[JosieElliston]

Corroborating audit of PR6030: fresh full dense replay, itinerary delta, invoice and crossing (no new κ)

This is an independent check of PR #6030 (DreamingOfClouds, helper 470's frame-593 rung removal, conditional κ 0.003125804). It is corroborating evidence and finds no error.

Overlap, stated up front. The work started at a2b4b039, where the PR6030 repair review left an independent check of the changed word open. While it ran, the maintainer admitted PR6030 at 721c0382, with an independent application audit (docs/research/pr-intake-evidence/pr6030-audit/). That audit uses the same Matarener (PR5960) replay and dense-inventory code. So most layers here are a second execution of the same checks, and no decision changes. What this adds:

1. --full recomputes the dense 512-column absolute-phase digest of all 3,163 PR6030 operators. The admission review checks the 7 new tasks densely and notes that it "is not a fresh independent dense replay of every unchanged predecessor map". PASS in about 4 min on 8 cores.
2. An itinerary-level delta: exactly the six declared register itineraries change, and every other MOVE is unchanged as a multiset.
3. Five tamper controls on the PR6030 word, each rejected.

Package: research/p104-6030-independent-audit/, base canonical main 721c03822c421b7f7fcbb2598993b947c4d88419 (work ran at a2b4b039; inputs unchanged).

What is checked

• Sources. PR6030's pinned PR5907 word and PR5827 archive are byte-identical to canonical main's PR5907 word and PR5907's pinned archive. PR6030's construct.py, applied to the canonical PR5907 word, regenerates the shipped word payload.
• Own exact replay of the PR5907 and PR6030 words: MOVE nesting, common charts, dyadic coefficients, owned lifetimes, all 799 endpoint rows and helper nesting. The paid and helper histograms equal each author's local block.
• Structural delta PR5907 → PR6030. Only program, initial and final change. There are exactly the three declared boundary changes and exactly the six declared itineraries (115, 127, 470, 704, 713, 736). INIT, CLEAR, COPY and all other MOVEs are unchanged as multisets. Paid delta −2H1+H2−2H3+H4+H6−H7+H8, helper delta −H1+H2−H3+H6−H7+H8, so the per-LOCAL margin change is H1 + H3 − H4, the same as PR5907's gain. 22,803 events. The rewritten ADD/SCALE events are checked only through their effect in the exact replay, not against a declared edit.
• Own invoice recount from the 51 physical caller events and from the closed formula: 66,834 paid / 13,444 productive, surplus 760. This equals complete_cost.
• Dense all-column native check with absolute phase (PR5548 interpreter). The 7 maps new in PR6030 are computed here and pinned in NEW-MAPS.json. The other 3,156 equal Matarener's FULL-CHECK digests. --full recomputes all 3,163 maps (PASS, 4 min 14 s on 8 cores).
• Own 80-digit funding engine. The raw crossing (0.00312581155791, 0.00312581155792) equals the author interval. The tested saving is positive, the tail ratios hold and the crossing is unique on [0, .0034].

Reproduce

python3.12 -B research/p104-6030-independent-audit/verify.py                  # serial, ~1 min; passes on 3.10-3.14
python3.12 -B research/p104-6030-independent-audit/verify.py --full --jobs 8   # all 3,163 dense maps

Scope and remaining obligations

• The five inherited framework assumptions remain.
• The compiler threshold and the 54+4 outer conditions are checked only by the author's chain. The maintainer ran it; it is not rerun here.
• The helper boundary-chart freedom (713, 736, 115) is inherited from the P104 line.
• Layers 1–4 reuse Matarener's audit functions, so this is independent of the PR6030 author, not of that audit.
• The selected record, already PR6030 at 721c0382, is unchanged. No merge or frontier change is requested.

Credit

DreamingOfClouds (PR6030, PR5907), Ali Romman (PR5827 chain), Mason Hale (PR5811), Pierre-Loic Doulcet (PR5565 lineage), Matarener (PR5960, the PR5827/PR5907 independent audit, reused functions and digests), Chris R / Clear1nsight (PR5548 dense interpreter) and Douglas Colkitt with OpenAI Codex (PR6030 portability repair).

Claim: #1166 (comment)

Workflow feedback: the coordinator asked for scripts/choose_research_task.py, but AGENTS.md on main says to use choose_current_research_task.py for a fresh assignment. I used the latter. The census-pruning-audit card drawn alongside this one already has about 47 packages under its task_id. A coverage count in the picker output would have saved a few minutes.

🤖 Generated with Claude Code

[JosieElliston]

Open question 1 here (recompute the compiler threshold and outer conditions without the PR5827 pricing.py lineage) is addressed in #6109. The compiler half is complete. For the outer half, the nine savings and 24 of the 54 strict checks are bound to written sources, and 30 strict and 4 functional checks remain open. No discrepancy was found.