D-TEH-3: calibration math → jc (drift, quorum, lift-gate comparisons); lab cronbach.rs deleted - #1143
Merged
Merged
Conversation
…:reliability The calibration MATH of the thinking-engine battery moves to jc per the ruling that jc is the single home of calibrated math (lift if correct, perfect in jc if not). - jc::drift: reencode_drift / reencode_batch (the re-encode convergence statistic, codec-agnostic — the round trip is a closure the caller supplies) and delta_summary (mean / mean|d| / max|d| / population sigma / fractions above two caller-named cut-offs). 8 tests. - jc::quorum: pairwise_agreement_u8 (per-pair 1 - sigma/sigma_max over k u8 lens tables), QuorumLevel with its 230/179/128 floors, and cronbach_report (alpha by delegation to reliability::cronbach_alpha + per-subject variances + the mean+sigma disagreement count). 6 tests. - jc::reliability: two lift-gate tests carrying the retired lab forms verbatim. Cronbach: same estimator, agrees to 1e-5 on the known-value fixture, the f32 copy loses the 1e7-shifted fixture the f64 form holds to 1e-9 (LIFT). Spearman: the retired copy ranked ties by position; tie-free fixtures cannot separate them, one tie does (1.000 vs 0.948683) (PERFECT-IN-JC, already there). No estimator changed. jc lib 135/135; the new modules are clippy -D warnings clean; fmt clean. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PFnYKqw6d7TTiB9cT8eFdK
…eted - cronbach.rs deleted (cronbach_alpha / variance_agreement_scores / QuorumLevel / cronbach_analysis now live in jc::reliability + jc::quorum). - ground_truth.rs: the private ordinal-rank Spearman and its three tests deleted; calibration::spearman_vs_ground_truth calls jc::reliability::spearman (tie-corrected), keeping its 0.0 fallback. - reencode_safety.rs: glue over jc::drift — the three codec wrappers pass round-trip closures, ReencodeSafety = jc::drift::ReencodeDrift, test_reencode_batch keeps its tuple shape; test_zipper_offsets and all 14 tests unchanged (the x256 proof runs through jc). - silu_correction.rs: CorrectionStats = jc::drift::DeltaSummary; correction_stats is an adapter with the cut-offs named MATERIAL_CORRECTION = 0.01 / LARGE_CORRECTION = 0.1. - examples/certify_jina_v5_7lane.rs: jc::reliability::cronbach_alpha, None -> NaN so an undefined alpha fails every >= verdict. - Cargo.toml: jc path dep; Cargo.lock adds jc and drops entries cargo no longer needs. Lab: cargo check --lib --examples clean except the pre-existing tts_stream_hhtld break (bgz-tensor private fn, untouched); tests for the three modules 23/23. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PFnYKqw6d7TTiB9cT8eFdK
…y, status/plan rows, TD for the copies outside jc Records the #1142 merge in LATEST_STATE; EPIPHANIES E-THE-LIFT-GATE-FOUND-A-TIE-BLIND-SPEARMAN-1; PR_ARC entry; STATUS_BOARD and plan §5 D-TEH-3 math half Shipped + §3 W2 result addendum; TECH_DEBT TD-RELIABILITY-COPIES-OUTSIDE-JC-1 (ndarray::hpc::reliability and perturbation-sim::stats); SUPERSESSION-INDEX regenerated last. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PFnYKqw6d7TTiB9cT8eFdK
|
Warning Review limit reachedNext included review available in 9 minutes. View limit detailsLimit details: You’ve used the included review currently available. Your 80 included PR review attempts over the past 7 days set your current allowance at 1 review per hour. Your organization has reached its usage spending cap. Adjust your spending cap in the billing tab. Review configuration: ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Essentials Run ID: ⛔ Files ignored due to path filters (1)
📒 Files selected for processing (21)
Comment |
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_da9bf458-0f7e-4792-8ffa-0338f771696e) |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: be8b1677f5
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".
…ta None - jc::quorum: max_u8_variance(k) = 255^2 * floor(k/2) * ceil(k/2) / k^2 replaces the even-k constant. The lifted source normalised every k by (255/2)^2, so a maximally split three-lens cell [0, 0, 255] scored 15 instead of 0; perfected in jc per the ruling. New two-sided test (k = 3 maximal -> 0, k = 2 maximal -> 0, k = 3 half-split strictly between) + doctest. 7/7 + doctests green. - lab silu_correction: MATERIAL_CORRECTION / LARGE_CORRECTION are f32 and promoted with the samples, so a correction of exactly 0.1f32 stays 'not more than' 0.1 as the retired code had it; correction_stats returns Option — None for a non-finite correction (an invalid run), Some(empty) only for an empty set. Test pins both; silu_crosscheck adapts (expect + one f32 cast). 10/10 green, both examples compile. - Board: LATEST_STATE / PR_ARC same-PR lines updated. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PFnYKqw6d7TTiB9cT8eFdK
…n main) awareness_dto.rs / persona.rs / world_model.rs were edited in #1142 and never run through rustfmt; the style workflow's format job walks the lab crate too, so main has been red on it since that merge. Formatting only. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PFnYKqw6d7TTiB9cT8eFdK
Owner
Author
|
CI note — Generated by Claude Code |
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.
What
W2 of
thinking-engine-harvest-closure-v1, deliverable D-TEH-3 (math half): the calibration battery is split by NATURE per §1d and the rulingE-JC-IS-THE-HOME-OF-ALL-CALIBRATED-MATH-1— MATH goes tojc, GLUE stays in the lab crate and calls jc, and every lab copy of a statistic dies in this PR.jc (new surface, no existing estimator changed)
jc::drift—reencode_drift(value, max_iterations, codec, round_trip)/reencode_batch(the re-encode convergence statistic; the round trip is a closure the caller supplies, so the statistic is codec-agnostic) anddelta_summary(deltas, material, large)(mean / mean|δ| / max|δ| / population σ / fraction above two caller-named cut-offs). 8 tests.jc::quorum—pairwise_agreement_u8(per-pair1 − σ/σ_max(k)overku8lens tables, withmax_u8_variance(k)the ceiling attainable for THATk— NOT an α per pair, and the name now says so),QuorumLevelwith its 230/179/128 floors,cronbach_report(α by delegation toreliability::cronbach_alpha+ per-subject variances + themean + σdisagreement count). 7 tests.jc::reliability— two lift-gate tests only, carrying the retired lab forms verbatim so the comparison outlives the deleted source.The lift gate — run on fixtures that can DISTINGUISH the copies
f32vs jcf64)1e-5(0.984615)1e7: jc moves< 1e-9(α is affine-invariant), thef32copy loses the fixture outright1e-6y = [1,2,2,3]: lab returns 1.000, jc returns the textbook 0.948683k:[0, 0, 255]atk = 3scored 15 under the lifted even-kceiling, now 0The Spearman result is the finding: the retired copy's own three tests were all tie-free, so "agrees on the existing tests" was true and meant nothing. A calibration battery comparing baked
u8lens distances against ground-truth cosines hits ties constantly, so the retired form was systematically optimistic exactly where it was used. Recorded asE-THE-LIFT-GATE-FOUND-A-TIE-BLIND-SPEARMAN-1.Lab crate (excluded from the workspace)
cronbach.rsdeleted;ground_truth.rslosesspearman_rank_correlation+ranks(+3 tests) and its calibration glue callsjc::reliability::spearman;reencode_safety.rsis glue overjc::drift(ReencodeSafety = jc::drift::ReencodeDrift, three codec wrappers pass closures,test_zipper_offsets+ all 14 tests verbatim — the x256 proof runs through jc);silu_correction.rshasCorrectionStats = jc::drift::DeltaSummarywith the cut-offs namedMATERIAL_CORRECTION/LARGE_CORRECTION(keptf32and promoted with the samples, so exactly0.1f32is not "more than" 0.1) andcorrection_stats -> Option(None= non-finite data,Some(empty)= genuinely empty);certify_jina_v5_7lane.rscalls jc's α (None→NaN, which fails every>=verdict honestly).Cargo.tomlgainsjc = { path = "../jc" }.Review folded in
Codex P2 ×3, all confirmed and fixed in
cfb9b3b: the odd-kquorum ceiling, thef32cut-off boundary, and the NaN-as-empty fallback — each with a two-sided test. Theformatjob was red on the first heads because three lab files from #1142 merged unformatted (main is red on the same job); the rustfmt-only fix is ported here as8c03eb5.Gates
jc lib 136/136 + doctests;
drift/quorum/reliabilityclippy-D warningsclean (the remaining clippy hits in jc are pre-existing files this PR does not touch);cargo fmt --all -- --checkclean. Lab:cargo check --lib --examplesclean except the pre-existingtts_stream_hhtldexample break (a bgz-tensor private fn, untouched, already underTD-THINKING-ENGINE-EXCLUDED-DEBT-1); lab tests for the three modules 24/24. CI: all eight checks green on8c03eb5.Not in this PR
semantic_chunker/spiral_segmenthalves of D-TEH-3 — gated on their own falsifiers, separate.siluandcosine_f32stay in the lab (activation and vector glue, not calibrated math).ndarray::hpc::reliabilityandperturbation-sim::statscarry the same four estimators (both tie-aware, both correct in the sense above). Filed asTD-RELIABILITY-COPIES-OUTSIDE-JC-1: the ruling names the ndarray direction as a deliberate PR with a bit-exactness gate, not a drive-by.Board
Records the #1142 merge in LATEST_STATE; EPIPHANIES entry; PR_ARC entry (with the review round); STATUS_BOARD + plan §5 D-TEH-3 (math half Shipped) + §3 W2 result addendum; TECH_DEBT entry; SUPERSESSION-INDEX regenerated last (byte-identical).
🤖 Generated with Claude Code
https://claude.ai/code/session_01PFnYKqw6d7TTiB9cT8eFdK