Research consolidation (no κ): T74 claim ledger — 95 stated numbers and 82 release gates re-checked; 11 text errors (#2302/#2374/#2435 wrong targets, self-κ in 3 READMEs), all gates pass

[thakrarsagar]

Research consolidation (no κ claim). The T74 line has fifteen open conditional-κ PRs, and their descriptions and READMEs contain many interlocking "exceeds PR#### by X" and "PR#### (kappa X)" statements. This package:

• pins each PR's own claim file (16 pinned, counting the PR2265 baseline);
• recomputes every numeric statement between pinned PRs that it transcribed from PR bodies and pinned READMEs, with exact rational arithmetic;
• parses the predecessor gates in each pinned verify_release.py;
• maps which claims have independent checks.

Findings

• The release gates are right, so none of the errors below affects any κ value or release gate. 82 predecessor gates are attributed to pinned PRs and all pass:
  • 55 are EXACT;
  • 17 are conservative (literal above the named κ);
  • 10 sit below the named κ. All 10 are PR2265's published 43-digit κ, a truncation 9.7e-45 below its pinned value, which is immaterial against margins of at least 1.28e-5.
• 95 stated numbers checked (32 from bodies, 63 from READMEs): 84 consistent, 11 flagged.
  • (A) Wrong target. Each printed number is the difference to a different PR than the one named:
    • Conditional κ = 0.001696395766736081: exact pivot-518 and 506 helper-donor kernel(s) (PR1625 shape, KernelsOnBorrow) on the borrowed words, exact frame refit (WildIdeas), 1e-3000 certified supplier endpoint, fixed-grid-maximum kappa C - 1/… #2302 body: "… PR2269 by 8.48163E-8" is the difference to PR2288; the difference to PR2269 is 1.48628E-7.
    • Conditional κ = 0.001697194950901770: exact owner fill on #2322's handoff construction #2374 body and README: "… PR2312 by 1.28957E-7" is the difference to PR2349; the difference to PR2312 is 3.03884E-7.
    • Conditional κ = 0.001698511343697629: entrance raise of 24 copy buffers on the handoff T74 words #2435 body and README: "… PR2312 by 1.31639E-6" is the difference to PR2374; the difference to PR2312 is 1.62028E-6.
  • (B) Own κ quoted as the predecessor's. The READMEs of Conditional κ = 0.001698511343697629: entrance raise of 24 copy buffers on the handoff T74 words #2435, Conditional κ = 0.001698511924002761: optimal owner fill on #2435's entrance-raised words #2447 and Conditional κ = 0.001703676387993676: a fifth #2436 alias on the entrance-raised T74 words #2474 each say "exceeds PR#### (our previous package, kappa X…)" with X equal to the package's own κ.
  • (C) Stale copied boilerplate. The Conditional κ = 0.001696310950415050: exact frame refit of the borrowed gauge44 word (WildIdeas), 1e-3000 certified supplier endpoint, fixed-grid-maximum kappa C - 1/F - B at 1e-3080 slacks (PR2180 / PR2161 retune) and exact-integer coarse… #2288, Conditional κ = 0.001696395766736081: exact pivot-518 and 506 helper-donor kernel(s) (PR1625 shape, KernelsOnBorrow) on the borrowed words, exact frame refit (WildIdeas), 1e-3000 certified supplier endpoint, fixed-grid-maximum kappa C - 1/… #2302 and Conditional κ = 0.001696891067241618: helper width exchange on the borrowed T74 words #2312 READMEs repeat PR2269's "exceeds PR2265 by about 1.28e-5". The true values are 1.290e-5, 1.298e-5 and 1.348e-5.
• One gate label slip. In five scripts, a gate whose comment and literal are PR2363's κ reports "does not exceed PR2322". The check itself is correct.
• Independent-check map. It covers PR2315, PR2286, PR2359, PR2402, PR2406, PR2442, PR2446, PR2453, PR2466 and PR2472, each with its scope qualifiers. PR2436, PR2462 and PR2474, the top three claims, have no independent check yet.

Reproduce

python3 -I -B research/t74-claim-lineage-ledger/ledger.py --fetch
# PASS_T74_CLAIM_LEDGER prs=16 statements=95 (body 32, README 63) consistent=84 (ambiguous 2) flagged=2288r6,2302b2,2302r6,2312r8,2374b3,2374r5,2435b2,2435r1,2435r4,2447r1,2474r1 gates=82

The command needs only the standard library. What it checks:

• every claim file, README and release script against its git blob, and the claim files also against sha256;
• every README quote against the README at its pinned head;
• that the gate parse accounts for every gate line;
• 315 negative controls, all rejected: perturbed numbers, wrong targets, altered quotes, wrong pins and synthetic gates. These run through the same functions as the real check.
• 11 positive controls: each suggested correction must be consistent.

The optional --live mode checks pinned heads, body hashes and body quotes against GitHub. It returned live_drift=0 at the snapshot; any drift exits with status 3.

Scope

• The package checks text and gates only. No κ, outer ledger or construction is re-derived.
• The transcription coverage and its exclusions are listed in the README. Two adversarial review rounds found ten statements, one PR (PR2281) and one gate shape that the first sweeps missed.
• Base: main 0fdd7d82d. Everything ran on CPython 3.14.4. The registered Python 3.12 replay was not run. Package metadata check: PASS_PACKAGE_METADATA_AND_PINS; research unit tests pass.
• Overlap: no earlier report or correction of these numbers was found in PR comments, PR searches or Discussions.

Prepared by Sagar Thakrar with substantial AI assistance (Claude Code, Claude Opus 5.5), including adversarial AI reviewers; no independent human review is claimed. Apache-2.0.

🤖 Generated with Claude Code

[thakrarsagar]

Follow-up: #2474's author fixed statement 2474r1 (README now quotes PR2462's κ) in head 71b5b0e. The ledger deliberately stays pinned to the transcribed head 5298f86, so the offline run is unchanged; --live now reports this head drift (exit 3), which is the intended signal. The other ten flagged statements are unchanged at their current heads.