Research: T185 dirty-entry and frame-refit screens on the #2863/#2894 words (negative result, no κ)

[digitalis-p]

Summary

Negative result. No κ claim. The selected record is unchanged.

This package bounds the most direct native-word lever left after #2863 proved the fixed-inventory packing optimum: letting more helpers enter dirty while every existing gate stays fixed. The words screened are:

• the PR2857 shared-kernel words of Attain the fixed-word packing optimum: conditional κ ≈ 0.0017201008054 #2863;
• the newer preload-move words of Conditional κ = 0.001720184528377469: two more staged quartets on #2863's shared-kernel words #2894 (Jeremy Kothe, 07:49 UTC).

Word	Result
PR2857 base/extra	0 of 184 eligible helpers have a frame-feasible dirty entry, in both direct and pre-only modes
PR2894 base/extra	0 of 182
One-hop relays, all four words	8 apparent relays (409, 411, 413, 415, 425, 427, 429, 431), 0 valid under an exact bilinear test
Frame refits (side check), all four words	0 improving: single-gate moves (345 candidates) and forward/backward cascades (2797/3009 on PR2857, 2803/3015 on PR2894)

• Calibration. On the older PR2833 words the same screen returns exactly Conditional κ = 0.0017193776586975866590: dirt-entry rewrite of A-phase helpers 460/462/478/480 on #2833 #2853's 16 quartet helpers: {460,461,468,469}, {462,463,470,471}, {478,479,490,491}, {480,481,492,493}.
• Frames are the obstruction. With every frame constraint removed, all 184 eligible helpers of the PR2857 words are algebraically cancellable (evidence/RELAXED-CENSUS.json).
• Consistency with Conditional κ = 0.001720184528377469: two more staged quartets on #2863's shared-kernel words #2894. Its two new quartets, (456,457,466,467) and (458,459,464,465), all have a zero-frame first gate in the PR2857 words. That puts them outside the fixed-gate class screened here, and Conditional κ = 0.001720184528377469: two more staged quartets on #2863's shared-kernel words #2894 reaches them by re-synthesising the preload. The verifier asserts both facts.
• Independent check of Conditional κ = 0.001720184528377469: two more staged quartets on #2863's shared-kernel words #2894's numbers. This package's ledger reproduces its local profile {1:2404, 2:642, 3:262, 4:128, 5:26, 8:8} and inventory {8:86, 9:232}.

Method

tools/wordlib.py is an independent, exact implementation:

• the PR2833/PR2857 frame-climb ledger;
• an adjoint pass giving every register's target-dirt response just before every program item. It reproduces the stored compensation for 402 of 408 live helpers on all six words; the six exceptions are warp-tainted initial-correction sources.

tools/entry_screen.py asks whether a helper's whole dirt response lies in the exact span of responses reachable by inserted gates r += u*q0:

• insertion at any time up to the helper's purity horizon, in either phase;
• partner cancellation;
• entry at a sub-frame of the first frame;
• frame feasibility over the full program timeline, including A/B pins, warps, restore and end frames;
• optional one-hop relays.

tools/refit_screens.py covers fixed-word frame moves. The card lists that class as a scope-change trigger (#1995/#2013), so it is labelled a side check.

Verification

All checks are standard-library only.

python3 scripts/research_package.py --package research/t185-dirty-entry-frame-screens --check-only
python3.12 scripts/research_package.py --package research/t185-dirty-entry-frame-screens

Check	Outcome
--check-only metadata	PASS
Registered replay, Python 3.12.13	PASS_REGISTERED_RESEARCH_PACKAGE ... outcome=negative-result (exit 0, 414 s)
core / chains2857 / chains2894, Python 3.14 local	94 / 218 / 213 s; timeouts 300 / 450 / 450
python3 -B -m unittest discover -s tests -p 'test_research_*.py'	133 tests OK (5 skipped)
scripts/check_research_commit.py --ref c291ad5	PASS (24 files)

The core part includes these controls:

• a corrupted compensation coefficient is detected;
• a broken frame chain is rejected;
• a de-optimised word (one climb split) is repaired by the refit screen;
• the frame-relaxed subset is entirely in span;
• the relay-validity test accepts six known relays of quartet helper 478.

make verify and the selected frontier1094 replay were not run. This package changes neither the selected construction nor any existing artifact, and CONTRIBUTING asks for affected checks only.

During development, four screen bugs produced false positives before publication. Each was caught by building the word literally or by the control. They are listed in the README; the published tools implement the corrected rules.

Independent reproduction of #2863

Recorded in evidence/T185-REPLAY.json.

• Full cold replay: PASS (exit 0, 1507 s; CPython 3.13.16, NumPy 2.5.4, SciPy 1.18.1, SymPy 1.14.0, mpmath 1.4.1). κ = 0.001720100805416885927944 − 10⁻⁴⁶ − 10⁻⁷⁰ exactly.
  • The release receipt differs from the published one only in a wall-clock seconds field and the manifest hashes derived from it. Substituting the published native hashes and seconds reproduces the published receipt exactly.
• Fast replay: all five stages pass.
  • The published head omits two .gitattributes files (* -text) that its MANIFEST lists, so --hash-only fails as published. Restoring them from Improve conditional κ to 0.00172009662868483 with shared kernels #2857 matches the pinned hash.
  • Under Python 3.14 the regenerated FILL.json.gz differs only in its gzip container; the decompressed payload is identical.

Scope, limits and overlap

• Not covered: changed gate schedules or preload/mixing blocks (the class Conditional κ = 0.001720184528377469: two more staged quartets on #2863's shared-kernel words #2894 uses), simultaneous multi-helper plans, relays longer than one hop, and operator-word corrections for warp-tainted helpers. A finite screen is not an impossibility theorem.
• Overlap. This complements Research: reproduce T185 and certify a precision-only kappa refinement #2875 (same account), whose screens covered adjoining one outside column and 94 entry/frame tests. It also overlaps lanes in the advisory structural investigation in Discussion Cost-aware source-partner helper assignments above PR2588 #2612; no priority is claimed.
• Open next questions are in TASK.json and the README:
  1. re-route the remaining 84 zero-frame-first helpers' preload;
  2. joint multi-helper plans, e.g. 479 via 502/503/490;
  3. warp-tainted helpers.
• Branch base. The branch is built directly on the fork's main (189b527, an ancestor of upstream main). The package was verified on upstream main 47e4094, and its tree object is identical in both commits. Only the 24 new package files change.

Credit

• Source words: Yiyang/Lyy-iiis (Improve conditional κ to 0.00172009662868483 with shared kernels #2857, shared-kernel words), Jeremy Kothe (Conditional κ = 0.001718657437166855: dirt-entry rewrite of helpers 530 and 542 on #2802's construction #2833 dirt-entry words and auditor conventions; Conditional κ = 0.001720184528377469: two more staged quartets on #2863's shared-kernel words #2894 preload-move words), vendored byte-for-byte with their NOTICE/CREDITS files.
• Census and control set: Philippe Van Velzen (Conditional κ = 0.0017193776586975866590: dirt-entry rewrite of A-phase helpers 460/462/478/480 on #2833 #2853).
• Packing optimum and dual prices: Dugongue (Attain the fixed-word packing optimum: conditional κ ≈ 0.0017201008054 #2863).
• Retained lineage: all contributors credited in those notices.

Prepared by digitalis-p with substantial Claude (Anthropic) assistance, including two delegated agents (a census survey and the #2863 cold replay). No independent human review is claimed. Apache-2.0.

🤖 Generated with Claude Code