Real FunctionBehavior episodes: opcode matching over-admits 99.7% on real code - #1014
Merged
Conversation
…real code PROBE-R2IL-REAL-EPISODES-1 (5/5 green) closes the named-not-built item from Phase 2: the frontier loop measured against the REAL ruff-side R2IL pass-1 harvest — 143 functions from real x86-64 ELF64 binaries, 17,557 typed FlatFact rows (ruff Release r2il-harvest-pass1; provenance FNV-pinned in ruff's committed PROVENANCE.md). Env-gated on R2IL_ORE_TSV; exits CORPUS ABSENT rather than fabricate. Headline (E5): of 380 real occurrences of the top recurring opcode trigram (int_add, copy, store), exactly 1 is dataflow-chained — a happy-path opcode matcher over-admits 99.7%, against a 31.5% adjacency base rate (SSA coverage total: 0 of 11,653 operand rows lack a ValueId). The top idiom is an addressing idiom, not a dataflow pipe: sequential adjacency is not composition, and the macro carrier on real code must be the def-use chain, never the linear opcode window. Confirms and sharpens Phase 2's R6 at real scale. Refuted pre-registration recorded (E3): "real top-1 trigram count > shuffled" failed (380 vs 387) — shuffling creates monotone copy-runs. The real structure is type-collapse: 97 distinct trigram types vs 264 shuffled (>2.7x), top-10 share 50.5% vs 33.6%. Gate rewritten to the measured fact, refutation kept in place. Stamp ceiling binds (E4): at 143 episodes the mod-64 Stamp drops ~26% of evidence for the widest idiom (64 revised / 23 CHOICE-dropped, e=0.999) — conservative, but now a measured capacity note for the Step-2 stamp residue item. Reader validity (E1/E2): probe-local TSV reader cross-validated EXACTLY against ruff's independently committed census (17,557 rows / 5 kinds / 9 opcodes; 143 episodes / 2 binaries / 5,340 ops). First E1 run failed on the CamelCase-vs-snake_case kind column — the gate working. Fences: no ruff_r2il import (separate workspace; schema cited, E1 catches misreads); evidence-read is a measurement, not a re-ingest; no mint, no BPE, no learner subsystem; the R2IL x BPE / OGAR-loco / V4 synthesis stays a three-IF hypothesis. Board hygiene same-commit: EPIPHANIES E-REAL-CODE-INTERLEAVES-THE-OPCODE-MACRO-IS-NOT-A-DATAFLOW-PIPE-1 + INTEGRATION_PLANS entry (assertion-verified writes). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KCGhDYoQBXs3poaR7sFuqp
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_3e62cc16-1250-4924-a830-e23b8aaa4d07) |
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars. ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
AdaWorldAPI
pushed a commit
that referenced
this pull request
Aug 24, 2026
…omain slot Two probes, one POC: PROBE-BPE-R2IL-LOCO-MICROCODE-1 (10/10) and the companion PROBE-STAMP-MORTON-CASCADE-1 (7/7, already landed in this branch's history, board entry now covers both together). Headline: BPE merges over 1,872 real def-use chain occurrences save 113.4 tokens/merge vs 50.1 for the same algorithm run over the linear opcode stream of the same episodes -- 2.3x denser compression per FnIndex domain slot. Confirms and quantifies #1014's "sequential adjacency is not composition" as a working compressor, not just a diagnostic. FnIndex domain ceiling corrected: DOMAIN_FLOOR=0x90 is const-asserted stored-byte ABI; ogar-ro already mints 22 of the 112-slot domain band, so real headroom for a BPE-learned macro is 0xA6..0xFF = 90 slots, not 255/248. B1 measures 7 R2IL atoms consumed; B2's 33 merges leave 50 free. B6 corrected mid-session, the most important fix in this POC: an earlier framing had this probe invent three self-built admission gates. The real architecture (verified in source) is MUL (contract::mul::GateDecision / Homeostasis / FlowState) as the real gate, and the autopoiesis triangle (ValueTenant::{FrozenStyle,LearnedStyle,ExploreStyle}) as RESONANCE-BASED THINKING, never an RL policy -- PROBE-METACOGNITIVE-TRIANGLE-1 (#998) already proved RungReceipt-only judging, FreeEnergyComparison:: minority_wins(), and Explore running in a counterfactual lane. B6 now ranks candidates via shipped TruthValue::revise + Stamp only, and is explicitly forbidden from freeze/admit/promote/gate vocabulary for its own action -- that belongs to MUL and the triangle, never this probe. Three exploratory cross-checks against shipped precedent, all measuring real effects on this corpus and all labelled exploratory, never architectural: INV1 (HighHeelBGZ's stride-as-role) 33 macros collapse into 21 role classes; INV2 (bgz17 LayeredScope's scent-prune-then-escalate) a cheap opcode-multiset check prunes 80.2% of candidates before the expensive exact walk; INV3 (BGZ-HHTL-D's shared-palette amortization) top macros carry 30-80 episodes-per-mint, confirming real cross-episode reuse. Process: the named next measurement (wider corpora) stayed blocked (r2sleigh absent). A container restart lost the first BPE-probe worker's entire write before it committed; re-dispatched from scratch. The re-dispatched worker introduced two real defects (an atom-count conflation printing chain-occurrence atom-slots, 5616, as if it were the 5340-atom corpus total; several raw Option/tuple Debug-format leaks in printed PASS lines) -- hand-fixed by the orchestrator after a worker fix-attempt only partially landed, reverified by recompiling to ground truth rather than trusting either self-report. Board hygiene same-commit: EPIPHANIES E-BPE-OVER-DEFUSE-CHAINS-BEATS-LINEAR-AND-FITS-LOCO-1 + INTEGRATION_PLANS entry (both assertion-verified writes). Meta-review pending -- three consecutive dispatches died to a transient server-side API 529 Overloaded error before any review content ran; this commit protects the verified, gate-passing work against a further container-restart loss while the review keeps retrying. Any finding it returns lands as a follow-up commit before the PR opens. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KCGhDYoQBXs3poaR7sFuqp
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Headline
PROBE-R2IL-REAL-EPISODES-1(5/5 green) closes the named-not-built item from Phase 2 (#1013): the frontier loop measured against the REAL ruff-side R2IL pass-1 harvest — 143 functions from real x86-64 ELF64 binaries, 17,557 typedFlatFactrows (ruff GitHub Releaser2il-harvest-pass1; provenance FNV-pinned in ruff's committed.claude/harvest/r2il/PROVENANCE.md). Env-gated onR2IL_ORE_TSV; the probe exits CORPUS ABSENT rather than fabricate a stand-in.E5: of 380 real occurrences of the top recurring opcode trigram
(int_add, copy, store), exactly 1 is dataflow-chained — a happy-path opcode matcher over-admits 99.7%, against a 31.5% adjacency base rate (SSA coverage is total: 0 of 11,653 operand rows lack a ValueId, so this is a dataflow fact, not missing coverage). The top idiom is an addressing idiom, not a dataflow pipe: sequential adjacency is not composition — on real machine code the macro carrier must be the DEF-USE CHAIN, never the linear opcode window. This confirms and sharpens Phase 2's toy R6 result at real scale.The other measured facts
(copy,copy,copy)runs. The real structure is type-collapse: 97 distinct trigram types vs 264 under shuffle (>2.7×), top-10 occupancy 50.5% vs 33.6%. Recurring idioms are real; top-1 count was the wrong statistic. The gate now asserts the measured facts (with a non-vacuous 2× margin) and pins the shuffle's top type to the predicted monotone-run artifact.(int_add,copy,store): 64 revised, 23 CHOICE-dropped (e=0.999). Conservative (never double-counts), but at real corpus size it discards ~26% of the widest idiom's evidence — a measured capacity note for the Step-2 stamp residue item.Honesty box
ruff_r2ilimport (separate cargo workspace); the schema is cited and E1 exists to catch misreads.Board hygiene (same commit)
EPIPHANIES.mdprepend:E-REAL-CODE-INTERLEAVES-THE-OPCODE-MACRO-IS-NOT-A-DATAFLOW-PIPE-1INTEGRATION_PLANS.mdprepend: R2IL REAL-EPISODE MEASUREMENTBoth writes assertion-verified.
Generated by Claude Code