Independent verification (no κ): #2530's source-partner words exact through cleanup and endpoint; stock, endpoint enclosure and κ composition reproduce; no defect found

[thakrarsagar]

Summary

Independent verification, no new κ. This PR adds research/t74-2530-source-partner-audit/, a source-bound
independent check of open PR #2530 by zxv (head b4012098; T74 source-partner lifetimes; conditional
κ ≈ 0.0017137379).

The checks use only the standard library. They read 38 pinned git objects from PR2530, PR2462, PR2435, PR2269 and
PR2474 as data, and never import or run PR code. No defect was found; every PR2530 number re-derived here
reproduces.

What this adds. Four concurrent audits cover parts of #2530: #2664, #2661, #2656 and #2669. As of 2026-10-15
22:26 UTC (heads re-confirmed), none of them checks the following for #2530:

1. Both words end to end. They are exact on all 646 dirty inputs, forward and dual, through the 4,529-item
   cleanup and the target endpoint. A dual-only mutant shows that the dual replay adds coverage: the forward replay
   accepts it and the dual rejects it.
2. Physical stock 66,814. This is a consistency check: this author contributed to the stock rule.
3. The endpoint and κ. A directed-rounding enclosure accepts C = 0.001713737921547504346752 and rejects
   C + 1e-24. The grid conditions for κ = C − 1e-70 − 1e-46 hold, given the author-checked outer inequalities.
4. The verifier overrides, diffed against PR2462's verifier.
5. R-g owner charts: 7,023 charts.

It also corroborates those audits with another implementation: the handoffs and gates (#2664), the profile recount
(#2656), the windows (#2661, #2669), and the fill R-a..R-e and idle histogram (#2661).

Fill order dependence. It records how #2530's fill depends on its call order:

• Under PR2462's order, 60,541 roles leave their windows.
• Of the 665 adjacent swaps of calls from different replicas, 238 keep the fill legal.

Findings for the PR2530 author (non-blocking):

• The 24-digit outer endpoint is stated only in the sibling t74-owner-consolidation README. It forgoes about
  7.0e-25 of κ.
• The prefix_l1 denominators are uncharged by prefix_bits. This is inherited and well within headroom.

What later claims can use (compared at heads; heads re-confirmed 2026-10-15 22:26 UTC):

• Word checks: Conditional κ = 0.0017137844912361143: repacked owner fill on #2530's source-partner words #2571 and Conditional κ = 0.001713879320009017: owner evacuation on #2530's source-partner words #2575.
• Order and window checks: Conditional κ = 0.001713879320009017: owner evacuation on #2530's source-partner words #2575, Conditional κ = 0.001714252039260245: swapped helper labels for lenders 80 and 88 on #2530's source-partner words, with an evacuated owner fill #2588 and Conditional κ = 0.0017148565406944334: pre-life hosting and a frame-refit alias on the #2588 T74 words #2679.
• Conditional κ = 0.0017137844912361143: repacked owner fill on #2530's source-partner words #2571's head uses an order shifted by s = 4; its PR body is stale.
• No later claim reuses Conditional κ ≈ 0.0017137379215475043 from source-partner lifetimes and physical rehosting #2530's fill. Conditional κ ≈ 0.001717794984118634663827: two hostee moves and seven retired-lender aliases on #2751's joint twelve-source words #2802 (opened 22:23 UTC) was not compared.

Not re-derived (author-checked):

• the native K_op identity for the 12 lenders applied in both the prelude and the handoff;
• the 63 outer inequalities;
• strict global admission;
• the binding of the words to the receipts' canonical_word_sha256 (a hash of a file the native replay produces);
• the receipts' prefix_l1 values.

Controls:

• 148 negative controls, 99 distinct. Each must be rejected by its named rule, and the run asserts that this rule
  fires.
• 3 positive controls, on PR2462's words and endpoint.
• 62 mutations of the checker itself (local-only): every one makes the run fail.

Reproduce

python3 -I -B research/t74-2530-source-partner-audit/verify.py --fetch

It needs about 2 min, about 1.1 GB of memory, git ≥ 2.29 and network access for --fetch. --fetch writes only the
fetched objects and refs/audit/prN. Expected final line:

PASS_T74_2530_INDEPENDENT_AUDIT words=2x646-inputs-fwd+dual profile={"1":2369,"2":632,"3":320,"4":118,"5":26,"7":8} roles=281052 idle={"1":21976,"3":6662,"5":440,"8":114,"16":172,"32":3,"48":26488} physical=66814 charts=7023 C=0.001713737921547504346752 kappa=C-1e-70-1e-46 negative_controls=148/148-rejected distinct_controls=99 positive_controls=3 pinned_inputs=38

python3 scripts/research_package.py --package research/t74-2530-source-partner-audit --check-only gives
PASS_PACKAGE_METADATA_AND_PINS. The registered command has been run only on CPython 3.14.4, not on 3.12.

Task, scope and overlap

Field	Value
Card	independent-moment (independent-verification), refined to PR2530
Assignment seed	63 (adversarial-audit track)
Priority revision	5
Base	db3edc00f732806f663bfb9d4000567f18b17386
Declared milestone	audit
Family expansion decision	none; no further fill, order or word optimisation proposed
Canonical thread	Discussion #2606, the existing PR2530 audit thread (status post); pointer on #2530

The overlap with #2664, #2661, #2656, #2669 and rcfc-jreed's audit of #2621 is described head by head in the README.
Their code was read, not replayed.

Frontier. The selected frontier is unchanged (κ = 0.001672660197076918, #1094). PR2530's disposition is unchanged.
The package is a finite, source-bound check in the lineage's free-operator, frame-climb and owner-rule model. It is not
a native re-execution and not a multiplication theorem.

Independence and conflicts

Vendored checkers. The package vendors two of this author's earlier checkers, pinned by blob and sha256:
words_check.py from #2315 and fill_check80.py from #2387. #2315 is now on main; its copy differs only in a
corrected docstring credit.

Credits to this author:

• PR2530 credits Independent verification (no new κ): record candidate #2269 T74 borrow43 — words, 437-helper fill and moment root reproduce; stale evidence, unpriced buffers, X2 open #2315 and Audit (no κ): exact owner-fill optima on the T74 handoff line — #2374's fill is optimal over all fills; #2435's fill leaves +5.8e-10 (pattern optimum certified) #2442, and cites Independent verification (no κ): #2447's owner fill is exactly the certified optimum for #2435's words; the same optimum in 4 role moves (fewer charts) #2466.
• PR2462 credits Audit (no κ): exact owner-fill optima on the T74 handoff line — #2374's fill is optimal over all fills; #2435's fill leaves +5.8e-10 (pattern optimum certified) #2442, Independent verification (no κ): #2435's words, 289,044-role owner map and endpoint reproduce; its change is frame-only (24 raised entrances), entrance delivery unchecked #2446, Audit (no new κ): side-condition coverage of the pending T74 outer at #2044 — 554 written conditions, none load-bearing false; or-True and g² review items resolved #2286 and Research audit: the selected #1094 local programs satisfy Lemma G's extended-label H1 (HPE-aware check, 0 violations); the selected chain runs no tag check (published checker needs a cold-HPE patch; the frame-only replay accepts three adapter mutants) (no new kappa) #1964.

None of these is used as a check here, except that the stock rule (Dugongue's #2309, with #2315's cleanup-buffer
finding) is checked and labelled as a consistency check.

Credits: zxv (#2530); jeremykothe2 (#2435, #2462, #2269, #2474); Tekalign/tekuuu (#2436);
Vlado Vojdanovski (#2322); Cinna Davis (#2226); Dugongue (#2309); Hao Yan (#858); Rohan Arun (the #1297 lineage).
Concurrent audits are by commdata2338 (#2664, #2661, #2656, #2669, Discussion #2606) and rcfc-jreed (#2621).

Prepared by Sagar Thakrar with substantial AI assistance (Claude Code, Anthropic, Claude Opus 5.5). Three rounds of AI
review shaped the overlap statements and the controls. No independent human review is claimed. Apache-2.0.

🤖 Generated with Claude Code