W6 RUN — the vector-sum dipole model is VOID by its own control, and the stranded stratum is empty for a structural reason - #940
Merged
Conversation
…un (dipole = neighbor + bow-wave deconvolution)
Per weather-w-probes-v1.md SS3. Tests the report SS10.2 vector-sum model:
D = c_geo*P_geo (far-field neighbor high) + c_bow*P_bow (relative-motion
bow wave) on CT-F14/F16's 19 stored storms, read from comet_tail_f16.json.
Ported spine() VERBATIM from l4_rail_probe.py (the brief names
comet_tail_f16.py; verified against the tree, the actual constrained
2-parameter dipole fit lives in l4_rail_probe.py -- ported from the real
location instead of the brief's file name). Every other primitive
(wrap_deg/err_deg/geom_ll/disk_mean_uv/circular) is verbatim from
comet_tail_f16.py / the SS0 statistics standard, named at each site.
Motion bearing recovered by algebra (no tracking): exact inversion of
CT-F16's stored err_deg = wrap(lp-(mth+pi/2)). v_rel = v_storm - v_env850
carries the bow term; the annulus neighbor search (600-2500km, lat 20-80N)
finds the strongest positive zonal-anomaly cell.
Bars pre-registered, controls FIRST:
B0 CONTROLS -- joint fit with per-storm PERMUTED P_bow ((i+7)%19) and
P_bow rotated +90deg; either control's joint R2 must stay <=
single-geo R2 + 0.03, else VOID.
B1 IDENTIFIABILITY -- joint R2_vec >= best single R2_vec + 0.10.
B2 SIGN -- c_bow > 0 AND c_geo > 0 (both in physically-predicted
directions).
B3 (descriptive) -- resultant (R_bar, mu, p) of residual bearings,
overall AND stratified by |v_storm| < / >= 8 m/s -- the stranded
stratum: if v_rel genuinely carries the bow term for weak-motion
storms, the stranded residuals should not be worse than the moving
ones.
B4 (descriptive) -- per-storm bearing(D) vs bearing(Dhat) table.
Checkpointed per storm (comet_tail_w6.partial.jsonl, resume-skip on t0),
tag-file heartbeat every 5 storms, per SS0's stranded-rescue protocol.
Offline smoke-tested before commit (no network): r2_vec/fit_joint recover
exact synthetic coefficients (c_geo=2, c_bow=3 -> R2=1.0 to 1e-9);
bearing_deg/wrap_deg/the +90deg rotation matrix/circular() all verified
against hand-computed expectations. Fixed one real bug found by the
smoke pass: an operator-precedence error in the annulus lat-mask (`*`
binds tighter than `&` in Python, silently producing a wrong boolean
combination) -- rewritten to compose the lat mask separately before
combining with `&`.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CcpLeEC3XK8Eye53GKBVvi
… storm, not just derived bearings The first W6 run completed (comet_tail_w6.json committed separately below) with a near-zero c_bow and wrong-signed c_geo. Manually audited one storm by re-fetching it (t0=54358): D=[-0.561,-0.469] Pa, P_bow=[-43.6,-19.3] Pa (dynamic pressure of an 8.9 m/s v_rel -- a physically sane magnitude), P_geo=[-0.340,0.237] Pa/km (A_H=906 Pa at d_H=2187 km -- also physically sane). No implementation bug found -- but the per-storm checkpoint row already carried these fields and the final JSON's per_storm table dropped them, so the finding above could only be verified by a live re-fetch, not from the committed artifact. Same lesson codex kept catching this session for other probes, self-caught here before it needed a reviewer to find it. Fix: per_storm now carries D/P_geo/P_bow/v_rel_ms/A_H_Pa/d_H_km/ theta_H_rad for every storm, not only the derived bearings. Re-running (the partial checkpoint from the first run self-deleted on success, so this refetches all 19 storms -- ~3 min per the brief's own cost estimate, not a burden). Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CcpLeEC3XK8Eye53GKBVvi
…ntrol, plus a reusable sample-composition lesson The dipole vector-sum model (report SS10.2: D = c_geo*P_geo + c_bow*P_bow) was fit on CT-F14/F16's 19 stored storms per weather-w-probes-v1.md SS3's pre-registered bars. Result: B0 VOIDS. Single-geo R2=-0.104 (worse than predicting the mean); both anti-vacuity controls (permuted P_bow, P_bow rotated +90deg) score -0.071/-0.062, both clearing the <=single-geo+0.03 ceiling of -0.074 -- two deliberately WRONG references score as well as or better than the real geo predictor. c_geo carries the wrong sign throughout (-0.41, predicted positive); c_bow ~= 0.0006 -- no measurable weight from the bow predictor at all. B1/B2 correctly report VOID per the pre-registered rule, not their own numbers. Checked for an implementation bug before calling this a clean negative (measurement-skeptic discipline): none found. One storm independently re-fetched and hand-audited; extended to the full sample by fixing a self-caught audit gap (the first run shipped only derived bearings, not the raw predictors -- fixed in 3436918, re-run bit-identical, confirming determinism and that the fix changed nothing but auditability). Every committed value is physically sane: A_H>0 always, d_H inside the 600-2500km annulus always, v_rel 2.8-27.9 m/s. |P_bow| averages 147x |D|, |P_geo| averages 0.61x -- a real Pa-vs-Pa/km scale disparity, but lstsq is scale-invariant per column, so the near-zero c_bow reflects a genuine absence of correlation, not a units artifact. The reusable lesson, filed as E-THE-DISPLACEMENT-FILTER-ATE-THE-STRANDED- STRATUM-1: B3's stranded stratum (|v_storm|<8 m/s) came back n=0 -- every storm has |v_storm|>=12.54 m/s. This is arithmetic, not physics: CT-F14's own qualifying filter (displacement_km>=250 over 6h) implies |v_storm|>=250km/6h=11.574 m/s for ANY admitted storm. A filter selected for one purpose (fast, cleanly-displaced storms) silently excludes exactly the storms a LATER, differently-motivated test needs -- knowable from the filter's own arithmetic before a single fetch, not checked until the stratum came back empty. Consequence: the vector-sum model AS SPECIFIED is disconfirmed on this sample, void by its own control. CT-F17's gate (W6 + independent audit) is now moot for THIS form of the model -- a fresh-sample verdict on a model that fails identifiability on the stored sample is not the next useful step. A revised model form needs its own W6-shaped test; a genuine stranded-rescue test needs a sample built without the displacement floor. Board homes: weather-w-probes-v1.md SS3 RUN section; EPIPHANIES new entry; PR_ARC_INVENTORY results entry (RUN complete, awaiting a PR); STATUS_BOARD D-W6 updated. Note: an earlier commit (afec59c) incorrectly claimed the first run's JSON was "committed separately below" -- it was never committed (superseded by this audited re-run before that could happen), corrected here rather than left standing. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CcpLeEC3XK8Eye53GKBVvi
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_d38ebd24-b3cb-4628-b7b5-ae03481841db) |
📝 WalkthroughWalkthroughThe PR adds and runs the W6 dipole-deconvolution weather probe. It stores ERA5-derived predictors and per-storm diagnostics. The run fails B0 anti-vacuity controls, marks B1/B2 as VOID, finds no stranded storms, and records the model as disconfirmed. ChangesW6 Weather Probe
Estimated code review effort: 4 (Complex) | ~45 minutes Sequence Diagram(s)sequenceDiagram
participant Probe as W6 probe
participant ERA5 as ERA5 Zarr
participant Models as model evaluation
participant Records as result records
Probe->>ERA5: fetch pressure and wind fields
ERA5-->>Probe: return weather data
Probe->>Models: provide measured bearings and predictors
Models-->>Probe: return coefficients and control metrics
Probe->>Records: write JSON, log, and board findings
Possibly related PRs
Suggested reviewers: Poem
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches📝 Generate docstrings
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 |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9affdc5828
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Comment on lines
+110
to
+111
| coef, *_ = np.linalg.lstsq(X, resid, rcond=None) | ||
| return coef |
There was a problem hiding this comment.
The least-squares coefficient points toward increasing pressure because resid = a*x + b*y, while both P_geo and P_bow are constructed as low-pole directions (away from the high and behind relative motion). Thus D should be -coef before fitting, or B2's expected signs must be reversed; as written, the published claim that the negative c_geo has the physically wrong sign is backwards (the committed F16 low-pole bearings are approximately bearing(coef) + 180°).
Useful? React with 👍 / 👎.
Comment on lines
+17
to
+20
| UNITS. c_geo and c_bow are dimensionless least-squares coefficients that | ||
| absorb P_geo's [Pa/km] and P_bow's [Pa] units into themselves -- their SIGN | ||
| is what B2 tests, not their magnitude, and no unit conversion is performed | ||
| or needed. |
There was a problem hiding this comment.
Because spine() regresses pressure in Pa on coordinates in km, D has units Pa/km; consequently c_geo is dimensionless but c_bow, which multiplies a Pa-valued dynamic pressure, has units km⁻¹. Calling both coefficients dimensionless enables the subsequent invalid interpretation of 0.0006 as intrinsically near zero and the dimensionally meaningless raw |P_bow|/|D| ratio; assess the fitted bow contribution or normalized effect instead.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Actionable comments posted: 4
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In @.claude/board/EPIPHANIES.md:
- Around line 15-16: Update the discussion of c_bow in EPIPHANIES.md to avoid
inferring absent correlation from its raw magnitude. Replace the claim that
c_bow ≈ 0.0006 indicates no measurable weight with a scale-aware measure, such
as a standardized coefficient, |c_bow·P_bow|, or an incremental fit statistic,
while preserving the separate c_geo sign observation.
In `@probes/weather-p1/comet_tail_w6.py`:
- Around line 230-234: Move the load_completed() call before the tag_path START
write so the checkpoint is successfully read before any run metadata is
appended. Preserve the existing done value and subsequent partial/tag file
handling.
- Around line 243-278: Wrap the per-storm processing block containing fetch(),
spine(), and predictor calls in try/except Exception. On any exception, append
an ERROR t0=<value> record to tf, flush the file, then re-raise the original
exception; preserve existing NO-VERDICT handling for valid predictor results.
- Around line 17-20: Correct the coefficient documentation and related
conclusions: in probes/weather-p1/comet_tail_w6.py lines 17-20, state that c_geo
is dimensionless while c_bow has units km⁻¹; in
.claude/plans/weather-w-probes-v1.md line 564, correct D to Pa/km and remove any
inference based solely on c_bow’s magnitude; in
.claude/board/PR_ARC_INVENTORY.md lines 21-24, remove the same magnitude-based
inference and rely on failed R² and B0 controls for the conclusion.
🪄 Autofix
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro Plus
Run ID: 332a2aa4-beaf-49b1-be5c-2d113b593ef6
📒 Files selected for processing (7)
.claude/board/EPIPHANIES.md.claude/board/PR_ARC_INVENTORY.md.claude/board/STATUS_BOARD.md.claude/plans/weather-w-probes-v1.mdprobes/weather-p1/comet_tail_w6.jsonprobes/weather-p1/comet_tail_w6.pyprobes/weather-p1/exec-runs/comet_tail_w6.txt
…nvention, unit-consistent contribution metric, error handling, checkpoint ordering All four real, verified before fixing: 1. (P2, both bots) SIGN CONVENTION. spine()'s raw coef is the gradient of INCREASING residual (toward the storm's HIGH side), while P_geo/P_bow both point toward the LOW side -- exactly the convention low_pole_bearing() makes explicit via its "(ph + pi) % (2*pi)" flip (comet_tail_f16.py:160). D = -spine(...) now, with the convention documented at both the module docstring and the call site. Verified offline before committing: negating the fit target flips c_geo/c_bow's signs EXACTLY and leaves R2 (hence every B0/B1 verdict) and the B3 residual-bearing resultant COMPLETELY UNCHANGED (both bearing(D) and bearing(Dhat) shift by 180 deg, canceling in their difference) -- so this fix changes the reported COEFFICIENT SIGNS and any prose asserting them, but not a single pass/fail verdict already on record. 2. (P2, both bots) UNITS. spine()/P_geo are Pa/km, P_bow is Pa -- so c_geo is dimensionless but c_bow carries km^-1, and OLS coefficients rescale inversely under column rescaling while leaving R2/fitted-values unchanged. The prior "c_bow~=0, no measurable weight" claim inferred absence of correlation from a raw coefficient magnitude that was never comparable across differently-unitted columns in the first place. Fixed with a dimensionally valid metric: |c_bow*P_bow| vs |D|, both now consistently Pa/km (the fitted CONTRIBUTION, not the raw coefficient), committed in the output JSON alongside a units_note on B2_sign. 3. (Major, CodeRabbit) No per-storm error handling -- a dead fetch/step would raise uncaught, violating SS0's "record it in the tag-file and stop" rule. Wrapped the per-storm block in try/except, appending ERROR t0=... before re-raising. 4. (Minor, CodeRabbit) load_completed() ran AFTER the tag file's START line was already written -- if the checkpoint read itself failed, the tag file would misleadingly show a successfully-started run. Reordered: load_completed() first, START written with the resumed count. Re-running with all four fixes -- R2/B0/B1/B3 verdicts are algebraically guaranteed unchanged (verified above), only B2's reported coefficient values and the new contribution metric are expected to differ from the prior commit's numbers. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CcpLeEC3XK8Eye53GKBVvi
…rative everywhere it appeared
Rerun confirms the algebra verified before committing the fix: R2 and every
B0/B1 verdict are bit-identical to the pre-fix run (r2_single_geo=
-0.10381057418374096, r2_permuted=-0.07139097664756067, r2_rotated90=
-0.062400718434540536, joint R2=-0.08608330921952212, B3 R_bar=
0.15311408050017475 -- all unchanged). Only the coefficient signs and the
new fitted-contribution metric differ, exactly as predicted:
c_geo: -0.407 (claimed wrong sign) -> +0.407 (CORRECT, matches the
physically predicted positive)
c_bow: +0.0006 (claimed near-zero) -> -0.0006 km^-1 (wrong sign, small)
fitted contribution (Pa/km, the unit-consistent metric): mean |D|=0.745,
geo contribution 25% of |D|, bow contribution 9% of |D| -- modest, not
"no weight"
Corrected the same two errors everywhere they had propagated:
weather-w-probes-v1.md's RUN table + prose (dated correction block, prior
text kept for the record), EPIPHANIES.md's E-THE-DISPLACEMENT-FILTER-ATE-
THE-STRANDED-STRATUM-1 entry (revised in place -- still unmerged, not yet
protected by the append-only rule), PR_ARC_INVENTORY.md's pending entry
(same). STATUS_BOARD.md checked and needed no change -- it never stated a
coefficient sign.
The headline finding is completely unchanged by either correction: B0 VOID,
the vector-sum model as specified is disconfirmed on this sample, the
stranded stratum is empty by CT-F14's own filter arithmetic. What changed
is only the supporting narrative around WHY c_geo/c_bow look the way they
do -- exactly the kind of correction this session's own falsifiability
discipline exists to catch before merge, caught here by codex + CodeRabbit
rather than missed.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CcpLeEC3XK8Eye53GKBVvi
AdaWorldAPI
pushed a commit
that referenced
this pull request
Aug 12, 2026
#940 was substantial (real probe code + real disconfirming results + a same-PR sign/units correction round caught by codex+CodeRabbit), so it gets its own MERGED entry per the mandatory board-hygiene rule -- not covered by the termination clause (that's for hygiene-only PRs). The entry records only what changed AFTER the pre-existing RUN-results entry was drafted (unmerged-PR content composed in place per the PR-878 allowance, since it predated this PR's existence): the sign-convention fix (D = -spine(...), verified R2/B0/B1/B3-invariant offline before committing, confirmed bit-identical on the actual re-run), the units fix (c_bow is km^-1, not dimensionless -- replaced with the dimensionally valid |c_bow*P_bow| vs |D| contribution metric), and two CodeRabbit robustness fixes (per-storm try/except with an ERROR tag-file record; load_completed() before the START line). The headline finding (B0 VOID, model disconfirmed, stranded stratum structurally empty) is unchanged by either correction -- stated explicitly rather than left implicit. LATEST_STATE shipped-table row added. STATUS_BOARD's D-W6 row was already checked earlier and confirmed to state no coefficient sign -- verified again here, no change needed. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CcpLeEC3XK8Eye53GKBVvi
AdaWorldAPI
added a commit
that referenced
this pull request
Aug 12, 2026
…arvest-rfii13 Board hygiene for merged #940 — arc entry + shipped row
AdaWorldAPI
pushed a commit
that referenced
this pull request
Aug 12, 2026
…lifier on the 25%/9% contribution figures The source RUN entry and the committed JSON's fitted_contribution_Pa_per_km keys report MEANS over the 19 storms (mean |c_geo*P_geo| / mean |c_bow*P_bow| vs mean |D|); the #940 summary rows dropped the qualifier. PR_ARC's merged entry gets an appended dated correction line (append-only rule -- the entry merged via #941 before the finding landed, not edited in place); LATEST_STATE's living shipped-table row is fixed in place (precedent: the N-vs-n_qualifying fix on #939). Fourth instance of the same summarization defect class this week: a qualifier or operand pairing true in the source, lost in the summary. The figure itself was never wrong -- its aggregation scope went unstated. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CcpLeEC3XK8Eye53GKBVvi
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.
The last exploratory W-probe runs, and it's a real, honest negative
comet_tail_w6.pytested the report §10.2 vector-sum model (D = c_geo·P_geo + c_bow·P_bow) on CT-F14/F16's 19 stored storms, perweather-w-probes-v1.md§3's pre-registered bars, controls first.B0 VOIDS — the model has nothing to identify
Two deliberately-wrong references score as well as or better than the real geo predictor alone. B1/B2 correctly report VOID per the pre-registered rule — not their own numbers.
⚠ Corrected during review (codex + CodeRabbit P2/Major, both real, both fixed before merge)
The first version of this PR asserted
c_geohad the physically wrong sign andc_bow ≈ 0("no measurable weight"). Both claims were backward:spine()'s raw coefficient points toward the storm's HIGH side (the gradient of increasing residual pressure), whileP_geo/P_boware both constructed pointing toward the LOW side — exactly the conventionlow_pole_bearing()makes explicit via its own(ph + π) % (2π)flip.Dis now−spine(...). Corrected:c_geo = +0.407— the physically predicted positive sign, correct.c_bow = −0.0006km⁻¹ — predicted positive, wrong sign, but small. Verified algebraically and numerically (bit-identical rerun) that this flip changes nothing about R² or the B0/B1 VOID verdicts — OLS is odd-symmetric in the fit target.D/P_geoare [Pa/km];P_bowis [Pa] — soc_geois dimensionless butc_bowcarries km⁻¹, and OLS coefficients rescale inversely under column rescaling while R²/fitted-values stay fixed. Raw|c_bow|(and the original|P_bow|/|D|≈ 147× comparison) was never valid evidence of "no weight." The dimensionally valid measure is the fitted contribution|c_bow·P_bow|vs|D|, both in Pa/km: mean|D|= 0.745, geo contribution = 25 % of|D|, bow contribution = 9 % of|D|— modest, not absent, and consistent throughout with R² < 0.Also fixed: no per-storm error handling (a dead fetch would raise uncaught, violating §0's "record it, then stop" rule — wrapped in try/except now) and the checkpoint being read after the tag file's START line was written.
None of this changes the headline finding. B0/B1's VOID verdicts, every R² value, and B3's residual-bearing resultant are bit-identical before and after — only the coefficient-sign narrative around B2 was wrong and is now corrected everywhere it appeared (this PR body, the plan's RUN section, the EPIPHANIES entry, the arc entry).
The reusable lesson —
E-THE-DISPLACEMENT-FILTER-ATE-THE-STRANDED-STRATUM-1B3's stranded stratum (
|v_storm| < 8 m/s) came back n=0 — every storm has|v_storm| ≥ 12.54 m/s. Arithmetic, not physics: CT-F14's own qualifying filter (displacement_km ≥ 250over 6h) implies|v_storm| ≥ 250 km/6h = 11.574 m/sfor any admitted storm. A filter selected for one purpose (fast, cleanly-displaced storms) silently excludes exactly the storms a later, differently-motivated test needs — knowable from the filter's own arithmetic before a single fetch. Filed as a reusable pre-flight check for any future probe reusing an existing filtered sample.Consequence for the report and CT-F17
The vector-sum model as specified is disconfirmed on this sample — void by its own control, not merely "unproven." CT-F17's gate (W6 + independent adversarial audit) is now moot for this form of the model. A revised model form needs its own W6-shaped mechanistic test first; a genuine stranded-rescue test needs a sample built without the displacement floor.
Board homes
weather-w-probes-v1.md§3 — full RUN section + dated correction blockEPIPHANIES.md— entry (revised in place; still unmerged when the fix landed)PR_ARC_INVENTORY.md— results entry (revised in place, same reason)STATUS_BOARD.md— D-W6 moved from Queued to the RUN verdict (checked — it never asserted a coefficient sign, needed no correction)Probe + docs — zero product code.
🤖 Generated with Claude Code
https://claude.ai/code/session_01CcpLeEC3XK8Eye53GKBVvi