines the two generated `summary.json` files and the
-audit record. Raw item predictions stay under `eval/data/`. Across the default
-and larger runs, 13,000 probability vectors and accuracies were independently
-checked; the default run’s 8,000 dictionary/lexical predictions reproduced exactly.
+| Gzip | 42.0% | 52.1% | 34.7% | 42.3% |
+| Character TF-IDF | 49.4% | 79.2% | 38.9% | 43.2% |
+| Window + character TF-IDF | 48.9% | 79.2% | 40.3% | 41.4% |
+| 75% TF-IDF + 25% Zstd (best observed blend) | 50.6% | 81.2% | 41.7% | 43.2% |
+| Jev 1.13.0, published | 86.6% | 100.0% | 98.6% | 73.0% |
+
+The best blend is the high end of an exploratory sweep, not an independently
+selected winner. It adds only three correct answers over character TF-IDF
+(+1.30 percentage points; paired scenario-bootstrap 95% interval −3.07 to +5.73
+points). This is not convincing evidence of an improvement. Uniform random
+choice averages 31.8% on these variable-sized label sets.
+
+The methods are fast but miss many reasoning decisions: the highlighted blend
+gets 5/19 long-policy questions and 5/18 multi-hop questions correct. Running
+all 54 scoring rules together took a median 1.53 ms per request on an Apple M5
+Max, including per-request TF-IDF fitting and feature extraction. This excludes
+loading, verification, and network/serving overhead. It is not an API latency.
+
+Every prediction function receives only inference fields. The harness scores
+answers afterward and checks every prediction again with reversed option order.
+All 12,474 predictions were independently checked with the upstream label scorer.
+No probabilities are invented from similarity scores; these are label accuracies,
+not official composite scores or a leaderboard submission.
+
+Each run writes per-item predictions and scores, the full per-family sweep,
+source/script hashes, software versions, timing, and paired bootstrap intervals
+to `eval/data/`. The checked-in summary retains the full accuracy sweep and
+selected family breakdowns; raw predictions stay in the ignored output folder.
+The Worker and deployed service are unchanged.
diff --git a/eval/compression-results.json b/eval/compression-results.json
deleted file mode 100644
index 3f9563e..0000000
--- a/eval/compression-results.json
+++ /dev/null
@@ -1,2655 +0,0 @@
-{
-  "default_run": {
-    "manifest": {
-      "args": {
-        "datasets": [
-          "agnews",
-          "banking77",
-          "sst2",
-          "emotion"
-        ],
-        "families": null,
-        "seed": 0,
-        "n": 500,
-        "train_cap": 10000,
-        "validation_n": 600,
-        "out": "eval/data/compression-verified",
-        "jevbench": "/tmp/classifier-gzip-jevbench"
-      },
-      "platform": "macOS-26.5.1-arm64-arm-64bit",
-      "processor": "arm",
-      "python": "3.12.14",
-      "zlib": "1.2.12",
-      "thread_environment": {
-        "OMP_NUM_THREADS": "1",
-        "OPENBLAS_NUM_THREADS": "1"
-      },
-      "timing": "Serial local warm inference: normalization, compression/features, probabilities, tie-breaking. Excludes training, network and data loading.",
-      "versions": {
-        "datasets": "5.0.1",
-        "numpy": "2.5.3",
-        "scipy": "1.18.1",
-        "scikit-learn": "1.9.1",
-        "zstandard": "0.25.0"
-      },
-      "script_sha256": "5342902063b3071509c854905aa18c2410c8202e74c1fcad2ab76ae89491a2f4"
-    },
-    "datasets": {
-      "agnews": {
-        "dataset": {
-          "repo": "fancyzhx/ag_news",
-          "hf_fingerprints": {
-            "train": "2bc1e2be15d7f2ee",
-            "test": "0ebcc57b6aea2fc9"
-          },
-          "train_n": 10000,
-          "validation_n": 600,
-          "test_n": 500,
-          "removed_duplicates_or_eval_overlaps": 161,
-          "hashes": {
-            "train": "04944cfb62e7665ceda0faf53c118d18743495868c573e315806570d7fa2c616",
-            "validation": "b8edd80e646b60c5ca1a76e2744cd41f336090666c93886b305374e69ad44ef6",
-            "test": "76c1f71538ffa2609b9dbb9bbf842e60d19b79470d1f4d0cc554d88d95ae47ed"
-          },
-          "labels": [
-            "World",
-            "Sports",
-            "Business",
-            "Sci/Tech"
-          ]
-        },
-        "validation_sweep": [
-          {
-            "family": "deflate",
-            "config": {
-              "codec": "deflate",
-              "size": 4096
-            },
-            "validation_accuracy": 0.5933333333333334,
-            "fit_s": 0.004283166999812238,
-            "dictionary_or_sample_bytes": 16384
-          },
-          {
-            "family": "deflate",
-            "config": {
-              "codec": "deflate",
-              "size": 32768
-            },
-            "validation_accuracy": 0.785,
-            "fit_s": 0.004138415999477729,
-            "dictionary_or_sample_bytes": 131072
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "trained": true
-            },
-            "validation_accuracy": 0.62,
-            "fit_s": 0.13306408299831674,
-            "dictionary_or_sample_bytes": 4096
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "trained": false
-            },
-            "validation_accuracy": 0.395,
-            "fit_s": 0.003096458996878937,
-            "dictionary_or_sample_bytes": 4096
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "trained": true
-            },
-            "validation_accuracy": 0.7116666666666667,
-            "fit_s": 0.148906624999654,
-            "dictionary_or_sample_bytes": 16384
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "trained": false
-            },
-            "validation_accuracy": 0.5316666666666666,
-            "fit_s": 0.0029460840014507994,
-            "dictionary_or_sample_bytes": 16384
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16384,
-              "trained": true
-            },
-            "validation_accuracy": 0.7816666666666666,
-            "fit_s": 0.12514475000352832,
-            "dictionary_or_sample_bytes": 65536
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16384,
-              "trained": false
-            },
-            "validation_accuracy": 0.6383333333333333,
-            "fit_s": 0.0035843750010826625,
-            "dictionary_or_sample_bytes": 65536
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 65536,
-              "level": 9
-            },
-            "validation_accuracy": 0.755,
-            "fit_s": 0.14723604200116824,
-            "dictionary_or_sample_bytes": 262144
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 262144,
-              "level": 9
-            },
-            "validation_accuracy": 0.7933333333333333,
-            "fit_s": 0.2483245409966912,
-            "dictionary_or_sample_bytes": 1048576
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "level": 9,
-              "trained": false
-            },
-            "validation_accuracy": 0.8466666666666667,
-            "fit_s": 0.003961707996495534,
-            "dictionary_or_sample_bytes": 2366771
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "shards": 4
-            },
-            "validation_accuracy": 0.7016666666666667,
-            "fit_s": 0.11391575000016019,
-            "dictionary_or_sample_bytes": 16384
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "shards": 8
-            },
-            "validation_accuracy": 0.7266666666666667,
-            "fit_s": 0.12054720900050597,
-            "dictionary_or_sample_bytes": 32768
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "shards": 4
-            },
-            "validation_accuracy": 0.7783333333333333,
-            "fit_s": 0.17654266599856783,
-            "dictionary_or_sample_bytes": 65536
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "shards": 8
-            },
-            "validation_accuracy": 0.7983333333333333,
-            "fit_s": 0.174204084003577,
-            "dictionary_or_sample_bytes": 131072
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "min"
-            },
-            "validation_accuracy": 0.8633333333333333,
-            "fit_s": 0.0033049999983632006,
-            "dictionary_or_sample_bytes": 2366763
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.895,
-            "fit_s": 0.0030992079991847277,
-            "dictionary_or_sample_bytes": 2366763
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "min"
-            },
-            "validation_accuracy": 0.8766666666666667,
-            "fit_s": 0.003387500000826549,
-            "dictionary_or_sample_bytes": 2366755
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.8833333333333333,
-            "fit_s": 0.003104374998656567,
-            "dictionary_or_sample_bytes": 2366755
-          },
-          {
-            "family": "gzip-knn",
-            "config": {
-              "k": 1
-            },
-            "validation_accuracy": 0.6116666666666667,
-            "fit_s": 0.010663459004717879,
-            "dictionary_or_sample_bytes": 232003
-          },
-          {
-            "family": "gzip-knn",
-            "config": {
-              "k": 5
-            },
-            "validation_accuracy": 0.6383333333333333,
-            "fit_s": 0.012955791004060302,
-            "dictionary_or_sample_bytes": 232003
-          },
-          {
-            "family": "tfidf-logistic",
-            "config": {},
-            "validation_accuracy": 0.9033333333333333,
-            "fit_s": 5.491404250002233,
-            "dictionary_or_sample_bytes": null
-          }
-        ],
-        "selected": {
-          "deflate": {
-            "family": "deflate",
-            "config": {
-              "codec": "deflate",
-              "size": 32768
-            },
-            "validation_accuracy": 0.785,
-            "fit_s": 0.004138415999477729,
-            "dictionary_or_sample_bytes": 131072,
-            "temperature": 4.29444227551508,
-            "accuracy": 0.766,
-            "macro_f1": 0.771227501914879,
-            "ece": 0.06008734299381312,
-            "nll": 0.6778785936410218,
-            "brier": 0.349022246408619,
-            "tie_rate": 0.03,
-            "p50_ms": 0.0734795,
-            "p95_ms": 0.09922049999999999,
-            "serial_items_per_s": 13096.827675922877,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.7269477983375824,
-              0.8009959046039771
-            ],
-            "confidence_at_least_90pct": {
-              "n": 180,
-              "coverage": 0.36,
-              "accuracy": 0.9222222222222223
-            }
-          },
-          "zstd": {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "level": 9,
-              "trained": false
-            },
-            "validation_accuracy": 0.8466666666666667,
-            "fit_s": 0.003961707996495534,
-            "dictionary_or_sample_bytes": 2366771,
-            "temperature": 4.637143191628681,
-            "accuracy": 0.808,
-            "macro_f1": 0.8072402391715388,
-            "ece": 0.057153539576628706,
-            "nll": 0.6757015422951654,
-            "brier": 0.2950965013185466,
-            "tie_rate": 0.022,
-            "p50_ms": 0.044521,
-            "p95_ms": 0.0784274,
-            "serial_items_per_s": 20026.02101066041,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.7711789076945584,
-              0.8401243272904052
-            ],
-            "confidence_at_least_90pct": {
-              "n": 263,
-              "coverage": 0.526,
-              "accuracy": 0.9125475285171103
-            }
-          },
-          "zstd-mixture": {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.895,
-            "fit_s": 0.0030992079991847277,
-            "dictionary_or_sample_bytes": 2366763,
-            "temperature": 3.410951603860984,
-            "accuracy": 0.848,
-            "macro_f1": 0.8480560127251228,
-            "ece": 0.04504076255490379,
-            "nll": 0.5399901438401714,
-            "brier": 0.22715867018883865,
-            "tie_rate": 0.006,
-            "p50_ms": 0.12854149999999998,
-            "p95_ms": 0.19065174999999993,
-            "serial_items_per_s": 7451.103215557868,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.8138851771600609,
-              0.8768080883424302
-            ],
-            "confidence_at_least_90pct": {
-              "n": 302,
-              "coverage": 0.604,
-              "accuracy": 0.9370860927152318
-            }
-          },
-          "gzip-knn": {
-            "family": "gzip-knn",
-            "config": {
-              "k": 5
-            },
-            "validation_accuracy": 0.6383333333333333,
-            "fit_s": 0.012955791004060302,
-            "dictionary_or_sample_bytes": 232003,
-            "temperature": 2.1518556436402267,
-            "accuracy": 0.618,
-            "macro_f1": 0.6199502440461163,
-            "ece": 0.10602590208369551,
-            "nll": 0.9974734324691097,
-            "brier": 0.5320566795008369,
-            "tie_rate": 0.2,
-            "p50_ms": 9.704708,
-            "p95_ms": 11.13455415,
-            "serial_items_per_s": 102.52232695558666,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.5746644739023564,
-              0.6595361161243504
-            ],
-            "confidence_at_least_90pct": {
-              "n": 0,
-              "coverage": 0.0,
-              "accuracy": null
-            }
-          },
-          "tfidf-logistic": {
-            "family": "tfidf-logistic",
-            "config": {},
-            "validation_accuracy": 0.9033333333333333,
-            "fit_s": 5.491404250002233,
-            "dictionary_or_sample_bytes": null,
-            "temperature": 0.7345152395906567,
-            "accuracy": 0.888,
-            "macro_f1": 0.8879638858343402,
-            "ece": 0.06050312885515615,
-            "nll": 0.44172785940001424,
-            "brier": 0.19518180715294506,
-            "tie_rate": 0.0,
-            "p50_ms": 0.5366455,
-            "p95_ms": 0.70514435,
-            "serial_items_per_s": 1814.9127050219765,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.8573456933034787,
-              0.9127376027165404
-            ],
-            "confidence_at_least_90pct": {
-              "n": 368,
-              "coverage": 0.736,
-              "accuracy": 0.9266304347826086
-            }
-          }
-        }
-      },
-      "banking77": {
-        "dataset": {
-          "repo": "legacy-datasets/banking77",
-          "hf_fingerprints": {
-            "train": "8360b37e52e4eb2a",
-            "test": "1d7c14b4c23350b9"
-          },
-          "train_n": 9392,
-          "validation_n": 600,
-          "test_n": 500,
-          "removed_duplicates_or_eval_overlaps": 11,
-          "hashes": {
-            "train": "3d092fadae6b94818c39469faa6c8997e81f07219855d5dd0a510fb78fd79fc6",
-            "validation": "4e12f6da49e8667d0a3464d75bd7c82ec2338a890264835660468ccad307fddd",
-            "test": "709b7b5509738c4bce13a37aa816b4934b835be78e91cf0c79792ba5b1716dae"
-          },
-          "labels": [
-            "activate_my_card",
-            "age_limit",
-            "apple_pay_or_google_pay",
-            "atm_support",
-            "automatic_top_up",
-            "balance_not_updated_after_bank_transfer",
-            "balance_not_updated_after_cheque_or_cash_deposit",
-            "beneficiary_not_allowed",
-            "cancel_transfer",
-            "card_about_to_expire",
-            "card_acceptance",
-            "card_arrival",
-            "card_delivery_estimate",
-            "card_linking",
-            "card_not_working",
-            "card_payment_fee_charged",
-            "card_payment_not_recognised",
-            "card_payment_wrong_exchange_rate",
-            "card_swallowed",
-            "cash_withdrawal_charge",
-            "cash_withdrawal_not_recognised",
-            "change_pin",
-            "compromised_card",
-            "contactless_not_working",
-            "country_support",
-            "declined_card_payment",
-            "declined_cash_withdrawal",
-            "declined_transfer",
-            "direct_debit_payment_not_recognised",
-            "disposable_card_limits",
-            "edit_personal_details",
-            "exchange_charge",
-            "exchange_rate",
-            "exchange_via_app",
-            "extra_charge_on_statement",
-            "failed_transfer",
-            "fiat_currency_support",
-            "get_disposable_virtual_card",
-            "get_physical_card",
-            "getting_spare_card",
-            "getting_virtual_card",
-            "lost_or_stolen_card",
-            "lost_or_stolen_phone",
-            "order_physical_card",
-            "passcode_forgotten",
-            "pending_card_payment",
-            "pending_cash_withdrawal",
-            "pending_top_up",
-            "pending_transfer",
-            "pin_blocked",
-            "receiving_money",
-            "Refund_not_showing_up",
-            "request_refund",
-            "reverted_card_payment?",
-            "supported_cards_and_currencies",
-            "terminate_account",
-            "top_up_by_bank_transfer_charge",
-            "top_up_by_card_charge",
-            "top_up_by_cash_or_cheque",
-            "top_up_failed",
-            "top_up_limits",
-            "top_up_reverted",
-            "topping_up_by_card",
-            "transaction_charged_twice",
-            "transfer_fee_charged",
-            "transfer_into_account",
-            "transfer_not_received_by_recipient",
-            "transfer_timing",
-            "unable_to_verify_identity",
-            "verify_my_identity",
-            "verify_source_of_funds",
-            "verify_top_up",
-            "virtual_card_not_working",
-            "visa_or_mastercard",
-            "why_verify_identity",
-            "wrong_amount_of_cash_received",
-            "wrong_exchange_rate_for_cash_withdrawal"
-          ]
-        },
-        "validation_sweep": [
-          {
-            "family": "deflate",
-            "config": {
-              "codec": "deflate",
-              "size": 4096
-            },
-            "validation_accuracy": 0.805,
-            "fit_s": 0.0074147080013062805,
-            "dictionary_or_sample_bytes": 304038
-          },
-          {
-            "family": "deflate",
-            "config": {
-              "codec": "deflate",
-              "size": 32768
-            },
-            "validation_accuracy": 0.7983333333333333,
-            "fit_s": 0.008205541002098471,
-            "dictionary_or_sample_bytes": 567780
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "trained": true
-            },
-            "validation_accuracy": 0.6466666666666666,
-            "fit_s": 0.05034095799783245,
-            "dictionary_or_sample_bytes": 78848
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "trained": false
-            },
-            "validation_accuracy": 0.6416666666666667,
-            "fit_s": 0.007061042000714224,
-            "dictionary_or_sample_bytes": 78848
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "trained": true
-            },
-            "validation_accuracy": 0.6933333333333334,
-            "fit_s": 0.08712499999819556,
-            "dictionary_or_sample_bytes": 286121
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "trained": false
-            },
-            "validation_accuracy": 0.7333333333333333,
-            "fit_s": 0.007319832999201026,
-            "dictionary_or_sample_bytes": 304038
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16384,
-              "trained": true
-            },
-            "validation_accuracy": 0.7133333333333334,
-            "fit_s": 0.12062429200159386,
-            "dictionary_or_sample_bytes": 411372
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16384,
-              "trained": false
-            },
-            "validation_accuracy": 0.7083333333333334,
-            "fit_s": 0.007289250002941117,
-            "dictionary_or_sample_bytes": 567780
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 65536,
-              "level": 9
-            },
-            "validation_accuracy": 0.7483333333333333,
-            "fit_s": 0.11870433299918659,
-            "dictionary_or_sample_bytes": 411372
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 262144,
-              "level": 9
-            },
-            "validation_accuracy": 0.7483333333333333,
-            "fit_s": 0.12351499999931548,
-            "dictionary_or_sample_bytes": 411372
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "level": 9,
-              "trained": false
-            },
-            "validation_accuracy": 0.7816666666666666,
-            "fit_s": 0.007592082998598926,
-            "dictionary_or_sample_bytes": 567780
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "shards": 4
-            },
-            "validation_accuracy": 0.72,
-            "fit_s": 0.12457012500090059,
-            "dictionary_or_sample_bytes": 281325
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "shards": 8
-            },
-            "training_error": "cannot train dict: Src size is incorrect"
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "shards": 4
-            },
-            "validation_accuracy": 0.7366666666666667,
-            "fit_s": 0.20517695900343824,
-            "dictionary_or_sample_bytes": 454504
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "shards": 8
-            },
-            "training_error": "cannot train dict: Src size is incorrect"
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "min"
-            },
-            "validation_accuracy": 0.8166666666666667,
-            "fit_s": 0.007201125001301989,
-            "dictionary_or_sample_bytes": 567626
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.835,
-            "fit_s": 0.007468750001862645,
-            "dictionary_or_sample_bytes": 567626
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "min"
-            },
-            "validation_accuracy": 0.8183333333333334,
-            "fit_s": 0.0075589999978546984,
-            "dictionary_or_sample_bytes": 567472
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.825,
-            "fit_s": 0.00782808299845783,
-            "dictionary_or_sample_bytes": 567472
-          },
-          {
-            "family": "gzip-knn",
-            "config": {
-              "k": 1
-            },
-            "validation_accuracy": 0.49833333333333335,
-            "fit_s": 0.011530167001183145,
-            "dictionary_or_sample_bytes": 52739
-          },
-          {
-            "family": "gzip-knn",
-            "config": {
-              "k": 5
-            },
-            "validation_accuracy": 0.45166666666666666,
-            "fit_s": 0.011287084002105985,
-            "dictionary_or_sample_bytes": 52739
-          },
-          {
-            "family": "tfidf-logistic",
-            "config": {},
-            "validation_accuracy": 0.9016666666666666,
-            "fit_s": 6.198137958999723,
-            "dictionary_or_sample_bytes": null
-          }
-        ],
-        "selected": {
-          "deflate": {
-            "family": "deflate",
-            "config": {
-              "codec": "deflate",
-              "size": 4096
-            },
-            "validation_accuracy": 0.805,
-            "fit_s": 0.0074147080013062805,
-            "dictionary_or_sample_bytes": 304038,
-            "temperature": 1.35753398137484,
-            "accuracy": 0.82,
-            "macro_f1": 0.7992486209415565,
-            "ece": 0.06290305365662752,
-            "nll": 0.748491242227106,
-            "brier": 0.28054942628044477,
-            "tie_rate": 0.094,
-            "p50_ms": 0.5353749999999999,
-            "p95_ms": 0.71514165,
-            "serial_items_per_s": 1798.0028453251189,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.7839246244600844,
-              0.8511956196801373
-            ],
-            "confidence_at_least_90pct": {
-              "n": 244,
-              "coverage": 0.488,
-              "accuracy": 0.9590163934426229
-            }
-          },
-          "zstd": {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "level": 9,
-              "trained": false
-            },
-            "validation_accuracy": 0.7816666666666666,
-            "fit_s": 0.007592082998598926,
-            "dictionary_or_sample_bytes": 567780,
-            "temperature": 1.5828442546700814,
-            "accuracy": 0.778,
-            "macro_f1": 0.7634533737731454,
-            "ece": 0.03987457002093967,
-            "nll": 0.8437418486929151,
-            "brier": 0.31676958909518704,
-            "tie_rate": 0.1,
-            "p50_ms": 0.14937499999999998,
-            "p95_ms": 0.3112707999999996,
-            "serial_items_per_s": 5918.135240090225,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.7395294756647782,
-              0.8122312364320395
-            ],
-            "confidence_at_least_90pct": {
-              "n": 234,
-              "coverage": 0.468,
-              "accuracy": 0.9700854700854701
-            }
-          },
-          "zstd-mixture": {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.835,
-            "fit_s": 0.007468750001862645,
-            "dictionary_or_sample_bytes": 567626,
-            "temperature": 1.35753398137484,
-            "accuracy": 0.806,
-            "macro_f1": 0.7948050126018442,
-            "ece": 0.027330381293454824,
-            "nll": 0.7245106932888512,
-            "brier": 0.27127663578583444,
-            "tie_rate": 0.026,
-            "p50_ms": 0.5831459999999999,
-            "p95_ms": 1.1346207999999984,
-            "serial_items_per_s": 1549.8975985756417,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.7690596512561212,
-              0.838274082202966
-            ],
-            "confidence_at_least_90pct": {
-              "n": 273,
-              "coverage": 0.546,
-              "accuracy": 0.9743589743589743
-            }
-          },
-          "gzip-knn": {
-            "family": "gzip-knn",
-            "config": {
-              "k": 1
-            },
-            "validation_accuracy": 0.49833333333333335,
-            "fit_s": 0.011530167001183145,
-            "dictionary_or_sample_bytes": 52739,
-            "temperature": 1.0782500762595328,
-            "accuracy": 0.548,
-            "macro_f1": 0.5242677003523842,
-            "ece": 0.06063191239609045,
-            "nll": 2.6533837260533817,
-            "brier": 0.7007323897060809,
-            "tie_rate": 0.0,
-            "p50_ms": 4.4909795,
-            "p95_ms": 5.387016699999999,
-            "serial_items_per_s": 217.7996094879138,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.5041745952330335,
-              0.5910934413879999
-            ],
-            "confidence_at_least_90pct": {
-              "n": 0,
-              "coverage": 0.0,
-              "accuracy": null
-            }
-          },
-          "tfidf-logistic": {
-            "family": "tfidf-logistic",
-            "config": {},
-            "validation_accuracy": 0.9016666666666666,
-            "fit_s": 6.198137958999723,
-            "dictionary_or_sample_bytes": null,
-            "temperature": 0.5402886714960652,
-            "accuracy": 0.908,
-            "macro_f1": 0.8941185048731366,
-            "ece": 0.02614868859476964,
-            "nll": 0.3168123349805611,
-            "brier": 0.1376886364442629,
-            "tie_rate": 0.0,
-            "p50_ms": 0.383479,
-            "p95_ms": 0.52263335,
-            "serial_items_per_s": 2489.7499360395686,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.8794606777802376,
-              0.9303176334985452
-            ],
-            "confidence_at_least_90pct": {
-              "n": 392,
-              "coverage": 0.784,
-              "accuracy": 0.9897959183673469
-            }
-          }
-        }
-      },
-      "sst2": {
-        "dataset": {
-          "repo": "stanfordnlp/sst2",
-          "hf_fingerprints": {
-            "train": "e7e6610fdbb3c277",
-            "validation": "c1ddc6497ec97f98",
-            "test": "14a60218fd61ec38"
-          },
-          "train_n": 10000,
-          "validation_n": 600,
-          "test_n": 500,
-          "removed_duplicates_or_eval_overlaps": 371,
-          "hashes": {
-            "train": "bb18ecae29bb9ea84b25405f6a0428096951a9a69efac09bb100c9facea1869d",
-            "validation": "fafee5f02c9c933efd0edf3c5c0b1621b8522669dfaca255450c809c6a815590",
-            "test": "380747f6d5a4f05ce8516c891e34d9dd29500beace1c0937b524634411bb287e"
-          },
-          "labels": [
-            "negative",
-            "positive"
-          ]
-        },
-        "validation_sweep": [
-          {
-            "family": "deflate",
-            "config": {
-              "codec": "deflate",
-              "size": 4096
-            },
-            "validation_accuracy": 0.5966666666666667,
-            "fit_s": 0.0017086660009226762,
-            "dictionary_or_sample_bytes": 8192
-          },
-          {
-            "family": "deflate",
-            "config": {
-              "codec": "deflate",
-              "size": 32768
-            },
-            "validation_accuracy": 0.6766666666666666,
-            "fit_s": 0.0017910410024342127,
-            "dictionary_or_sample_bytes": 65536
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "trained": true
-            },
-            "validation_accuracy": 0.5166666666666667,
-            "fit_s": 0.027700541002559476,
-            "dictionary_or_sample_bytes": 2048
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "trained": false
-            },
-            "validation_accuracy": 0.5066666666666667,
-            "fit_s": 0.0017736669979058206,
-            "dictionary_or_sample_bytes": 2048
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "trained": true
-            },
-            "validation_accuracy": 0.5833333333333334,
-            "fit_s": 0.04100783300236799,
-            "dictionary_or_sample_bytes": 8192
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "trained": false
-            },
-            "validation_accuracy": 0.5483333333333333,
-            "fit_s": 0.002048333000857383,
-            "dictionary_or_sample_bytes": 8192
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16384,
-              "trained": true
-            },
-            "validation_accuracy": 0.6283333333333333,
-            "fit_s": 0.038298833002045285,
-            "dictionary_or_sample_bytes": 32768
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16384,
-              "trained": false
-            },
-            "validation_accuracy": 0.5933333333333334,
-            "fit_s": 0.0018218339973827824,
-            "dictionary_or_sample_bytes": 32768
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 65536,
-              "level": 9
-            },
-            "validation_accuracy": 0.6433333333333333,
-            "fit_s": 0.046571332997700665,
-            "dictionary_or_sample_bytes": 131072
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 262144,
-              "level": 9
-            },
-            "validation_accuracy": 0.7066666666666667,
-            "fit_s": 0.08784708299936028,
-            "dictionary_or_sample_bytes": 437677
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "level": 9,
-              "trained": false
-            },
-            "validation_accuracy": 0.7283333333333334,
-            "fit_s": 0.0018446670001139864,
-            "dictionary_or_sample_bytes": 536488
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "shards": 4
-            },
-            "validation_accuracy": 0.5583333333333333,
-            "fit_s": 0.029137124998669606,
-            "dictionary_or_sample_bytes": 8192
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "shards": 8
-            },
-            "validation_accuracy": 0.58,
-            "fit_s": 0.031216542003676295,
-            "dictionary_or_sample_bytes": 16384
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "shards": 4
-            },
-            "validation_accuracy": 0.6216666666666667,
-            "fit_s": 0.0438428749985178,
-            "dictionary_or_sample_bytes": 32768
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "shards": 8
-            },
-            "validation_accuracy": 0.6416666666666667,
-            "fit_s": 0.04907000000093831,
-            "dictionary_or_sample_bytes": 65536
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "min"
-            },
-            "validation_accuracy": 0.725,
-            "fit_s": 0.00196666600095341,
-            "dictionary_or_sample_bytes": 536484
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.7533333333333333,
-            "fit_s": 0.001909541002532933,
-            "dictionary_or_sample_bytes": 536484
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "min"
-            },
-            "validation_accuracy": 0.7366666666666667,
-            "fit_s": 0.0017700829994282685,
-            "dictionary_or_sample_bytes": 536480
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.7616666666666667,
-            "fit_s": 0.001830375003919471,
-            "dictionary_or_sample_bytes": 536480
-          },
-          {
-            "family": "gzip-knn",
-            "config": {
-              "k": 1
-            },
-            "validation_accuracy": 0.5733333333333334,
-            "fit_s": 0.005904542005737312,
-            "dictionary_or_sample_bytes": 52471
-          },
-          {
-            "family": "gzip-knn",
-            "config": {
-              "k": 5
-            },
-            "validation_accuracy": 0.595,
-            "fit_s": 0.006380707993230317,
-            "dictionary_or_sample_bytes": 52471
-          },
-          {
-            "family": "tfidf-logistic",
-            "config": {},
-            "validation_accuracy": 0.84,
-            "fit_s": 0.5434237500012387,
-            "dictionary_or_sample_bytes": null
-          }
-        ],
-        "selected": {
-          "deflate": {
-            "family": "deflate",
-            "config": {
-              "codec": "deflate",
-              "size": 32768
-            },
-            "validation_accuracy": 0.6766666666666666,
-            "fit_s": 0.0017910410024342127,
-            "dictionary_or_sample_bytes": 65536,
-            "temperature": 6.304134293587772,
-            "accuracy": 0.656,
-            "macro_f1": 0.655332724153962,
-            "ece": 0.02910858412013078,
-            "nll": 0.6184185625087876,
-            "brier": 0.43000688230669276,
-            "tie_rate": 0.074,
-            "p50_ms": 0.0276875,
-            "p95_ms": 0.036502099999999996,
-            "serial_items_per_s": 35514.841972093716,
-            "n": 500,
-            "auroc": 0.7180215804303278,
-            "accuracy_wilson_95": [
-              0.6133133706304251,
-              0.6963077483879331
-            ],
-            "confidence_at_least_90pct": {
-              "n": 11,
-              "coverage": 0.022,
-              "accuracy": 0.9090909090909091
-            }
-          },
-          "zstd": {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "level": 9,
-              "trained": false
-            },
-            "validation_accuracy": 0.7283333333333334,
-            "fit_s": 0.0018446670001139864,
-            "dictionary_or_sample_bytes": 536488,
-            "temperature": 10.790252743818396,
-            "accuracy": 0.65,
-            "macro_f1": 0.6499985999944,
-            "ece": 0.03593030036559715,
-            "nll": 0.618717418902004,
-            "brier": 0.43029277027337204,
-            "tie_rate": 0.086,
-            "p50_ms": 0.013625,
-            "p95_ms": 0.022675299999999992,
-            "serial_items_per_s": 69523.71113120501,
-            "n": 500,
-            "auroc": 0.7180696080942622,
-            "accuracy_wilson_95": [
-              0.6071920689703412,
-              0.6905205454703878
-            ],
-            "confidence_at_least_90pct": {
-              "n": 2,
-              "coverage": 0.004,
-              "accuracy": 1.0
-            }
-          },
-          "zstd-mixture": {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.7616666666666667,
-            "fit_s": 0.001830375003919471,
-            "dictionary_or_sample_bytes": 536480,
-            "temperature": 2.925419034377105,
-            "accuracy": 0.718,
-            "macro_f1": 0.7175922031413361,
-            "ece": 0.04554804951123206,
-            "nll": 0.5547712132787098,
-            "brier": 0.37369763190577826,
-            "tie_rate": 0.028,
-            "p50_ms": 0.041292,
-            "p95_ms": 0.06892435,
-            "serial_items_per_s": 22944.298676765586,
-            "n": 500,
-            "auroc": 0.7985319544057375,
-            "accuracy_wilson_95": [
-              0.6770114417638189,
-              0.7556642245567071
-            ],
-            "confidence_at_least_90pct": {
-              "n": 72,
-              "coverage": 0.144,
-              "accuracy": 0.9166666666666666
-            }
-          },
-          "gzip-knn": {
-            "family": "gzip-knn",
-            "config": {
-              "k": 5
-            },
-            "validation_accuracy": 0.595,
-            "fit_s": 0.006380707993230317,
-            "dictionary_or_sample_bytes": 52471,
-            "temperature": 5.007191993770399,
-            "accuracy": 0.548,
-            "macro_f1": 0.5474134478283856,
-            "ece": 0.013168762839443831,
-            "nll": 0.6913485232561188,
-            "brier": 0.4975115063432846,
-            "tie_rate": 0.0,
-            "p50_ms": 5.481375,
-            "p95_ms": 6.29362325,
-            "serial_items_per_s": 182.7036415967676,
-            "n": 500,
-            "auroc": 0.5551037397540983,
-            "accuracy_wilson_95": [
-              0.5041745952330335,
-              0.5910934413879999
-            ],
-            "confidence_at_least_90pct": {
-              "n": 0,
-              "coverage": 0.0,
-              "accuracy": null
-            }
-          },
-          "tfidf-logistic": {
-            "family": "tfidf-logistic",
-            "config": {},
-            "validation_accuracy": 0.84,
-            "fit_s": 0.5434237500012387,
-            "dictionary_or_sample_bytes": null,
-            "temperature": 0.7345152395906567,
-            "accuracy": 0.796,
-            "macro_f1": 0.7959706197692469,
-            "ece": 0.03953913908774128,
-            "nll": 0.41708729362573255,
-            "brier": 0.26258361413544473,
-            "tie_rate": 0.0,
-            "p50_ms": 0.412729,
-            "p95_ms": 0.5067076999999999,
-            "serial_items_per_s": 2391.932066718238,
-            "n": 500,
-            "auroc": 0.8961321721311475,
-            "accuracy_wilson_95": [
-              0.7584839354593685,
-              0.8290022903703368
-            ],
-            "confidence_at_least_90pct": {
-              "n": 243,
-              "coverage": 0.486,
-              "accuracy": 0.9382716049382716
-            }
-          }
-        }
-      },
-      "emotion": {
-        "dataset": {
-          "repo": "dair-ai/emotion",
-          "hf_fingerprints": {
-            "train": "27c5e4f1e47ffa42",
-            "validation": "5cae643edfd5aaf3",
-            "test": "bd45cf0a5e0b9690"
-          },
-          "train_n": 10000,
-          "validation_n": 600,
-          "test_n": 500,
-          "removed_duplicates_or_eval_overlaps": 42,
-          "hashes": {
-            "train": "5bcc8e86df660889cb24bfd3a261f5528d8fb3b142f4693c73fd739338b9cb86",
-            "validation": "bda5d620b7ea488ba509f07b3591f772feb319b8a396d7812b26f89421143e7a",
-            "test": "62fe848e5c2808aed7e0077efd032834d274b6eb6e7d323e76be32ef97b1feea"
-          },
-          "labels": [
-            "sadness",
-            "joy",
-            "love",
-            "anger",
-            "fear",
-            "surprise"
-          ]
-        },
-        "validation_sweep": [
-          {
-            "family": "deflate",
-            "config": {
-              "codec": "deflate",
-              "size": 4096
-            },
-            "validation_accuracy": 0.32666666666666666,
-            "fit_s": 0.002004999994824175,
-            "dictionary_or_sample_bytes": 24576
-          },
-          {
-            "family": "deflate",
-            "config": {
-              "codec": "deflate",
-              "size": 32768
-            },
-            "validation_accuracy": 0.43833333333333335,
-            "fit_s": 0.00225687499914784,
-            "dictionary_or_sample_bytes": 196608
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "trained": true
-            },
-            "validation_accuracy": 0.33,
-            "fit_s": 0.04342029100371292,
-            "dictionary_or_sample_bytes": 6144
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "trained": false
-            },
-            "validation_accuracy": 0.235,
-            "fit_s": 0.0020858330026385374,
-            "dictionary_or_sample_bytes": 6144
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "trained": true
-            },
-            "validation_accuracy": 0.445,
-            "fit_s": 0.05890566699963529,
-            "dictionary_or_sample_bytes": 24576
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "trained": false
-            },
-            "validation_accuracy": 0.275,
-            "fit_s": 0.002086583997879643,
-            "dictionary_or_sample_bytes": 24576
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16384,
-              "trained": true
-            },
-            "validation_accuracy": 0.44666666666666666,
-            "fit_s": 0.058133709004323464,
-            "dictionary_or_sample_bytes": 98304
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16384,
-              "trained": false
-            },
-            "validation_accuracy": 0.37666666666666665,
-            "fit_s": 0.0023508340018452145,
-            "dictionary_or_sample_bytes": 98304
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 65536,
-              "level": 9
-            },
-            "validation_accuracy": 0.4033333333333333,
-            "fit_s": 0.11329087500052992,
-            "dictionary_or_sample_bytes": 349551
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 262144,
-              "level": 9
-            },
-            "validation_accuracy": 0.43666666666666665,
-            "fit_s": 0.15255850000539795,
-            "dictionary_or_sample_bytes": 856643
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "level": 9,
-              "trained": false
-            },
-            "validation_accuracy": 0.44166666666666665,
-            "fit_s": 0.0022252499984460883,
-            "dictionary_or_sample_bytes": 979245
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "shards": 4
-            },
-            "validation_accuracy": 0.3616666666666667,
-            "fit_s": 0.0483227920049103,
-            "dictionary_or_sample_bytes": 24576
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "shards": 8
-            },
-            "validation_accuracy": 0.3883333333333333,
-            "fit_s": 0.0537706250033807,
-            "dictionary_or_sample_bytes": 49152
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "shards": 4
-            },
-            "validation_accuracy": 0.515,
-            "fit_s": 0.07201379199977964,
-            "dictionary_or_sample_bytes": 98304
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "shards": 8
-            },
-            "validation_accuracy": 0.48,
-            "fit_s": 0.08771720799995819,
-            "dictionary_or_sample_bytes": 195367
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "min"
-            },
-            "validation_accuracy": 0.5066666666666667,
-            "fit_s": 0.002359584002988413,
-            "dictionary_or_sample_bytes": 979233
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.535,
-            "fit_s": 0.0026667080019251443,
-            "dictionary_or_sample_bytes": 979233
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "min"
-            },
-            "validation_accuracy": 0.5033333333333333,
-            "fit_s": 0.002307417002157308,
-            "dictionary_or_sample_bytes": 979221
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.5533333333333333,
-            "fit_s": 0.00223658300092211,
-            "dictionary_or_sample_bytes": 979221
-          },
-          {
-            "family": "gzip-knn",
-            "config": {
-              "k": 1
-            },
-            "validation_accuracy": 0.25833333333333336,
-            "fit_s": 0.007107000004907604,
-            "dictionary_or_sample_bytes": 98188
-          },
-          {
-            "family": "gzip-knn",
-            "config": {
-              "k": 5
-            },
-            "validation_accuracy": 0.24166666666666667,
-            "fit_s": 0.006843499999376945,
-            "dictionary_or_sample_bytes": 98188
-          },
-          {
-            "family": "tfidf-logistic",
-            "config": {},
-            "validation_accuracy": 0.8416666666666667,
-            "fit_s": 3.909456957997463,
-            "dictionary_or_sample_bytes": null
-          }
-        ],
-        "selected": {
-          "deflate": {
-            "family": "deflate",
-            "config": {
-              "codec": "deflate",
-              "size": 32768
-            },
-            "validation_accuracy": 0.43833333333333335,
-            "fit_s": 0.00225687499914784,
-            "dictionary_or_sample_bytes": 196608,
-            "temperature": 3.410951603860984,
-            "accuracy": 0.406,
-            "macro_f1": 0.3684907264569877,
-            "ece": 0.03192085479609739,
-            "nll": 1.486202601279427,
-            "brier": 0.7179354684289194,
-            "tie_rate": 0.168,
-            "p50_ms": 0.0647915,
-            "p95_ms": 0.10161844999999997,
-            "serial_items_per_s": 14590.575182943197,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.3638296860034575,
-              0.44960374228035244
-            ],
-            "confidence_at_least_90pct": {
-              "n": 1,
-              "coverage": 0.002,
-              "accuracy": 1.0
-            }
-          },
-          "zstd": {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16384,
-              "trained": true
-            },
-            "validation_accuracy": 0.44666666666666666,
-            "fit_s": 0.058133709004323464,
-            "dictionary_or_sample_bytes": 98304,
-            "temperature": 2.709220452261488,
-            "accuracy": 0.454,
-            "macro_f1": 0.4018783545119495,
-            "ece": 0.05846382416841619,
-            "nll": 1.463362818214318,
-            "brier": 0.698117185244948,
-            "tie_rate": 0.214,
-            "p50_ms": 0.0115,
-            "p95_ms": 0.017131249999999997,
-            "serial_items_per_s": 81165.0627909159,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.4108749466266584,
-              0.49782651827818475
-            ],
-            "confidence_at_least_90pct": {
-              "n": 2,
-              "coverage": 0.004,
-              "accuracy": 0.5
-            }
-          },
-          "zstd-mixture": {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.5533333333333333,
-            "fit_s": 0.00223658300092211,
-            "dictionary_or_sample_bytes": 979221,
-            "temperature": 2.925419034377105,
-            "accuracy": 0.562,
-            "macro_f1": 0.46353991429511865,
-            "ece": 0.08457185931239918,
-            "nll": 1.2537522096435618,
-            "brier": 0.597552891264581,
-            "tie_rate": 0.038,
-            "p50_ms": 0.103604,
-            "p95_ms": 0.2117881499999997,
-            "serial_items_per_s": 8727.203975946568,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.5182021184950548,
-              0.6048524288071132
-            ],
-            "confidence_at_least_90pct": {
-              "n": 22,
-              "coverage": 0.044,
-              "accuracy": 0.8181818181818182
-            }
-          },
-          "gzip-knn": {
-            "family": "gzip-knn",
-            "config": {
-              "k": 1
-            },
-            "validation_accuracy": 0.25833333333333336,
-            "fit_s": 0.007107000004907604,
-            "dictionary_or_sample_bytes": 98188,
-            "temperature": 8.570386453309192,
-            "accuracy": 0.298,
-            "macro_f1": 0.2781904296547765,
-            "ece": 0.0427759626994248,
-            "nll": 1.743637872594331,
-            "brier": 0.814830939581835,
-            "tie_rate": 0.0,
-            "p50_ms": 5.6842915000000005,
-            "p95_ms": 6.971593449999999,
-            "serial_items_per_s": 171.43746654479273,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.25957253815100145,
-              0.3395078077354835
-            ],
-            "confidence_at_least_90pct": {
-              "n": 0,
-              "coverage": 0.0,
-              "accuracy": null
-            }
-          },
-          "tfidf-logistic": {
-            "family": "tfidf-logistic",
-            "config": {},
-            "validation_accuracy": 0.8416666666666667,
-            "fit_s": 3.909456957997463,
-            "dictionary_or_sample_bytes": null,
-            "temperature": 0.5834042638846706,
-            "accuracy": 0.852,
-            "macro_f1": 0.7837710838613873,
-            "ece": 0.030185522477487146,
-            "nll": 0.4129957238398832,
-            "brier": 0.22078743602615492,
-            "tie_rate": 0.0,
-            "p50_ms": 0.426187,
-            "p95_ms": 0.5887381499999998,
-            "serial_items_per_s": 2240.786993074041,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.818193200072725,
-              0.8804390684815191
-            ],
-            "confidence_at_least_90pct": {
-              "n": 322,
-              "coverage": 0.644,
-              "accuracy": 0.9472049689440993
-            }
-          }
-        }
-      }
-    },
-    "jevbench": {
-      "easy": {
-        "n": 48,
-        "sha256": "231df3c2c8e88a1a8c137ebe85de96ba70fabd330849098ac7b3c52c70b7172b",
-        "accuracy": 0.4583333333333333,
-        "uniform_chance": 0.284375,
-        "p50_ms": 0.074271,
-        "families": {
-          "extraction": {
-            "n": 12,
-            "accuracy": 0.4166666666666667
-          },
-          "fact": {
-            "n": 12,
-            "accuracy": 0.5833333333333334
-          },
-          "intent": {
-            "n": 12,
-            "accuracy": 0.4166666666666667
-          },
-          "tool_selection": {
-            "n": 12,
-            "accuracy": 0.4166666666666667
-          }
-        }
-      },
-      "original": {
-        "n": 72,
-        "sha256": "5c2414edb3006b8bfcb70fda433f0f9ca015759433849f8d3104328a1f7c4180",
-        "accuracy": 0.3611111111111111,
-        "uniform_chance": 0.3111111111111111,
-        "p50_ms": 0.0727085,
-        "families": {
-          "adequacy": {
-            "n": 12,
-            "accuracy": 0.4166666666666667
-          },
-          "extraction": {
-            "n": 12,
-            "accuracy": 0.3333333333333333
-          },
-          "intent": {
-            "n": 12,
-            "accuracy": 0.16666666666666666
-          },
-          "ordinal": {
-            "n": 12,
-            "accuracy": 0.4166666666666667
-          },
-          "policy": {
-            "n": 12,
-            "accuracy": 0.5
-          },
-          "routing": {
-            "n": 12,
-            "accuracy": 0.3333333333333333
-          }
-        }
-      },
-      "hard": {
-        "n": 111,
-        "sha256": "89e9e6becb33ed88c1de7d42dcc87531b2fb64cfaef4e1986faf7c37b3f80ebb",
-        "accuracy": 0.42342342342342343,
-        "uniform_chance": 0.3361861861861862,
-        "p50_ms": 0.183583,
-        "families": {
-          "adversarial": {
-            "n": 6,
-            "accuracy": 0.6666666666666666
-          },
-          "ambiguous": {
-            "n": 7,
-            "accuracy": 0.42857142857142855
-          },
-          "judge_hard": {
-            "n": 17,
-            "accuracy": 0.4117647058823529
-          },
-          "long_policy": {
-            "n": 19,
-            "accuracy": 0.21052631578947367
-          },
-          "multi_hop": {
-            "n": 18,
-            "accuracy": 0.2222222222222222
-          },
-          "probability": {
-            "n": 10,
-            "accuracy": 0.5
-          },
-          "routing_hard": {
-            "n": 5,
-            "accuracy": 1.0
-          },
-          "temporal_numeric": {
-            "n": 15,
-            "accuracy": 0.3333333333333333
-          },
-          "tradeoff": {
-            "n": 6,
-            "accuracy": 1.0
-          },
-          "trap": {
-            "n": 8,
-            "accuracy": 0.5
-          }
-        }
-      },
-      "source_commit": "2fa63fa3226cb369795525ed011800f57dcbd894",
-      "method": "Zero-shot symmetric gzip NCD against label and rubric; no training or calibration. Public diagnostic, not an official score."
-    }
-  },
-  "large_training_run": {
-    "manifest": {
-      "args": {
-        "datasets": [
-          "agnews",
-          "sst2"
-        ],
-        "families": [
-          "zstd",
-          "zstd-mixture",
-          "tfidf-logistic"
-        ],
-        "seed": 0,
-        "n": 500,
-        "train_cap": 120000,
-        "validation_n": 600,
-        "out": "eval/data/compression-large",
-        "jevbench": null
-      },
-      "platform": "macOS-26.5.1-arm64-arm-64bit",
-      "processor": "arm",
-      "python": "3.12.14",
-      "zlib": "1.2.12",
-      "thread_environment": {
-        "OMP_NUM_THREADS": "1",
-        "OPENBLAS_NUM_THREADS": "1"
-      },
-      "timing": "Serial local warm inference: normalization, compression/features, probabilities, tie-breaking. Excludes training, network and data loading.",
-      "versions": {
-        "datasets": "5.0.1",
-        "numpy": "2.5.3",
-        "scipy": "1.18.1",
-        "scikit-learn": "1.9.1",
-        "zstandard": "0.25.0"
-      },
-      "script_sha256": "5342902063b3071509c854905aa18c2410c8202e74c1fcad2ab76ae89491a2f4"
-    },
-    "datasets": {
-      "agnews": {
-        "dataset": {
-          "repo": "fancyzhx/ag_news",
-          "hf_fingerprints": {
-            "train": "2bc1e2be15d7f2ee",
-            "test": "0ebcc57b6aea2fc9"
-          },
-          "train_n": 119239,
-          "validation_n": 600,
-          "test_n": 500,
-          "removed_duplicates_or_eval_overlaps": 161,
-          "hashes": {
-            "train": "d1fb37dc340dc5ed27d4162ef9a98b4e83a80b5dc958e3e4e97101aa0f9e62ed",
-            "validation": "89cea8648589dcdb7aacc5e04f4efcd1b6058e5946f346d672c492fe8b03d403",
-            "test": "76c1f71538ffa2609b9dbb9bbf842e60d19b79470d1f4d0cc554d88d95ae47ed"
-          },
-          "labels": [
-            "World",
-            "Sports",
-            "Business",
-            "Sci/Tech"
-          ]
-        },
-        "validation_sweep": [
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "trained": true
-            },
-            "validation_accuracy": 0.655,
-            "fit_s": 1.095241957998951,
-            "dictionary_or_sample_bytes": 4096
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "trained": false
-            },
-            "validation_accuracy": 0.38333333333333336,
-            "fit_s": 0.05557016700186068,
-            "dictionary_or_sample_bytes": 4096
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "trained": true
-            },
-            "validation_accuracy": 0.7733333333333333,
-            "fit_s": 1.6625714999972843,
-            "dictionary_or_sample_bytes": 16384
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "trained": false
-            },
-            "validation_accuracy": 0.5033333333333333,
-            "fit_s": 0.0551289579962031,
-            "dictionary_or_sample_bytes": 16384
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16384,
-              "trained": true
-            },
-            "validation_accuracy": 0.805,
-            "fit_s": 1.376491791997978,
-            "dictionary_or_sample_bytes": 65536
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16384,
-              "trained": false
-            },
-            "validation_accuracy": 0.6483333333333333,
-            "fit_s": 0.05820737500471296,
-            "dictionary_or_sample_bytes": 65536
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 65536,
-              "level": 9
-            },
-            "validation_accuracy": 0.795,
-            "fit_s": 1.1389680420033983,
-            "dictionary_or_sample_bytes": 262144
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 262144,
-              "level": 9
-            },
-            "validation_accuracy": 0.7983333333333333,
-            "fit_s": 1.0690593339968473,
-            "dictionary_or_sample_bytes": 1048576
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "level": 9,
-              "trained": false
-            },
-            "validation_accuracy": 0.855,
-            "fit_s": 0.05941937500028871,
-            "dictionary_or_sample_bytes": 28247565
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "shards": 4
-            },
-            "validation_accuracy": 0.7283333333333334,
-            "fit_s": 1.1204033750036615,
-            "dictionary_or_sample_bytes": 16384
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "shards": 8
-            },
-            "validation_accuracy": 0.73,
-            "fit_s": 1.1286926250031684,
-            "dictionary_or_sample_bytes": 32768
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "shards": 4
-            },
-            "validation_accuracy": 0.8066666666666666,
-            "fit_s": 1.700393332997919,
-            "dictionary_or_sample_bytes": 65536
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "shards": 8
-            },
-            "validation_accuracy": 0.8166666666666667,
-            "fit_s": 1.7476613329999964,
-            "dictionary_or_sample_bytes": 131072
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "min"
-            },
-            "validation_accuracy": 0.8766666666666667,
-            "fit_s": 0.0598740830027964,
-            "dictionary_or_sample_bytes": 28247557
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.8983333333333333,
-            "fit_s": 0.057473291002679616,
-            "dictionary_or_sample_bytes": 28247557
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "min"
-            },
-            "validation_accuracy": 0.8966666666666666,
-            "fit_s": 0.058699958004581276,
-            "dictionary_or_sample_bytes": 28247549
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.9116666666666666,
-            "fit_s": 0.060205792004126124,
-            "dictionary_or_sample_bytes": 28247549
-          },
-          {
-            "family": "tfidf-logistic",
-            "config": {},
-            "validation_accuracy": 0.9366666666666666,
-            "fit_s": 72.26334416699683,
-            "dictionary_or_sample_bytes": null
-          }
-        ],
-        "selected": {
-          "zstd": {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "level": 9,
-              "trained": false
-            },
-            "validation_accuracy": 0.855,
-            "fit_s": 0.05941937500028871,
-            "dictionary_or_sample_bytes": 28247565,
-            "temperature": 5.838236970939662,
-            "accuracy": 0.856,
-            "macro_f1": 0.8578350680660315,
-            "ece": 0.07148579683576123,
-            "nll": 0.573052244855724,
-            "brier": 0.2274947566483206,
-            "tie_rate": 0.02,
-            "p50_ms": 0.0481665,
-            "p95_ms": 0.09512054999999997,
-            "serial_items_per_s": 16460.801843188412,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.822508879125479,
-              0.8840623924805179
-            ],
-            "confidence_at_least_90pct": {
-              "n": 246,
-              "coverage": 0.492,
-              "accuracy": 0.9471544715447154
-            }
-          },
-          "zstd-mixture": {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.9116666666666666,
-            "fit_s": 0.060205792004126124,
-            "dictionary_or_sample_bytes": 28247549,
-            "temperature": 3.683149054535093,
-            "accuracy": 0.9,
-            "macro_f1": 0.8997427715689791,
-            "ece": 0.05852817346189611,
-            "nll": 0.4590178473955788,
-            "brier": 0.17604696824424826,
-            "tie_rate": 0.004,
-            "p50_ms": 0.24593700000000002,
-            "p95_ms": 0.4126940499999994,
-            "serial_items_per_s": 3567.4229415418463,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.8705774917999974,
-              0.9233228133752799
-            ],
-            "confidence_at_least_90pct": {
-              "n": 360,
-              "coverage": 0.72,
-              "accuracy": 0.9472222222222222
-            }
-          },
-          "tfidf-logistic": {
-            "family": "tfidf-logistic",
-            "config": {},
-            "validation_accuracy": 0.9366666666666666,
-            "fit_s": 72.26334416699683,
-            "dictionary_or_sample_bytes": null,
-            "temperature": 0.9247663594340342,
-            "accuracy": 0.912,
-            "macro_f1": 0.9127427169277376,
-            "ece": 0.02759829499776148,
-            "nll": 0.28636336613317387,
-            "brier": 0.14006859532256022,
-            "tie_rate": 0.0,
-            "p50_ms": 0.593167,
-            "p95_ms": 0.7779229499999999,
-            "serial_items_per_s": 1645.5464238809047,
-            "n": 500,
-            "accuracy_wilson_95": [
-              0.8839229514479334,
-              0.9337943628826022
-            ],
-            "confidence_at_least_90pct": {
-              "n": 397,
-              "coverage": 0.794,
-              "accuracy": 0.9697732997481109
-            }
-          }
-        }
-      },
-      "sst2": {
-        "dataset": {
-          "repo": "stanfordnlp/sst2",
-          "hf_fingerprints": {
-            "train": "e7e6610fdbb3c277",
-            "validation": "c1ddc6497ec97f98",
-            "test": "14a60218fd61ec38"
-          },
-          "train_n": 66378,
-          "validation_n": 600,
-          "test_n": 500,
-          "removed_duplicates_or_eval_overlaps": 371,
-          "hashes": {
-            "train": "488b7c60c5677461e842fc5ecd32774c1ddbdd5878c8e6c793c7430eeb5720e1",
-            "validation": "745798b437dae118d5c5f3498399b935f25a45040fb07d6423a8d18964f91fd7",
-            "test": "380747f6d5a4f05ce8516c891e34d9dd29500beace1c0937b524634411bb287e"
-          },
-          "labels": [
-            "negative",
-            "positive"
-          ]
-        },
-        "validation_sweep": [
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "trained": true
-            },
-            "validation_accuracy": 0.57,
-            "fit_s": 0.1833836249934393,
-            "dictionary_or_sample_bytes": 2048
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "trained": false
-            },
-            "validation_accuracy": 0.54,
-            "fit_s": 0.01816158300061943,
-            "dictionary_or_sample_bytes": 2048
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "trained": true
-            },
-            "validation_accuracy": 0.61,
-            "fit_s": 0.2721342500008177,
-            "dictionary_or_sample_bytes": 8192
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "trained": false
-            },
-            "validation_accuracy": 0.56,
-            "fit_s": 0.019221125003241468,
-            "dictionary_or_sample_bytes": 8192
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16384,
-              "trained": true
-            },
-            "validation_accuracy": 0.6816666666666666,
-            "fit_s": 0.2559416250005597,
-            "dictionary_or_sample_bytes": 32768
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16384,
-              "trained": false
-            },
-            "validation_accuracy": 0.6333333333333333,
-            "fit_s": 0.018674541999644134,
-            "dictionary_or_sample_bytes": 32768
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 65536,
-              "level": 9
-            },
-            "validation_accuracy": 0.6616666666666666,
-            "fit_s": 0.2445404169993708,
-            "dictionary_or_sample_bytes": 131072
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 262144,
-              "level": 9
-            },
-            "validation_accuracy": 0.7466666666666667,
-            "fit_s": 0.2638120420015184,
-            "dictionary_or_sample_bytes": 524288
-          },
-          {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "level": 9,
-              "trained": false
-            },
-            "validation_accuracy": 0.7816666666666666,
-            "fit_s": 0.018079166002280544,
-            "dictionary_or_sample_bytes": 3567328
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "shards": 4
-            },
-            "validation_accuracy": 0.6133333333333333,
-            "fit_s": 0.1946094590020948,
-            "dictionary_or_sample_bytes": 8192
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 1024,
-              "shards": 8
-            },
-            "validation_accuracy": 0.5866666666666667,
-            "fit_s": 0.19280641600198578,
-            "dictionary_or_sample_bytes": 16384
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "shards": 4
-            },
-            "validation_accuracy": 0.6583333333333333,
-            "fit_s": 0.2784915840020403,
-            "dictionary_or_sample_bytes": 32768
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 4096,
-              "shards": 8
-            },
-            "validation_accuracy": 0.66,
-            "fit_s": 0.28512387500086334,
-            "dictionary_or_sample_bytes": 65536
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "min"
-            },
-            "validation_accuracy": 0.7916666666666666,
-            "fit_s": 0.01775899999483954,
-            "dictionary_or_sample_bytes": 3567324
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.79,
-            "fit_s": 0.0159552499972051,
-            "dictionary_or_sample_bytes": 3567324
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "min"
-            },
-            "validation_accuracy": 0.7883333333333333,
-            "fit_s": 0.017087957996409386,
-            "dictionary_or_sample_bytes": 3567320
-          },
-          {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 5,
-              "trained": false,
-              "level": 9,
-              "reduction": "mean"
-            },
-            "validation_accuracy": 0.79,
-            "fit_s": 0.016865915997186676,
-            "dictionary_or_sample_bytes": 3567320
-          },
-          {
-            "family": "tfidf-logistic",
-            "config": {},
-            "validation_accuracy": 0.915,
-            "fit_s": 3.577201125001011,
-            "dictionary_or_sample_bytes": null
-          }
-        ],
-        "selected": {
-          "zstd": {
-            "family": "zstd",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "level": 9,
-              "trained": false
-            },
-            "validation_accuracy": 0.7816666666666666,
-            "fit_s": 0.018079166002280544,
-            "dictionary_or_sample_bytes": 3567328,
-            "temperature": 8.570386453309192,
-            "accuracy": 0.662,
-            "macro_f1": 0.6608591301137025,
-            "ece": 0.04250272619027978,
-            "nll": 0.6126490882855867,
-            "brier": 0.4222055784678397,
-            "tie_rate": 0.09,
-            "p50_ms": 0.015396,
-            "p95_ms": 0.025002049999999998,
-            "serial_items_per_s": 62596.52383976092,
-            "n": 500,
-            "auroc": 0.7419633709016394,
-            "accuracy_wilson_95": [
-              0.6194419387868491,
-              0.7020876848091384
-            ],
-            "confidence_at_least_90pct": {
-              "n": 6,
-              "coverage": 0.012,
-              "accuracy": 0.6666666666666666
-            }
-          },
-          "zstd-mixture": {
-            "family": "zstd-mixture",
-            "config": {
-              "codec": "zstd",
-              "size": 16777216,
-              "shards": 3,
-              "trained": false,
-              "level": 9,
-              "reduction": "min"
-            },
-            "validation_accuracy": 0.7916666666666666,
-            "fit_s": 0.01775899999483954,
-            "dictionary_or_sample_bytes": 3567324,
-            "temperature": 7.3504331266672,
-            "accuracy": 0.7,
-            "macro_f1": 0.6996106954613179,
-            "ece": 0.059913157093607276,
-            "nll": 0.6007650774839849,
-            "brier": 0.4103150298882698,
-            "tie_rate": 0.084,
-            "p50_ms": 0.034,
-            "p95_ms": 0.06059134999999999,
-            "serial_items_per_s": 28267.394171399006,
-            "n": 500,
-            "auroc": 0.7595174820696722,
-            "accuracy_wilson_95": [
-              0.6584314090816449,
-              0.7385187435059938
-            ],
-            "confidence_at_least_90pct": {
-              "n": 13,
-              "coverage": 0.026,
-              "accuracy": 0.8461538461538461
-            }
-          },
-          "tfidf-logistic": {
-            "family": "tfidf-logistic",
-            "config": {},
-            "validation_accuracy": 0.915,
-            "fit_s": 3.577201125001011,
-            "dictionary_or_sample_bytes": null,
-            "temperature": 0.7345152395906567,
-            "accuracy": 0.85,
-            "macro_f1": 0.8499513842484965,
-            "ece": 0.038211722573768335,
-            "nll": 0.3873007503417493,
-            "brier": 0.22026970652824607,
-            "tie_rate": 0.0,
-            "p50_ms": 0.43174999999999997,
-            "p95_ms": 0.5471291999999999,
-            "serial_items_per_s": 2263.669087689429,
-            "n": 500,
-            "auroc": 0.9229956454918032,
-            "accuracy_wilson_95": [
-              0.8160382472597502,
-              0.8786245197686174
-            ],
-            "confidence_at_least_90pct": {
-              "n": 320,
-              "coverage": 0.64,
-              "accuracy": 0.940625
-            }
-          }
-        }
-      }
-    }
-  },
-  "verification": {
-    "predictions_validated": 10000,
-    "dictionary_and_lexical_predictions_reproduce_exactly": true,
-    "script_hash_matches": true,
-    "jevbench": {
-      "official_label_scoring_matches": true,
-      "option_reversal_invariant": true,
-      "results": [
-        {
-          "tier": "easy",
-          "n": 48,
-          "correct": 22,
-          "accuracy": 0.4583333333333333
-        },
-        {
-          "tier": "hard",
-          "n": 111,
-          "correct": 47,
-          "accuracy": 0.42342342342342343
-        },
-        {
-          "tier": "original",
-          "n": 72,
-          "correct": 26,
-          "accuracy": 0.3611111111111111
-        }
-      ]
-    },
-    "large_training_predictions_validated": 3000,
-    "large_training_uses_identical_test_sample": true
-  }
-}
diff --git a/eval/compression.py b/eval/compression.py
deleted file mode 100644
index 35cad0c..0000000
--- a/eval/compression.py
+++ /dev/null
@@ -1,371 +0,0 @@
-#!/usr/bin/env python3
-# /// script
-# requires-python = ">=3.12"
-# dependencies = ["datasets==5.0.1", "numpy==2.5.3", "scipy==1.18.1", "scikit-learn==1.9.1", "zstandard==0.25.0"]
-# ///
-"""Compression classification experiment. Run with: uv run eval/compression.py.
-
-All tuning uses held-back training rows. Each family gets one final test run.
-Raw predictions, validation sweeps, split hashes and environment are saved.
-"""
-import argparse
-import gzip
-import hashlib
-import importlib.metadata
-import json
-import os
-import platform
-import random
-import subprocess
-import time
-import zlib
-from pathlib import Path
-
-import numpy as np
-import zstandard as zstd
-from datasets import load_dataset
-from scipy.special import softmax
-from sklearn.feature_extraction.text import TfidfVectorizer
-from sklearn.pipeline import FeatureUnion
-from sklearn.linear_model import LogisticRegression
-from sklearn.metrics import accuracy_score, f1_score, log_loss, roc_auc_score
-from sklearn.model_selection import train_test_split
-
-DATASETS = {
-    "agnews": ("fancyzhx/ag_news", None, "text", "test"),
-    "banking77": ("legacy-datasets/banking77", None, "text", "test"),
-    "sst2": ("stanfordnlp/sst2", None, "sentence", "validation"),
-    "emotion": ("dair-ai/emotion", "split", "text", "test"),
-}
-
-
-def normalize(text):
-    return " ".join(text.lower().split())
-
-
-def digest(value):
-    return hashlib.sha256(json.dumps(value, ensure_ascii=False, sort_keys=True).encode()).hexdigest()
-
-
-def load_split(name, seed, n, train_cap, val_n):
-    repo, config, field, test = DATASETS[name]
-    ds = load_dataset(repo, config)
-    labels = ds["train"].features["label"].names
-    heldout = [(normalize(r[field]), int(r["label"])) for r in ds[test]]
-    rng = random.Random(seed)
-    rng.shuffle(heldout)
-    evaluation = heldout[:n]
-    seen = {x for x, _ in heldout}
-    pool = []
-    duplicates = 0
-    for r in ds["train"]:
-        x, y = normalize(r[field]), int(r["label"])
-        if x in seen:
-            duplicates += 1
-            continue
-        seen.add(x)
-        pool.append((x, y))
-    rng.shuffle(pool)
-    pool = pool[:train_cap + val_n]
-    train, val = train_test_split(pool, test_size=val_n, random_state=seed,
-                                  stratify=[r[1] for r in pool])
-    assert not ({r[0] for r in train} & {r[0] for r in val})
-    assert not ({r[0] for r in train + val} & {r[0] for r in heldout})
-    assert set(y for _, y in train) == set(range(len(labels)))
-    meta = {"repo": repo, "hf_fingerprints": {k: v._fingerprint for k, v in ds.items()},
-            "train_n": len(train), "validation_n": len(val), "test_n": len(evaluation),
-            "removed_duplicates_or_eval_overlaps": duplicates,
-            "hashes": {k: digest(v) for k, v in [("train", train), ("validation", val), ("test", evaluation)]},
-            "labels": labels}
-    return train, val, evaluation, meta
-
-
-class Dictionaries:
-    def __init__(self, rows, classes, codec, size, shards=1, seed=42, trained=True, level=3, reduction="min"):
-        self.codec, self.models, self.bytes = codec, [], 0
-        self.reduction = reduction
-        rng = random.Random(seed)
-        for label in range(classes):
-            texts = [x.encode() for x, y in rows if y == label]
-            rng.shuffle(texts)
-            models = []
-            for i in range(shards):
-                samples = texts[i::shards]
-                if not samples:
-                    raise ValueError("empty class shard")
-                if codec == "zstd":
-                    if trained:
-                        # Do not silently change the algorithm if training fails.
-                        dictionary = zstd.train_dictionary(size, samples, threads=0)
-                    else:
-                        dictionary = zstd.ZstdCompressionDict(b"\n".join(samples)[-size:], dict_type=zstd.DICT_TYPE_RAWCONTENT)
-                    self.bytes += len(dictionary.as_bytes())
-                    models.append(zstd.ZstdCompressor(level=level, dict_data=dictionary))
-                else:
-                    dictionary = b"\n".join(samples)[-size:]
-                    self.bytes += len(dictionary)
-                    models.append(zlib.compressobj(level=6, wbits=-15, zdict=dictionary))
-            self.models.append(models)
-
-    def score(self, text):
-        x = text.encode()
-        scores = []
-        for models in self.models:
-            lengths = []
-            for model in models:
-                if self.codec == "zstd":
-                    lengths.append(len(model.compress(x)))
-                else:
-                    c = model.copy()
-                    lengths.append(len(c.compress(x) + c.flush()))
-            scores.append(-float(np.mean(lengths)) if self.reduction == "mean" else -min(lengths))
-        return np.asarray(scores, dtype=float)
-
-
-class GzipKNN:
-    def __init__(self, rows, classes, k=5, cap=1000, seed=42):
-        rng = random.Random(seed)
-        rows = list(rows)
-        rng.shuffle(rows)
-        # Balanced memory avoids a majority-class prior from subsampling.
-        self.samples = []
-        for label in range(classes):
-            self.samples += [(x.encode(), y) for x, y in rows if y == label][:max(1, cap // classes)]
-        self.lengths = [len(gzip.compress(x, mtime=0)) for x, _ in self.samples]
-        self.tie_keys = [hashlib.sha256(x).hexdigest() for x, _ in self.samples]
-        self.classes, self.k = classes, k
-        self.bytes = sum(len(x) for x, _ in self.samples)
-
-    def score(self, text):
-        x = text.encode()
-        cx = len(gzip.compress(x, mtime=0))
-        distances = [(len(gzip.compress(z + b" " + x, mtime=0)) - min(cx, cz)) / max(cx, cz)
-                     for (z, _), cz in zip(self.samples, self.lengths)]
-        nearest = np.lexsort((self.tie_keys, distances))[:self.k]
-        votes = np.full(self.classes, 0.01)
-        for i in nearest:
-            votes[self.samples[i][1]] += 1
-        return np.log(votes)
-
-
-class Lexical:
-    def __init__(self, rows, classes):
-        self.vectorizer = FeatureUnion([
-            ("word", TfidfVectorizer(ngram_range=(1, 2), sublinear_tf=True, min_df=2)),
-            ("char", TfidfVectorizer(analyzer="char", ngram_range=(3, 5), sublinear_tf=True, min_df=2, max_features=100000)),
-        ])
-        x = self.vectorizer.fit_transform([x for x, _ in rows])
-        self.classifier = LogisticRegression(C=4, max_iter=500).fit(x, [y for _, y in rows])
-        self.bytes = None
-
-    def score(self, text):
-        return self.classifier.predict_log_proba(self.vectorizer.transform([text]))[0]
-
-
-def scores_for(model, rows, temp=1):
-    scores, times = [], []
-    for text, _ in rows:
-        start = time.perf_counter_ns()
-        score = model.score(normalize(text))
-        softmax(score / temp)
-        predictions([score], [text])
-        scores.append(score)
-        times.append((time.perf_counter_ns() - start) / 1e6)
-    return np.asarray(scores), times
-
-
-def predictions(scores, texts):
-    # Exact ties use a reproducible text hash instead of always selecting class 0.
-    out = []
-    for score, text in zip(scores, texts):
-        ties = np.flatnonzero(score == score.max())
-        index = int(hashlib.sha256(text.encode()).hexdigest(), 16) % len(ties)
-        out.append(int(ties[index]))
-    return np.asarray(out)
-
-
-def temperature(scores, gold):
-    choices = np.geomspace(.05, 100, 100)
-    losses = [log_loss(gold, softmax(scores / t, axis=1), labels=np.arange(scores.shape[1])) for t in choices]
-    return float(choices[np.argmin(losses)])
-
-
-def metrics(scores, rows, times, temp):
-    gold = np.asarray([y for _, y in rows])
-    pred = predictions(scores, [x for x, _ in rows])
-    prob = softmax(scores / temp, axis=1)
-    conf = prob.max(axis=1)
-    correct = pred == gold
-    ece = 0.
-    for i in range(10):
-        mask = (conf >= i / 10) & ((conf < (i + 1) / 10) if i < 9 else (conf <= 1))
-        if mask.any():
-            ece += mask.mean() * abs(correct[mask].mean() - conf[mask].mean())
-    result = {"accuracy": float(correct.mean()), "macro_f1": float(f1_score(gold, pred, labels=np.arange(scores.shape[1]), average="macro", zero_division=0)),
-              "ece": float(ece), "nll": float(log_loss(gold, prob, labels=np.arange(scores.shape[1]))),
-              "brier": float(np.mean(np.sum((prob - np.eye(scores.shape[1])[gold]) ** 2, axis=1))),
-              "tie_rate": float(np.mean((scores == scores.max(axis=1, keepdims=True)).sum(axis=1) > 1)),
-              "p50_ms": float(np.median(times)), "p95_ms": float(np.percentile(times, 95)),
-              "serial_items_per_s": float(1000 / np.mean(times)), "n": len(rows)}
-    if scores.shape[1] == 2:
-        result["auroc"] = float(roc_auc_score(gold, prob[:, 1]))
-    n, p, z = len(rows), float(correct.mean()), 1.96
-    center = (p + z * z / (2 * n)) / (1 + z * z / n)
-    radius = z * np.sqrt(p * (1 - p) / n + z * z / (4 * n * n)) / (1 + z * z / n)
-    result["accuracy_wilson_95"] = [float(center - radius), float(center + radius)]
-    mask = conf >= .9
-    result["confidence_at_least_90pct"] = {
-        "n": int(mask.sum()), "coverage": float(mask.mean()),
-        "accuracy": float(correct[mask].mean()) if mask.any() else None,
-    }
-    return result, [{"text_sha256": hashlib.sha256(x.encode()).hexdigest(), "gold": int(y),
-                      "prediction": int(p), "probabilities": ps.tolist(), "ms": ms}
-                     for (x, y), p, ps, ms in zip(rows, pred, prob, times)]
-
-
-def run_dataset(name, args, outdir):
-    train, val, test, meta = load_split(name, args.seed, args.n, args.train_cap, args.validation_n)
-    classes = len(meta["labels"])
-    configs = [("deflate", {"codec": "deflate", "size": s}) for s in [4096, 32768]]
-    configs += [("zstd", {"codec": "zstd", "size": s, "trained": trained})
-                for s in [1024, 4096, 16384] for trained in [True, False]]
-    configs += [("zstd", {"codec": "zstd", "size": s, "level": 9})
-                for s in [65536, 262144]]
-    configs += [("zstd", {"codec": "zstd", "size": 16777216, "level": 9, "trained": False})]
-    configs += [("zstd-mixture", {"codec": "zstd", "size": s, "shards": shards})
-                for s in [1024, 4096] for shards in [4, 8]]
-    configs += [("zstd-mixture", {"codec": "zstd", "size": 16777216, "shards": shards,
-                                  "trained": False, "level": 9, "reduction": reduction})
-                for shards in [3, 5] for reduction in ["min", "mean"]]
-    configs += [("gzip-knn", {"k": k}) for k in [1, 5]]
-    configs += [("tfidf-logistic", {})]
-    if args.families:
-        configs = [(family, config) for family, config in configs if family in args.families]
-    best, sweep = {}, []
-    for family, config in configs:
-        start = time.perf_counter()
-        try:
-            if family == "gzip-knn":
-                model = GzipKNN(train, classes, seed=args.seed, **config)
-            elif family == "tfidf-logistic":
-                model = Lexical(train, classes)
-            else:
-                model = Dictionaries(train, classes, seed=args.seed, **config)
-        except zstd.ZstdError as exc:
-            sweep.append({"family": family, "config": config, "training_error": str(exc)})
-            continue
-        fit_s = time.perf_counter() - start
-        scores, times = scores_for(model, val)
-        pred = predictions(scores, [x for x, _ in val])
-        acc = float(accuracy_score([y for _, y in val], pred))
-        entry = {"family": family, "config": config, "validation_accuracy": acc,
-                 "fit_s": fit_s, "dictionary_or_sample_bytes": model.bytes}
-        sweep.append(entry)
-        print(name, family, config, f"validation={acc:.3f}", flush=True)
-        if family not in best or acc > best[family][0]["validation_accuracy"]:
-            best[family] = entry, model, scores
-    report = {"dataset": meta, "validation_sweep": sweep, "selected": {}}
-    for family, (entry, model, scores) in best.items():
-        temp = temperature(scores, [y for _, y in val])
-        test_scores, times = scores_for(model, test, temp)
-        summary, rows = metrics(test_scores, test, times, temp)
-        # OOD-like inputs must yield a finite score for every allowed class.
-        for text in ["", "🙂 café 你好", "x" * 50000]:
-            edge = model.score(text)
-            assert edge.shape == (classes,) and np.isfinite(edge).all()
-        report["selected"][family] = {**entry, "temperature": temp, **summary}
-        (outdir / f"{name}-{family}-predictions.json").write_text(json.dumps(rows))
-        print(name, family, "TEST", json.dumps(summary), flush=True)
-    (outdir / f"{name}.json").write_text(json.dumps(report, indent=2))
-    return report
-
-
-def predict_jevbench(state, question, labels):
-    if not isinstance(state, str):
-        state = json.dumps(state, ensure_ascii=False, sort_keys=True)
-    x = normalize(state + "\n" + question["instructions"]).encode()
-    cx = len(gzip.compress(x, mtime=0))
-    criteria = question.get("criteria") or {}
-    if isinstance(criteria, list):
-        criteria = {str(i): description for i, description in enumerate(criteria)}
-    scores = []
-    for label in labels:
-        key = {"yes": "true", "no": "false"}.get(label, label) if question["type"] == "noul" else label
-        description = criteria.get(key)
-        if description is None:
-            description = ""
-        z = normalize(label.replace("_", " ") + ": " + str(description)).encode()
-        cz = len(gzip.compress(z, mtime=0))
-        # Symmetric NCD: both concat directions, normalized for label length.
-        joint = min(len(gzip.compress(x + b" " + z, mtime=0)), len(gzip.compress(z + b" " + x, mtime=0)))
-        scores.append(-(joint - min(cx, cz)) / max(cx, cz))
-    scores = np.asarray(scores)
-    ties = np.flatnonzero(scores == scores.max())
-    # Choose by label name hash so option permutation does not change ties.
-    index = min(ties, key=lambda i: hashlib.sha256(x + labels[i].encode()).hexdigest())
-    return labels[index], len(ties) > 1
-
-
-def jevbench(root, outdir):
-    results = {}
-    for tier in ["easy", "original", "hard"]:
-        path = root / "datasets" / "public" / f"{tier}.jsonl"
-        rows = [json.loads(line) for line in path.read_text().splitlines() if line.strip()]
-        predictions_out = []
-        if any(r["expected"] is None or r.get("provenance", {}).get("exclude_reason") for r in rows):
-            raise ValueError("This diagnostic requires fully scored public cohorts")
-        for row in rows:
-            start = time.perf_counter_ns()
-            prediction, tie = predict_jevbench(row["state"], row["question"], row["labels"])
-            predictions_out.append({"id": row["id"], "group": row.get("group"), "family": row["family"],
-                                    "gold": str(row["expected"]), "prediction": prediction,
-                                    "tie": tie, "chance": 1 / len(row["labels"]),
-                                    "ms": (time.perf_counter_ns() - start) / 1e6})
-        family_scores = {}
-        for family in sorted({r["family"] for r in predictions_out}):
-            group = [r for r in predictions_out if r["family"] == family]
-            family_scores[family] = {"n": len(group), "accuracy": np.mean([r["gold"] == r["prediction"] for r in group])}
-        results[tier] = {"n": len(rows), "sha256": hashlib.sha256(path.read_bytes()).hexdigest(),
-                         "accuracy": float(np.mean([r["gold"] == r["prediction"] for r in predictions_out])),
-                         "uniform_chance": float(np.mean([r["chance"] for r in predictions_out])),
-                         "p50_ms": float(np.median([r["ms"] for r in predictions_out])), "families": family_scores}
-        (outdir / f"jevbench-{tier}-predictions.json").write_text(json.dumps(predictions_out, indent=2))
-    results["source_commit"] = subprocess.check_output(["git", "-C", str(root), "rev-parse", "HEAD"], text=True).strip()
-    results["method"] = "Zero-shot symmetric gzip NCD against label and rubric; no training or calibration. Public diagnostic, not an official score."
-    (outdir / "jevbench.json").write_text(json.dumps(results, indent=2))
-    print("JEVBENCH", json.dumps(results), flush=True)
-    return results
-
-
-def main():
-    ap = argparse.ArgumentParser(description=__doc__)
-    ap.add_argument("--datasets", nargs="+", choices=DATASETS, default=list(DATASETS))
-    ap.add_argument("--families", nargs="+", choices=["deflate", "zstd", "zstd-mixture", "gzip-knn", "tfidf-logistic"])
-    ap.add_argument("--seed", type=int, default=0)
-    ap.add_argument("--n", type=int, default=500)
-    ap.add_argument("--train-cap", type=int, default=10000)
-    ap.add_argument("--validation-n", type=int, default=600)
-    ap.add_argument("--out", type=Path, default=Path(__file__).parent / "data" / "compression")
-    ap.add_argument("--jevbench", type=Path)
-    args = ap.parse_args()
-    if min(args.n, args.train_cap, args.validation_n) <= 0:
-        ap.error("sample counts must be positive")
-    args.out.mkdir(parents=True, exist_ok=False)
-    metadata = {"args": {k: str(v) if isinstance(v, Path) else v for k, v in vars(args).items()},
-                "platform": platform.platform(), "processor": platform.processor(),
-                "python": platform.python_version(), "zlib": zlib.ZLIB_RUNTIME_VERSION,
-                "thread_environment": {k: os.environ.get(k) for k in ["OMP_NUM_THREADS", "OPENBLAS_NUM_THREADS"]},
-                "timing": "Serial local warm inference: normalization, compression/features, probabilities, tie-breaking. Excludes training, network and data loading.",
-                "versions": {k: importlib.metadata.version(k) for k in ["datasets", "numpy", "scipy", "scikit-learn", "zstandard"]},
-                "script_sha256": hashlib.sha256(Path(__file__).read_bytes()).hexdigest()}
-    (args.out / "manifest.json").write_text(json.dumps(metadata, indent=2))
-    summary = {"manifest": metadata, "datasets": {}}
-    for name in args.datasets:
-        summary["datasets"][name] = run_dataset(name, args, args.out)
-    if args.jevbench:
-        summary["jevbench"] = jevbench(args.jevbench, args.out)
-    (args.out / "summary.json").write_text(json.dumps(summary, indent=2))
-
-
-if __name__ == "__main__":
-    main()
diff --git a/eval/jevbench-zero-shot-results.json b/eval/jevbench-zero-shot-results.json
new file mode 100644
index 0000000..212ca1e
--- /dev/null
+++ b/eval/jevbench-zero-shot-results.json
@@ -0,0 +1,2093 @@
+{
+  "manifest": {
+    "source_sha256": {
+      "easy": "231df3c2c8e88a1a8c137ebe85de96ba70fabd330849098ac7b3c52c70b7172b",
+      "original": "5c2414edb3006b8bfcb70fda433f0f9ca015759433849f8d3104328a1f7c4180",
+      "hard": "89e9e6becb33ed88c1de7d42dcc87531b2fb64cfaef4e1986faf7c37b3f80ebb"
+    },
+    "script_sha256": "7162b92992aa7ac4706b03e6ee957c5c987059c219810464e75aeb9bdcb02eed",
+    "platform": "macOS-26.5.1-arm64-arm-64bit",
+    "python": "3.12.14",
+    "features": [
+      "word_cosine",
+      "word_window_cosine",
+      "char_cosine",
+      "gzip_ncd",
+      "deflate_gain",
+      "zstd_gain",
+      "label_word_cosine"
+    ],
+    "versions": {
+      "numpy": "2.5.3",
+      "scipy": "1.18.1",
+      "scikit-learn": "1.9.1",
+      "zstandard": "0.25.0"
+    },
+    "all_methods_p50_ms": 1.533333,
+    "all_methods_p95_ms": 11.855979000000001,
+    "option_reversal_predictions_checked": 12474,
+    "note": "Public zero-shot diagnostic only; no gold answers or labeled examples enter prediction. IDF is computed per request. All methods are fixed; multiple exploratory comparisons, no official sealed evaluation. Timing includes every method together, excluding verification and loading."
+  },
+  "uniform_chance": 0.3176046176046176,
+  "versus_tfidf": {
+    "tfidf-word": {
+      "accuracy_difference": -0.03463203463203463,
+      "scenario_bootstrap_95": [
+        -0.07860262008733625,
+        0.008928571428571428
+      ]
+    },
+    "tfidf-window": {
+      "accuracy_difference": -0.04329004329004329,
+      "scenario_bootstrap_95": [
+        -0.08334205020920502,
+        -0.004273504273504274
+      ]
+    },
+    "tfidf-char": {
+      "accuracy_difference": 0.004329004329004329,
+      "scenario_bootstrap_95": [
+        -0.047619047619047616,
+        0.05803571428571429
+      ]
+    },
+    "tfidf-label": {
+      "accuracy_difference": -0.0735930735930736,
+      "scenario_bootstrap_95": [
+        -0.15625679347826088,
+        0.00847457627118644
+      ]
+    },
+    "gzip": {
+      "accuracy_difference": -0.06926406926406926,
+      "scenario_bootstrap_95": [
+        -0.14979060709761585,
+        0.013215859030837005
+      ]
+    },
+    "deflate": {
+      "accuracy_difference": -0.09956709956709957,
+      "scenario_bootstrap_95": [
+        -0.17467787416531222,
+        -0.021929824561403508
+      ]
+    },
+    "zstd": {
+      "accuracy_difference": -0.09090909090909091,
+      "scenario_bootstrap_95": [
+        -0.1630901287553648,
+        -0.017167381974248927
+      ]
+    },
+    "tfidf-word+gzip:0.25": {
+      "accuracy_difference": -0.05194805194805195,
+      "scenario_bootstrap_95": [
+        -0.12664784522872072,
+        0.025423728813559324
+      ]
+    },
+    "tfidf-word+gzip:0.5": {
+      "accuracy_difference": -0.03896103896103896,
+      "scenario_bootstrap_95": [
+        -0.09333333333333334,
+        0.013217320962145705
+      ]
+    },
+    "tfidf-word+gzip:0.75": {
+      "accuracy_difference": -0.008658008658008658,
+      "scenario_bootstrap_95": [
+        -0.05194805194805195,
+        0.035398230088495575
+      ]
+    },
+    "tfidf-word+deflate:0.25": {
+      "accuracy_difference": -0.08658008658008658,
+      "scenario_bootstrap_95": [
+        -0.16240822320117473,
+        -0.009048754636989946
+      ]
+    },
+    "tfidf-word+deflate:0.5": {
+      "accuracy_difference": -0.06060606060606061,
+      "scenario_bootstrap_95": [
+        -0.11555555555555555,
+        -0.00851063829787234
+      ]
+    },
+    "tfidf-word+deflate:0.75": {
+      "accuracy_difference": -0.03463203463203463,
+      "scenario_bootstrap_95": [
+        -0.07488986784140969,
+        0.004566733693603109
+      ]
+    },
+    "tfidf-word+zstd:0.25": {
+      "accuracy_difference": -0.09523809523809523,
+      "scenario_bootstrap_95": [
+        -0.1630901287553648,
+        -0.02643171806167401
+      ]
+    },
+    "tfidf-word+zstd:0.5": {
+      "accuracy_difference": -0.05627705627705628,
+      "scenario_bootstrap_95": [
+        -0.11013215859030837,
+        0.0
+      ]
+    },
+    "tfidf-word+zstd:0.75": {
+      "accuracy_difference": -0.030303030303030304,
+      "scenario_bootstrap_95": [
+        -0.07017543859649122,
+        0.008849557522123894
+      ]
+    },
+    "tfidf-window+gzip:0.25": {
+      "accuracy_difference": -0.04329004329004329,
+      "scenario_bootstrap_95": [
+        -0.11666666666666667,
+        0.03389830508474576
+      ]
+    },
+    "tfidf-window+gzip:0.5": {
+      "accuracy_difference": -0.03896103896103896,
+      "scenario_bootstrap_95": [
+        -0.09210526315789473,
+        0.013452914798206279
+      ]
+    },
+    "tfidf-window+gzip:0.75": {
+      "accuracy_difference": -0.021645021645021644,
+      "scenario_bootstrap_95": [
+        -0.06222222222222222,
+        0.01832262247026242
+      ]
+    },
+    "tfidf-window+deflate:0.25": {
+      "accuracy_difference": -0.08658008658008658,
+      "scenario_bootstrap_95": [
+        -0.16033755274261605,
+        -0.01293103448275862
+      ]
+    },
+    "tfidf-window+deflate:0.5": {
+      "accuracy_difference": -0.04329004329004329,
+      "scenario_bootstrap_95": [
+        -0.09649122807017543,
+        0.012450736991961628
+      ]
+    },
+    "tfidf-window+deflate:0.75": {
+      "accuracy_difference": -0.04329004329004329,
+      "scenario_bootstrap_95": [
+        -0.08299619640387275,
+        -0.004328537841468883
+      ]
+    },
+    "tfidf-window+zstd:0.25": {
+      "accuracy_difference": -0.08658008658008658,
+      "scenario_bootstrap_95": [
+        -0.15418502202643172,
+        -0.021092632805465875
+      ]
+    },
+    "tfidf-window+zstd:0.5": {
+      "accuracy_difference": -0.05627705627705628,
+      "scenario_bootstrap_95": [
+        -0.1091703056768559,
+        -0.004366812227074236
+      ]
+    },
+    "tfidf-window+zstd:0.75": {
+      "accuracy_difference": -0.03896103896103896,
+      "scenario_bootstrap_95": [
+        -0.07555555555555556,
+        -0.00423728813559322
+      ]
+    },
+    "tfidf-char+gzip:0.25": {
+      "accuracy_difference": -0.03896103896103896,
+      "scenario_bootstrap_95": [
+        -0.11688311688311688,
+        0.039473684210526314
+      ]
+    },
+    "tfidf-char+gzip:0.5": {
+      "accuracy_difference": -0.030303030303030304,
+      "scenario_bootstrap_95": [
+        -0.09583333333333334,
+        0.035398230088495575
+      ]
+    },
+    "tfidf-char+gzip:0.75": {
+      "accuracy_difference": -0.008658008658008658,
+      "scenario_bootstrap_95": [
+        -0.06809238043280597,
+        0.04911264814221645
+      ]
+    },
+    "tfidf-char+deflate:0.25": {
+      "accuracy_difference": -0.09523809523809523,
+      "scenario_bootstrap_95": [
+        -0.16956521739130434,
+        -0.020918322873082318
+      ]
+    },
+    "tfidf-char+deflate:0.5": {
+      "accuracy_difference": -0.05627705627705628,
+      "scenario_bootstrap_95": [
+        -0.1227292663476874,
+        0.008438818565400843
+      ]
+    },
+    "tfidf-char+deflate:0.75": {
+      "accuracy_difference": -0.012987012987012988,
+      "scenario_bootstrap_95": [
+        -0.06722783283717074,
+        0.04310344827586207
+      ]
+    },
+    "tfidf-char+zstd:0.25": {
+      "accuracy_difference": -0.09090909090909091,
+      "scenario_bootstrap_95": [
+        -0.15450643776824036,
+        -0.02631291657090328
+      ]
+    },
+    "tfidf-char+zstd:0.5": {
+      "accuracy_difference": -0.05627705627705628,
+      "scenario_bootstrap_95": [
+        -0.11063829787234042,
+        -0.004166666666666667
+      ]
+    },
+    "tfidf-char+zstd:0.75": {
+      "accuracy_difference": 0.0,
+      "scenario_bootstrap_95": [
+        -0.05240174672489083,
+        0.05454736440030555
+      ]
+    },
+    "tfidf-label+gzip:0.25": {
+      "accuracy_difference": -0.06926406926406926,
+      "scenario_bootstrap_95": [
+        -0.14977973568281938,
+        0.01276595744680851
+      ]
+    },
+    "tfidf-label+gzip:0.5": {
+      "accuracy_difference": -0.03896103896103896,
+      "scenario_bootstrap_95": [
+        -0.11392405063291139,
+        0.03508771929824561
+      ]
+    },
+    "tfidf-label+gzip:0.75": {
+      "accuracy_difference": -0.030303030303030304,
+      "scenario_bootstrap_95": [
+        -0.10460251046025104,
+        0.043478260869565216
+      ]
+    },
+    "tfidf-label+deflate:0.25": {
+      "accuracy_difference": -0.09956709956709957,
+      "scenario_bootstrap_95": [
+        -0.17672413793103448,
+        -0.02127659574468085
+      ]
+    },
+    "tfidf-label+deflate:0.5": {
+      "accuracy_difference": -0.06493506493506493,
+      "scenario_bootstrap_95": [
+        -0.13393533549783548,
+        0.004291845493562232
+      ]
+    },
+    "tfidf-label+deflate:0.75": {
+      "accuracy_difference": -0.03896103896103896,
+      "scenario_bootstrap_95": [
+        -0.10666666666666667,
+        0.030042918454935622
+      ]
+    },
+    "tfidf-label+zstd:0.25": {
+      "accuracy_difference": -0.06926406926406926,
+      "scenario_bootstrap_95": [
+        -0.13964209843030803,
+        0.0
+      ]
+    },
+    "tfidf-label+zstd:0.5": {
+      "accuracy_difference": -0.05194805194805195,
+      "scenario_bootstrap_95": [
+        -0.11914893617021277,
+        0.01694915254237288
+      ]
+    },
+    "tfidf-label+zstd:0.75": {
+      "accuracy_difference": -0.03896103896103896,
+      "scenario_bootstrap_95": [
+        -0.10300765988567909,
+        0.02575107296137339
+      ]
+    },
+    "tfidf+gzip:0.25": {
+      "accuracy_difference": -0.047619047619047616,
+      "scenario_bootstrap_95": [
+        -0.12196151075434804,
+        0.02643464192429141
+      ]
+    },
+    "tfidf+gzip:0.5": {
+      "accuracy_difference": -0.03896103896103896,
+      "scenario_bootstrap_95": [
+        -0.10177885520617587,
+        0.022323930973734748
+      ]
+    },
+    "tfidf+gzip:0.75": {
+      "accuracy_difference": 0.0,
+      "scenario_bootstrap_95": [
+        -0.047619047619047616,
+        0.04782608695652174
+      ]
+    },
+    "tfidf+deflate:0.25": {
+      "accuracy_difference": -0.08658008658008658,
+      "scenario_bootstrap_95": [
+        -0.16017316017316016,
+        -0.012875536480686695
+      ]
+    },
+    "tfidf+deflate:0.5": {
+      "accuracy_difference": -0.07792207792207792,
+      "scenario_bootstrap_95": [
+        -0.14102564102564102,
+        -0.013391369047619069
+      ]
+    },
+    "tfidf+deflate:0.75": {
+      "accuracy_difference": -0.012987012987012988,
+      "scenario_bootstrap_95": [
+        -0.05932834475297512,
+        0.034782608695652174
+      ]
+    },
+    "tfidf+zstd:0.25": {
+      "accuracy_difference": -0.08225108225108226,
+      "scenario_bootstrap_95": [
+        -0.14530102790014685,
+        -0.021274341868013014
+      ]
+    },
+    "tfidf+zstd:0.5": {
+      "accuracy_difference": -0.06060606060606061,
+      "scenario_bootstrap_95": [
+        -0.11441004629283955,
+        -0.008657075682937766
+      ]
+    },
+    "tfidf+zstd:0.75": {
+      "accuracy_difference": 0.017316017316017316,
+      "scenario_bootstrap_95": [
+        -0.013215859030837005,
+        0.04954954954954955
+      ]
+    },
+    "all-four": {
+      "accuracy_difference": -0.06493506493506493,
+      "scenario_bootstrap_95": [
+        -0.12334801762114538,
+        -0.004347826086956522
+      ]
+    }
+  },
+  "published_jev_reference": {
+    "source_sha256": "c772b3b85809f483449ec2a6ae0272347371ed5f40f7faff8394ecf4ee22ae7c",
+    "note": "Published Jev 1.13.0 outcomes on these exact public IDs; not rerun.",
+    "n": 231,
+    "correct": 200,
+    "accuracy": 0.8658008658008658,
+    "by_tier": {
+      "easy": {
+        "n": 48,
+        "correct": 48,
+        "accuracy": 1.0
+      },
+      "hard": {
+        "n": 111,
+        "correct": 81,
+        "accuracy": 0.7297297297297297
+      },
+      "original": {
+        "n": 72,
+        "correct": 71,
+        "accuracy": 0.9861111111111112
+      }
+    },
+    "by_family": {
+      "adequacy": {
+        "n": 12,
+        "correct": 12,
+        "accuracy": 1.0
+      },
+      "adversarial": {
+        "n": 6,
+        "correct": 6,
+        "accuracy": 1.0
+      },
+      "ambiguous": {
+        "n": 7,
+        "correct": 6,
+        "accuracy": 0.8571428571428571
+      },
+      "extraction": {
+        "n": 24,
+        "correct": 24,
+        "accuracy": 1.0
+      },
+      "fact": {
+        "n": 12,
+        "correct": 12,
+        "accuracy": 1.0
+      },
+      "intent": {
+        "n": 24,
+        "correct": 24,
+        "accuracy": 1.0
+      },
+      "judge_hard": {
+        "n": 17,
+        "correct": 13,
+        "accuracy": 0.7647058823529411
+      },
+      "long_policy": {
+        "n": 19,
+        "correct": 12,
+        "accuracy": 0.631578947368421
+      },
+      "multi_hop": {
+        "n": 18,
+        "correct": 15,
+        "accuracy": 0.8333333333333334
+      },
+      "ordinal": {
+        "n": 12,
+        "correct": 12,
+        "accuracy": 1.0
+      },
+      "policy": {
+        "n": 12,
+        "correct": 11,
+        "accuracy": 0.9166666666666666
+      },
+      "probability": {
+        "n": 10,
+        "correct": 7,
+        "accuracy": 0.7
+      },
+      "routing": {
+        "n": 12,
+        "correct": 12,
+        "accuracy": 1.0
+      },
+      "routing_hard": {
+        "n": 5,
+        "correct": 5,
+        "accuracy": 1.0
+      },
+      "temporal_numeric": {
+        "n": 15,
+        "correct": 4,
+        "accuracy": 0.26666666666666666
+      },
+      "tool_selection": {
+        "n": 12,
+        "correct": 12,
+        "accuracy": 1.0
+      },
+      "tradeoff": {
+        "n": 6,
+        "correct": 5,
+        "accuracy": 0.8333333333333334
+      },
+      "trap": {
+        "n": 8,
+        "correct": 8,
+        "accuracy": 1.0
+      }
+    }
+  },
+  "methods": {
+    "tfidf-word": {
+      "n": 231,
+      "correct": 105,
+      "accuracy": 0.45454545454545453,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 38,
+          "accuracy": 0.7916666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 41,
+          "accuracy": 0.36936936936936937
+        },
+        "original": {
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      }
+    },
+    "tfidf-window": {
+      "n": 231,
+      "correct": 103,
+      "accuracy": 0.4458874458874459,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 38,
+          "accuracy": 0.7916666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 39,
+          "accuracy": 0.35135135135135137
+        },
+        "original": {
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      }
+    },
+    "tfidf-char": {
+      "n": 231,
+      "correct": 114,
+      "accuracy": 0.4935064935064935,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 38,
+          "accuracy": 0.7916666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 48,
+          "accuracy": 0.43243243243243246
+        },
+        "original": {
+          "n": 72,
+          "correct": 28,
+          "accuracy": 0.3888888888888889
+        }
+      }
+    },
+    "tfidf-label": {
+      "n": 231,
+      "correct": 96,
+      "accuracy": 0.4155844155844156,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 35,
+          "accuracy": 0.7291666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 35,
+          "accuracy": 0.3153153153153153
+        },
+        "original": {
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      }
+    },
+    "tfidf": {
+      "n": 231,
+      "correct": 113,
+      "accuracy": 0.48917748917748916,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 38,
+          "accuracy": 0.7916666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 46,
+          "accuracy": 0.4144144144144144
+        },
+        "original": {
+          "n": 72,
+          "correct": 29,
+          "accuracy": 0.4027777777777778
+        }
+      }
+    },
+    "gzip": {
+      "n": 231,
+      "correct": 97,
+      "accuracy": 0.4199134199134199,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 25,
+          "accuracy": 0.5208333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 47,
+          "accuracy": 0.42342342342342343
+        },
+        "original": {
+          "n": 72,
+          "correct": 25,
+          "accuracy": 0.3472222222222222
+        }
+      }
+    },
+    "deflate": {
+      "n": 231,
+      "correct": 90,
+      "accuracy": 0.38961038961038963,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 32,
+          "accuracy": 0.6666666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 34,
+          "accuracy": 0.3063063063063063
+        },
+        "original": {
+          "n": 72,
+          "correct": 24,
+          "accuracy": 0.3333333333333333
+        }
+      }
+    },
+    "zstd": {
+      "n": 231,
+      "correct": 92,
+      "accuracy": 0.39826839826839827,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 25,
+          "accuracy": 0.5208333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 37,
+          "accuracy": 0.3333333333333333
+        },
+        "original": {
+          "n": 72,
+          "correct": 30,
+          "accuracy": 0.4166666666666667
+        }
+      }
+    },
+    "tfidf-word+gzip:0.25": {
+      "n": 231,
+      "correct": 101,
+      "accuracy": 0.43722943722943725,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 30,
+          "accuracy": 0.625
+        },
+        "hard": {
+          "n": 111,
+          "correct": 44,
+          "accuracy": 0.3963963963963964
+        },
+        "original": {
+          "n": 72,
+          "correct": 27,
+          "accuracy": 0.375
+        }
+      }
+    },
+    "tfidf-word+gzip:0.5": {
+      "n": 231,
+      "correct": 104,
+      "accuracy": 0.45021645021645024,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 34,
+          "accuracy": 0.7083333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 44,
+          "accuracy": 0.3963963963963964
+        },
+        "original": {
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      }
+    },
+    "tfidf-word+gzip:0.75": {
+      "n": 231,
+      "correct": 111,
+      "accuracy": 0.4805194805194805,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 40,
+          "accuracy": 0.8333333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 44,
+          "accuracy": 0.3963963963963964
+        },
+        "original": {
+          "n": 72,
+          "correct": 27,
+          "accuracy": 0.375
+        }
+      }
+    },
+    "tfidf-word+deflate:0.25": {
+      "n": 231,
+      "correct": 93,
+      "accuracy": 0.4025974025974026,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 33,
+          "accuracy": 0.6875
+        },
+        "hard": {
+          "n": 111,
+          "correct": 35,
+          "accuracy": 0.3153153153153153
+        },
+        "original": {
+          "n": 72,
+          "correct": 25,
+          "accuracy": 0.3472222222222222
+        }
+      }
+    },
+    "tfidf-word+deflate:0.5": {
+      "n": 231,
+      "correct": 99,
+      "accuracy": 0.42857142857142855,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 37,
+          "accuracy": 0.7708333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 34,
+          "accuracy": 0.3063063063063063
+        },
+        "original": {
+          "n": 72,
+          "correct": 28,
+          "accuracy": 0.3888888888888889
+        }
+      }
+    },
+    "tfidf-word+deflate:0.75": {
+      "n": 231,
+      "correct": 105,
+      "accuracy": 0.45454545454545453,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 38,
+          "accuracy": 0.7916666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 40,
+          "accuracy": 0.36036036036036034
+        },
+        "original": {
+          "n": 72,
+          "correct": 27,
+          "accuracy": 0.375
+        }
+      }
+    },
+    "tfidf-word+zstd:0.25": {
+      "n": 231,
+      "correct": 91,
+      "accuracy": 0.3939393939393939,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 29,
+          "accuracy": 0.6041666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 38,
+          "accuracy": 0.34234234234234234
+        },
+        "original": {
+          "n": 72,
+          "correct": 24,
+          "accuracy": 0.3333333333333333
+        }
+      }
+    },
+    "tfidf-word+zstd:0.5": {
+      "n": 231,
+      "correct": 100,
+      "accuracy": 0.4329004329004329,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 35,
+          "accuracy": 0.7291666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 40,
+          "accuracy": 0.36036036036036034
+        },
+        "original": {
+          "n": 72,
+          "correct": 25,
+          "accuracy": 0.3472222222222222
+        }
+      }
+    },
+    "tfidf-word+zstd:0.75": {
+      "n": 231,
+      "correct": 106,
+      "accuracy": 0.4588744588744589,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 38,
+          "accuracy": 0.7916666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 41,
+          "accuracy": 0.36936936936936937
+        },
+        "original": {
+          "n": 72,
+          "correct": 27,
+          "accuracy": 0.375
+        }
+      }
+    },
+    "tfidf-window+gzip:0.25": {
+      "n": 231,
+      "correct": 103,
+      "accuracy": 0.4458874458874459,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 30,
+          "accuracy": 0.625
+        },
+        "hard": {
+          "n": 111,
+          "correct": 46,
+          "accuracy": 0.4144144144144144
+        },
+        "original": {
+          "n": 72,
+          "correct": 27,
+          "accuracy": 0.375
+        }
+      }
+    },
+    "tfidf-window+gzip:0.5": {
+      "n": 231,
+      "correct": 104,
+      "accuracy": 0.45021645021645024,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 34,
+          "accuracy": 0.7083333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 44,
+          "accuracy": 0.3963963963963964
+        },
+        "original": {
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      }
+    },
+    "tfidf-window+gzip:0.75": {
+      "n": 231,
+      "correct": 108,
+      "accuracy": 0.4675324675324675,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 40,
+          "accuracy": 0.8333333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 41,
+          "accuracy": 0.36936936936936937
+        },
+        "original": {
+          "n": 72,
+          "correct": 27,
+          "accuracy": 0.375
+        }
+      }
+    },
+    "tfidf-window+deflate:0.25": {
+      "n": 231,
+      "correct": 93,
+      "accuracy": 0.4025974025974026,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 33,
+          "accuracy": 0.6875
+        },
+        "hard": {
+          "n": 111,
+          "correct": 35,
+          "accuracy": 0.3153153153153153
+        },
+        "original": {
+          "n": 72,
+          "correct": 25,
+          "accuracy": 0.3472222222222222
+        }
+      }
+    },
+    "tfidf-window+deflate:0.5": {
+      "n": 231,
+      "correct": 103,
+      "accuracy": 0.4458874458874459,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 37,
+          "accuracy": 0.7708333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 38,
+          "accuracy": 0.34234234234234234
+        },
+        "original": {
+          "n": 72,
+          "correct": 28,
+          "accuracy": 0.3888888888888889
+        }
+      }
+    },
+    "tfidf-window+deflate:0.75": {
+      "n": 231,
+      "correct": 103,
+      "accuracy": 0.4458874458874459,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 38,
+          "accuracy": 0.7916666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 38,
+          "accuracy": 0.34234234234234234
+        },
+        "original": {
+          "n": 72,
+          "correct": 27,
+          "accuracy": 0.375
+        }
+      }
+    },
+    "tfidf-window+zstd:0.25": {
+      "n": 231,
+      "correct": 93,
+      "accuracy": 0.4025974025974026,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 29,
+          "accuracy": 0.6041666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 40,
+          "accuracy": 0.36036036036036034
+        },
+        "original": {
+          "n": 72,
+          "correct": 24,
+          "accuracy": 0.3333333333333333
+        }
+      }
+    },
+    "tfidf-window+zstd:0.5": {
+      "n": 231,
+      "correct": 100,
+      "accuracy": 0.4329004329004329,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 35,
+          "accuracy": 0.7291666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 40,
+          "accuracy": 0.36036036036036034
+        },
+        "original": {
+          "n": 72,
+          "correct": 25,
+          "accuracy": 0.3472222222222222
+        }
+      }
+    },
+    "tfidf-window+zstd:0.75": {
+      "n": 231,
+      "correct": 104,
+      "accuracy": 0.45021645021645024,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 38,
+          "accuracy": 0.7916666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 39,
+          "accuracy": 0.35135135135135137
+        },
+        "original": {
+          "n": 72,
+          "correct": 27,
+          "accuracy": 0.375
+        }
+      }
+    },
+    "tfidf-char+gzip:0.25": {
+      "n": 231,
+      "correct": 104,
+      "accuracy": 0.45021645021645024,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 27,
+          "accuracy": 0.5625
+        },
+        "hard": {
+          "n": 111,
+          "correct": 48,
+          "accuracy": 0.43243243243243246
+        },
+        "original": {
+          "n": 72,
+          "correct": 29,
+          "accuracy": 0.4027777777777778
+        }
+      }
+    },
+    "tfidf-char+gzip:0.5": {
+      "n": 231,
+      "correct": 106,
+      "accuracy": 0.4588744588744589,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 31,
+          "accuracy": 0.6458333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 48,
+          "accuracy": 0.43243243243243246
+        },
+        "original": {
+          "n": 72,
+          "correct": 27,
+          "accuracy": 0.375
+        }
+      }
+    },
+    "tfidf-char+gzip:0.75": {
+      "n": 231,
+      "correct": 111,
+      "accuracy": 0.4805194805194805,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 37,
+          "accuracy": 0.7708333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 48,
+          "accuracy": 0.43243243243243246
+        },
+        "original": {
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      }
+    },
+    "tfidf-char+deflate:0.25": {
+      "n": 231,
+      "correct": 91,
+      "accuracy": 0.3939393939393939,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 32,
+          "accuracy": 0.6666666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 34,
+          "accuracy": 0.3063063063063063
+        },
+        "original": {
+          "n": 72,
+          "correct": 25,
+          "accuracy": 0.3472222222222222
+        }
+      }
+    },
+    "tfidf-char+deflate:0.5": {
+      "n": 231,
+      "correct": 100,
+      "accuracy": 0.4329004329004329,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 36,
+          "accuracy": 0.75
+        },
+        "hard": {
+          "n": 111,
+          "correct": 36,
+          "accuracy": 0.32432432432432434
+        },
+        "original": {
+          "n": 72,
+          "correct": 28,
+          "accuracy": 0.3888888888888889
+        }
+      }
+    },
+    "tfidf-char+deflate:0.75": {
+      "n": 231,
+      "correct": 110,
+      "accuracy": 0.47619047619047616,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 38,
+          "accuracy": 0.7916666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 44,
+          "accuracy": 0.3963963963963964
+        },
+        "original": {
+          "n": 72,
+          "correct": 28,
+          "accuracy": 0.3888888888888889
+        }
+      }
+    },
+    "tfidf-char+zstd:0.25": {
+      "n": 231,
+      "correct": 92,
+      "accuracy": 0.39826839826839827,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 28,
+          "accuracy": 0.5833333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 38,
+          "accuracy": 0.34234234234234234
+        },
+        "original": {
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      }
+    },
+    "tfidf-char+zstd:0.5": {
+      "n": 231,
+      "correct": 100,
+      "accuracy": 0.4329004329004329,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 34,
+          "accuracy": 0.7083333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 41,
+          "accuracy": 0.36936936936936937
+        },
+        "original": {
+          "n": 72,
+          "correct": 25,
+          "accuracy": 0.3472222222222222
+        }
+      }
+    },
+    "tfidf-char+zstd:0.75": {
+      "n": 231,
+      "correct": 113,
+      "accuracy": 0.48917748917748916,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 38,
+          "accuracy": 0.7916666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 47,
+          "accuracy": 0.42342342342342343
+        },
+        "original": {
+          "n": 72,
+          "correct": 28,
+          "accuracy": 0.3888888888888889
+        }
+      }
+    },
+    "tfidf-label+gzip:0.25": {
+      "n": 231,
+      "correct": 97,
+      "accuracy": 0.4199134199134199,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 27,
+          "accuracy": 0.5625
+        },
+        "hard": {
+          "n": 111,
+          "correct": 45,
+          "accuracy": 0.40540540540540543
+        },
+        "original": {
+          "n": 72,
+          "correct": 25,
+          "accuracy": 0.3472222222222222
+        }
+      }
+    },
+    "tfidf-label+gzip:0.5": {
+      "n": 231,
+      "correct": 104,
+      "accuracy": 0.45021645021645024,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 34,
+          "accuracy": 0.7083333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 45,
+          "accuracy": 0.40540540540540543
+        },
+        "original": {
+          "n": 72,
+          "correct": 25,
+          "accuracy": 0.3472222222222222
+        }
+      }
+    },
+    "tfidf-label+gzip:0.75": {
+      "n": 231,
+      "correct": 106,
+      "accuracy": 0.4588744588744589,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 39,
+          "accuracy": 0.8125
+        },
+        "hard": {
+          "n": 111,
+          "correct": 42,
+          "accuracy": 0.3783783783783784
+        },
+        "original": {
+          "n": 72,
+          "correct": 25,
+          "accuracy": 0.3472222222222222
+        }
+      }
+    },
+    "tfidf-label+deflate:0.25": {
+      "n": 231,
+      "correct": 90,
+      "accuracy": 0.38961038961038963,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 33,
+          "accuracy": 0.6875
+        },
+        "hard": {
+          "n": 111,
+          "correct": 33,
+          "accuracy": 0.2972972972972973
+        },
+        "original": {
+          "n": 72,
+          "correct": 24,
+          "accuracy": 0.3333333333333333
+        }
+      }
+    },
+    "tfidf-label+deflate:0.5": {
+      "n": 231,
+      "correct": 98,
+      "accuracy": 0.42424242424242425,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 38,
+          "accuracy": 0.7916666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 34,
+          "accuracy": 0.3063063063063063
+        },
+        "original": {
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      }
+    },
+    "tfidf-label+deflate:0.75": {
+      "n": 231,
+      "correct": 104,
+      "accuracy": 0.45021645021645024,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 39,
+          "accuracy": 0.8125
+        },
+        "hard": {
+          "n": 111,
+          "correct": 39,
+          "accuracy": 0.35135135135135137
+        },
+        "original": {
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      }
+    },
+    "tfidf-label+zstd:0.25": {
+      "n": 231,
+      "correct": 97,
+      "accuracy": 0.4199134199134199,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 30,
+          "accuracy": 0.625
+        },
+        "hard": {
+          "n": 111,
+          "correct": 38,
+          "accuracy": 0.34234234234234234
+        },
+        "original": {
+          "n": 72,
+          "correct": 29,
+          "accuracy": 0.4027777777777778
+        }
+      }
+    },
+    "tfidf-label+zstd:0.5": {
+      "n": 231,
+      "correct": 101,
+      "accuracy": 0.43722943722943725,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 35,
+          "accuracy": 0.7291666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 35,
+          "accuracy": 0.3153153153153153
+        },
+        "original": {
+          "n": 72,
+          "correct": 31,
+          "accuracy": 0.4305555555555556
+        }
+      }
+    },
+    "tfidf-label+zstd:0.75": {
+      "n": 231,
+      "correct": 104,
+      "accuracy": 0.45021645021645024,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 37,
+          "accuracy": 0.7708333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 37,
+          "accuracy": 0.3333333333333333
+        },
+        "original": {
+          "n": 72,
+          "correct": 30,
+          "accuracy": 0.4166666666666667
+        }
+      }
+    },
+    "tfidf+gzip:0.25": {
+      "n": 231,
+      "correct": 102,
+      "accuracy": 0.44155844155844154,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 29,
+          "accuracy": 0.6041666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 46,
+          "accuracy": 0.4144144144144144
+        },
+        "original": {
+          "n": 72,
+          "correct": 27,
+          "accuracy": 0.375
+        }
+      }
+    },
+    "tfidf+gzip:0.5": {
+      "n": 231,
+      "correct": 104,
+      "accuracy": 0.45021645021645024,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 34,
+          "accuracy": 0.7083333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 46,
+          "accuracy": 0.4144144144144144
+        },
+        "original": {
+          "n": 72,
+          "correct": 24,
+          "accuracy": 0.3333333333333333
+        }
+      }
+    },
+    "tfidf+gzip:0.75": {
+      "n": 231,
+      "correct": 113,
+      "accuracy": 0.48917748917748916,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 40,
+          "accuracy": 0.8333333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 47,
+          "accuracy": 0.42342342342342343
+        },
+        "original": {
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      }
+    },
+    "tfidf+deflate:0.25": {
+      "n": 231,
+      "correct": 93,
+      "accuracy": 0.4025974025974026,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 34,
+          "accuracy": 0.7083333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 34,
+          "accuracy": 0.3063063063063063
+        },
+        "original": {
+          "n": 72,
+          "correct": 25,
+          "accuracy": 0.3472222222222222
+        }
+      }
+    },
+    "tfidf+deflate:0.5": {
+      "n": 231,
+      "correct": 95,
+      "accuracy": 0.41125541125541126,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 36,
+          "accuracy": 0.75
+        },
+        "hard": {
+          "n": 111,
+          "correct": 33,
+          "accuracy": 0.2972972972972973
+        },
+        "original": {
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      }
+    },
+    "tfidf+deflate:0.75": {
+      "n": 231,
+      "correct": 110,
+      "accuracy": 0.47619047619047616,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 37,
+          "accuracy": 0.7708333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 47,
+          "accuracy": 0.42342342342342343
+        },
+        "original": {
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      }
+    },
+    "tfidf+zstd:0.25": {
+      "n": 231,
+      "correct": 94,
+      "accuracy": 0.4069264069264069,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 29,
+          "accuracy": 0.6041666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 39,
+          "accuracy": 0.35135135135135137
+        },
+        "original": {
+          "n": 72,
+          "correct": 26,
+          "accuracy": 0.3611111111111111
+        }
+      }
+    },
+    "tfidf+zstd:0.5": {
+      "n": 231,
+      "correct": 99,
+      "accuracy": 0.42857142857142855,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 35,
+          "accuracy": 0.7291666666666666
+        },
+        "hard": {
+          "n": 111,
+          "correct": 39,
+          "accuracy": 0.35135135135135137
+        },
+        "original": {
+          "n": 72,
+          "correct": 25,
+          "accuracy": 0.3472222222222222
+        }
+      }
+    },
+    "tfidf+zstd:0.75": {
+      "n": 231,
+      "correct": 117,
+      "accuracy": 0.5064935064935064,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 39,
+          "accuracy": 0.8125
+        },
+        "hard": {
+          "n": 111,
+          "correct": 48,
+          "accuracy": 0.43243243243243246
+        },
+        "original": {
+          "n": 72,
+          "correct": 30,
+          "accuracy": 0.4166666666666667
+        }
+      }
+    },
+    "all-four": {
+      "n": 231,
+      "correct": 98,
+      "accuracy": 0.42424242424242425,
+      "by_tier": {
+        "easy": {
+          "n": 48,
+          "correct": 37,
+          "accuracy": 0.7708333333333334
+        },
+        "hard": {
+          "n": 111,
+          "correct": 36,
+          "accuracy": 0.32432432432432434
+        },
+        "original": {
+          "n": 72,
+          "correct": 25,
+          "accuracy": 0.3472222222222222
+        }
+      }
+    }
+  },
+  "family_detail": {
+    "tfidf-char": {
+      "adequacy": {
+        "n": 12,
+        "correct": 5,
+        "accuracy": 0.4166666666666667
+      },
+      "adversarial": {
+        "n": 6,
+        "correct": 4,
+        "accuracy": 0.6666666666666666
+      },
+      "ambiguous": {
+        "n": 7,
+        "correct": 0,
+        "accuracy": 0.0
+      },
+      "extraction": {
+        "n": 24,
+        "correct": 15,
+        "accuracy": 0.625
+      },
+      "fact": {
+        "n": 12,
+        "correct": 8,
+        "accuracy": 0.6666666666666666
+      },
+      "intent": {
+        "n": 24,
+        "correct": 10,
+        "accuracy": 0.4166666666666667
+      },
+      "judge_hard": {
+        "n": 17,
+        "correct": 7,
+        "accuracy": 0.4117647058823529
+      },
+      "long_policy": {
+        "n": 19,
+        "correct": 6,
+        "accuracy": 0.3157894736842105
+      },
+      "multi_hop": {
+        "n": 18,
+        "correct": 5,
+        "accuracy": 0.2777777777777778
+      },
+      "ordinal": {
+        "n": 12,
+        "correct": 8,
+        "accuracy": 0.6666666666666666
+      },
+      "policy": {
+        "n": 12,
+        "correct": 6,
+        "accuracy": 0.5
+      },
+      "probability": {
+        "n": 10,
+        "correct": 4,
+        "accuracy": 0.4
+      },
+      "routing": {
+        "n": 12,
+        "correct": 3,
+        "accuracy": 0.25
+      },
+      "routing_hard": {
+        "n": 5,
+        "correct": 5,
+        "accuracy": 1.0
+      },
+      "temporal_numeric": {
+        "n": 15,
+        "correct": 8,
+        "accuracy": 0.5333333333333333
+      },
+      "tool_selection": {
+        "n": 12,
+        "correct": 11,
+        "accuracy": 0.9166666666666666
+      },
+      "tradeoff": {
+        "n": 6,
+        "correct": 5,
+        "accuracy": 0.8333333333333334
+      },
+      "trap": {
+        "n": 8,
+        "correct": 4,
+        "accuracy": 0.5
+      }
+    },
+    "tfidf": {
+      "adequacy": {
+        "n": 12,
+        "correct": 5,
+        "accuracy": 0.4166666666666667
+      },
+      "adversarial": {
+        "n": 6,
+        "correct": 3,
+        "accuracy": 0.5
+      },
+      "ambiguous": {
+        "n": 7,
+        "correct": 2,
+        "accuracy": 0.2857142857142857
+      },
+      "extraction": {
+        "n": 24,
+        "correct": 16,
+        "accuracy": 0.6666666666666666
+      },
+      "fact": {
+        "n": 12,
+        "correct": 10,
+        "accuracy": 0.8333333333333334
+      },
+      "intent": {
+        "n": 24,
+        "correct": 9,
+        "accuracy": 0.375
+      },
+      "judge_hard": {
+        "n": 17,
+        "correct": 7,
+        "accuracy": 0.4117647058823529
+      },
+      "long_policy": {
+        "n": 19,
+        "correct": 6,
+        "accuracy": 0.3157894736842105
+      },
+      "multi_hop": {
+        "n": 18,
+        "correct": 5,
+        "accuracy": 0.2777777777777778
+      },
+      "ordinal": {
+        "n": 12,
+        "correct": 8,
+        "accuracy": 0.6666666666666666
+      },
+      "policy": {
+        "n": 12,
+        "correct": 7,
+        "accuracy": 0.5833333333333334
+      },
+      "probability": {
+        "n": 10,
+        "correct": 4,
+        "accuracy": 0.4
+      },
+      "routing": {
+        "n": 12,
+        "correct": 2,
+        "accuracy": 0.16666666666666666
+      },
+      "routing_hard": {
+        "n": 5,
+        "correct": 5,
+        "accuracy": 1.0
+      },
+      "temporal_numeric": {
+        "n": 15,
+        "correct": 6,
+        "accuracy": 0.4
+      },
+      "tool_selection": {
+        "n": 12,
+        "correct": 10,
+        "accuracy": 0.8333333333333334
+      },
+      "tradeoff": {
+        "n": 6,
+        "correct": 4,
+        "accuracy": 0.6666666666666666
+      },
+      "trap": {
+        "n": 8,
+        "correct": 4,
+        "accuracy": 0.5
+      }
+    },
+    "gzip": {
+      "adequacy": {
+        "n": 12,
+        "correct": 5,
+        "accuracy": 0.4166666666666667
+      },
+      "adversarial": {
+        "n": 6,
+        "correct": 3,
+        "accuracy": 0.5
+      },
+      "ambiguous": {
+        "n": 7,
+        "correct": 3,
+        "accuracy": 0.42857142857142855
+      },
+      "extraction": {
+        "n": 24,
+        "correct": 11,
+        "accuracy": 0.4583333333333333
+      },
+      "fact": {
+        "n": 12,
+        "correct": 8,
+        "accuracy": 0.6666666666666666
+      },
+      "intent": {
+        "n": 24,
+        "correct": 7,
+        "accuracy": 0.2916666666666667
+      },
+      "judge_hard": {
+        "n": 17,
+        "correct": 7,
+        "accuracy": 0.4117647058823529
+      },
+      "long_policy": {
+        "n": 19,
+        "correct": 4,
+        "accuracy": 0.21052631578947367
+      },
+      "multi_hop": {
+        "n": 18,
+        "correct": 4,
+        "accuracy": 0.2222222222222222
+      },
+      "ordinal": {
+        "n": 12,
+        "correct": 5,
+        "accuracy": 0.4166666666666667
+      },
+      "policy": {
+        "n": 12,
+        "correct": 6,
+        "accuracy": 0.5
+      },
+      "probability": {
+        "n": 10,
+        "correct": 5,
+        "accuracy": 0.5
+      },
+      "routing": {
+        "n": 12,
+        "correct": 3,
+        "accuracy": 0.25
+      },
+      "routing_hard": {
+        "n": 5,
+        "correct": 5,
+        "accuracy": 1.0
+      },
+      "temporal_numeric": {
+        "n": 15,
+        "correct": 5,
+        "accuracy": 0.3333333333333333
+      },
+      "tool_selection": {
+        "n": 12,
+        "correct": 5,
+        "accuracy": 0.4166666666666667
+      },
+      "tradeoff": {
+        "n": 6,
+        "correct": 6,
+        "accuracy": 1.0
+      },
+      "trap": {
+        "n": 8,
+        "correct": 5,
+        "accuracy": 0.625
+      }
+    },
+    "tfidf+zstd:0.75": {
+      "adequacy": {
+        "n": 12,
+        "correct": 5,
+        "accuracy": 0.4166666666666667
+      },
+      "adversarial": {
+        "n": 6,
+        "correct": 3,
+        "accuracy": 0.5
+      },
+      "ambiguous": {
+        "n": 7,
+        "correct": 1,
+        "accuracy": 0.14285714285714285
+      },
+      "extraction": {
+        "n": 24,
+        "correct": 16,
+        "accuracy": 0.6666666666666666
+      },
+      "fact": {
+        "n": 12,
+        "correct": 10,
+        "accuracy": 0.8333333333333334
+      },
+      "intent": {
+        "n": 24,
+        "correct": 10,
+        "accuracy": 0.4166666666666667
+      },
+      "judge_hard": {
+        "n": 17,
+        "correct": 7,
+        "accuracy": 0.4117647058823529
+      },
+      "long_policy": {
+        "n": 19,
+        "correct": 8,
+        "accuracy": 0.42105263157894735
+      },
+      "multi_hop": {
+        "n": 18,
+        "correct": 4,
+        "accuracy": 0.2222222222222222
+      },
+      "ordinal": {
+        "n": 12,
+        "correct": 9,
+        "accuracy": 0.75
+      },
+      "policy": {
+        "n": 12,
+        "correct": 7,
+        "accuracy": 0.5833333333333334
+      },
+      "probability": {
+        "n": 10,
+        "correct": 5,
+        "accuracy": 0.5
+      },
+      "routing": {
+        "n": 12,
+        "correct": 2,
+        "accuracy": 0.16666666666666666
+      },
+      "routing_hard": {
+        "n": 5,
+        "correct": 5,
+        "accuracy": 1.0
+      },
+      "temporal_numeric": {
+        "n": 15,
+        "correct": 6,
+        "accuracy": 0.4
+      },
+      "tool_selection": {
+        "n": 12,
+        "correct": 10,
+        "accuracy": 0.8333333333333334
+      },
+      "tradeoff": {
+        "n": 6,
+        "correct": 5,
+        "accuracy": 0.8333333333333334
+      },
+      "trap": {
+        "n": 8,
+        "correct": 4,
+        "accuracy": 0.5
+      }
+    }
+  },
+  "verification": {
+    "official_scorer_matches": 12474,
+    "best_hybrid_vs_char_tfidf": {
+      "accuracy_difference": 0.012987012987012988,
+      "scenario_bootstrap_95": [
+        -0.03070175438596491,
+        0.05726872246696035
+      ]
+    },
+    "exact_repeatability_predictions_and_scores": 12474
+  },
+  "selection_note": "The highlighted hybrid is the highest accuracy of 45 fixed blends in an exploratory sweep. It was not independently selected or validated. No gold answers or labeled examples are inputs to any prediction method."
+}
diff --git a/eval/jevbench_hybrid.py b/eval/jevbench_hybrid.py
new file mode 100644
index 0000000..f11308e
--- /dev/null
+++ b/eval/jevbench_hybrid.py
@@ -0,0 +1,198 @@
+#!/usr/bin/env python3
+# /// script
+# requires-python = ">=3.12"
+# dependencies = ["numpy==2.5.3", "scipy==1.18.1", "scikit-learn==1.9.1", "zstandard==0.25.0"]
+# ///
+"""TF-IDF + compression on JevBench's public typed decisions.
+
+uv run --python 3.12 eval/jevbench_hybrid.py --jevbench /path/to/jevbench
+
+Zero-shot only: no labeled examples, learned weights, retrieval corpus, or
+calibration. Each TF-IDF vocabulary/IDF uses only the current request. Fixed
+scoring rules are evaluated side by side; no winner is trained on the answers.
+"""
+import argparse
+import gzip
+import hashlib
+import importlib.metadata
+import json
+import platform
+import time
+import zlib
+from pathlib import Path
+
+import numpy as np
+import zstandard as zstd
+from sklearn.feature_extraction.text import TfidfVectorizer
+
+FEATURES = ["word_cosine", "word_window_cosine", "char_cosine", "gzip_ncd",
+            "deflate_gain", "zstd_gain", "label_word_cosine"]
+
+
+def normalized(text):
+    return " ".join(text.lower().split())
+
+
+def centered(values):
+    v = np.asarray(values, dtype=float)
+    std = v.std(axis=0)
+    return np.where(std < 1e-10, 0, (v - v.mean(axis=0)) / np.maximum(std, 1e-10))
+
+
+def inference_text(task):
+    state = task["state"]
+    if not isinstance(state, str):
+        state = json.dumps(state, ensure_ascii=False, sort_keys=True)
+    return normalized(state + "\n" + task["question"]["instructions"])
+
+
+def option_texts(task):
+    q = task["question"]
+    criteria = q.get("criteria") or {}
+    if isinstance(criteria, list):
+        criteria = {str(i): value for i, value in enumerate(criteria)}
+    options = []
+    for label in task["labels"]:
+        key = {"yes": "true", "no": "false"}.get(label, label) if q["type"] == "noul" else label
+        options.append(normalized(label.replace("_", " ") + ": " + str(criteria.get(key) or "")))
+    return options
+
+
+def features(task):
+    text, options = inference_text(task), option_texts(task)
+    words = text.split()
+    chunks = [" ".join(words[i:i + 160]) for i in range(0, len(words), 80)] or [""]
+    docs = [text] + options + chunks + [x.replace("_", " ") for x in task["labels"]]
+    vectors = TfidfVectorizer(ngram_range=(1, 2), token_pattern=r"(?u)\b\w+\b").fit_transform(docs)
+    n = len(options)
+    word = (vectors[1:n+1] @ vectors[0].T).toarray().ravel()
+    window = (vectors[1:n+1] @ vectors[n+1:n+1+len(chunks)].T).toarray().max(axis=1)
+    label = (vectors[-n:] @ vectors[0].T).toarray().ravel()
+    chars = TfidfVectorizer(analyzer="char", ngram_range=(3, 5), max_features=30000).fit_transform([text] + options)
+    char = (chars[1:] @ chars[0].T).toarray().ravel()
+    x = text.encode()
+    cx = len(gzip.compress(x, mtime=0))
+    raw = zlib.compressobj(level=6, wbits=-15, zdict=x[-32768:])
+    zd = zstd.ZstdCompressor(level=3, dict_data=zstd.ZstdCompressionDict(x, dict_type=zstd.DICT_TYPE_RAWCONTENT))
+    zplain = zstd.ZstdCompressor(level=3)
+    values = []
+    for option in options:
+        y = option.encode()
+        cy = len(gzip.compress(y, mtime=0))
+        joint = min(len(gzip.compress(x + b" " + y, mtime=0)), len(gzip.compress(y + b" " + x, mtime=0)))
+        c = raw.copy()
+        conditioned = len(c.compress(y) + c.flush())
+        plain = zlib.compressobj(level=6, wbits=-15)
+        unconditioned = len(plain.compress(y) + plain.flush())
+        values.append([-(joint - min(cx, cy)) / max(cx, cy),
+                       (unconditioned - conditioned) / max(len(y), 1),
+                       (len(zplain.compress(y)) - len(zd.compress(y))) / max(len(y), 1)])
+    a = np.asarray(values)
+    return np.column_stack([word, window, char, a, label])
+
+
+def pick(scores, task):
+    scores = np.asarray(scores)
+    ties = np.flatnonzero(scores == scores.max())
+    # JevBench's distribution scorer resolves ties by label name.
+    return int(min(ties, key=lambda i: task["labels"][i]))
+
+
+def methods(f):
+    word, window, char, gz, df, zs, label = f.T
+    tfidf = centered((centered(window) + centered(char)) / 2)
+    out = {
+        "tfidf-word": centered(word), "tfidf-window": centered(window),
+        "tfidf-char": centered(char), "tfidf-label": centered(label),
+        "tfidf": tfidf, "gzip": centered(gz),
+        "deflate": centered(df), "zstd": centered(zs),
+    }
+    for lexical in ["tfidf-word", "tfidf-window", "tfidf-char", "tfidf-label", "tfidf"]:
+        for compressor in ["gzip", "deflate", "zstd"]:
+            for weight in [.25, .5, .75]:
+                out[f"{lexical}+{compressor}:{weight:g}"] = weight * out[lexical] + (1-weight) * out[compressor]
+    out["all-four"] = (tfidf + out["gzip"] + out["deflate"] + out["zstd"]) / 4
+    return {name: np.round(scores, 10) for name, scores in out.items()}
+
+
+def summarize(rows):
+    def score(group):
+        return {"n": len(group), "correct": sum(r["correct"] for r in group),
+                "accuracy": sum(r["correct"] for r in group) / len(group)}
+    return {**score(rows),
+            "by_tier": {t: score([r for r in rows if r["tier"] == t]) for t in sorted({r["tier"] for r in rows})},
+            "by_family": {f: score([r for r in rows if r["family"] == f]) for f in sorted({r["family"] for r in rows})}}
+
+
+def paired_interval(a, b, seed=0):
+    groups = sorted({r["group"] for r in a})
+    delta = {g: [int(x["correct"]) - int(y["correct"]) for x, y in zip(a, b) if x["group"] == g] for g in groups}
+    rng = np.random.default_rng(seed)
+    samples = []
+    for _ in range(5000):
+        chosen = rng.choice(groups, len(groups), replace=True)
+        differences = [v for g in chosen for v in delta[g]]
+        samples.append(np.mean(differences))
+    return {"accuracy_difference": float(np.mean([int(x["correct"]) - int(y["correct"]) for x, y in zip(a, b)])),
+            "scenario_bootstrap_95": np.percentile(samples, [2.5, 97.5]).tolist()}
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--jevbench", required=True, type=Path)
+    ap.add_argument("--out", type=Path, default=Path(__file__).parent / "data" / "jevbench-zero-shot")
+    args = ap.parse_args()
+    args.out.mkdir(parents=True, exist_ok=False)
+    tasks, source_hashes = [], {}
+    for tier in ["easy", "original", "hard"]:
+        path = args.jevbench / "datasets" / "public" / f"{tier}.jsonl"
+        source_hashes[tier] = hashlib.sha256(path.read_bytes()).hexdigest()
+        for line in path.read_text().splitlines():
+            row = json.loads(line)
+            assert row["expected"] is not None and not row.get("provenance", {}).get("exclude_reason")
+            row["tier"] = tier
+            tasks.append(row)
+    outputs, elapsed = {}, []
+    for task in tasks:
+        view = {k: task[k] for k in ["state", "question", "labels"]}
+        start = time.perf_counter_ns()
+        f = features(view)
+        scores = methods(f)
+        predicted = {name: view["labels"][pick(values, view)] for name, values in scores.items()}
+        elapsed.append((time.perf_counter_ns() - start) / 1e6)
+        assert f.shape == (len(view["labels"]), len(FEATURES)) and np.isfinite(f).all()
+        reversed_view = {**view, "labels": list(reversed(view["labels"]))}
+        reversed_features = features(reversed_view)
+        assert np.allclose(f, reversed_features[::-1], atol=1e-12)
+        reversed_scores = methods(reversed_features)
+        for name, prediction in predicted.items():
+            assert prediction == reversed_view["labels"][pick(reversed_scores[name], reversed_view)], (task["id"], name)
+            outputs.setdefault(name, []).append({"id": task["id"], "group": task.get("group") or task["id"],
+                "tier": task["tier"], "family": task["family"], "gold": str(task["expected"]),
+                "prediction": prediction, "correct": prediction == str(task["expected"]),
+                "labels": view["labels"], "scores": scores[name].tolist()})
+    published_path = args.jevbench / "results/v1.2/jevbench-v1.2-per-task.json"
+    published = json.loads(published_path.read_text())["systems"]["jev-1.13.0"]["public_tasks"]
+    reference = [{**r, "correct": published[r["id"]][0] == "c"} for r in outputs["tfidf"]]
+    report = {"manifest": {"source_sha256": source_hashes,
+              "script_sha256": hashlib.sha256(Path(__file__).read_bytes()).hexdigest(),
+              "platform": platform.platform(), "python": platform.python_version(), "features": FEATURES,
+              "versions": {p: importlib.metadata.version(p) for p in ["numpy", "scipy", "scikit-learn", "zstandard"]},
+              "all_methods_p50_ms": float(np.median(elapsed)), "all_methods_p95_ms": float(np.percentile(elapsed, 95)),
+              "option_reversal_predictions_checked": sum(len(v) for v in outputs.values()),
+              "note": "Public zero-shot diagnostic only; no gold answers or labeled examples enter prediction. IDF is computed per request. All methods are fixed; multiple exploratory comparisons, no official sealed evaluation. Timing includes every method together, excluding verification and loading."},
+              "uniform_chance": float(np.mean([1/len(t["labels"]) for t in tasks])),
+              "methods": {name: summarize(rows) for name, rows in outputs.items()},
+              "versus_tfidf": {name: paired_interval(rows, outputs["tfidf"]) for name, rows in outputs.items() if name != "tfidf"},
+              "published_jev_reference": {"source_sha256": hashlib.sha256(published_path.read_bytes()).hexdigest(),
+                                          "note": "Published Jev 1.13.0 outcomes on these exact public IDs; not rerun.", **summarize(reference)},
+              "oracle_choose_tfidf_or_gzip": {"note": "Diagnostic only: uses gold answers to pick the correct method. Not a deployable model or a bound on other hybrids.",
+                                             "accuracy": float(np.mean([a["correct"] or b["correct"] for a,b in zip(outputs["tfidf"],outputs["gzip"])]))}}
+    for name, rows in outputs.items():
+        (args.out / f"{name}-predictions.json").write_text(json.dumps(rows, indent=2))
+        print(name, json.dumps(report["methods"][name]), flush=True)
+    (args.out / "summary.json").write_text(json.dumps(report, indent=2) + "\n")
+
+
+if __name__ == "__main__":
+    main()