Independent verification: supplier-admission premises of the #2265 substitution traced into PR2260 (0 FALSE / 0 MISSING-PROOF; no new kappa)

[thakrarsagar]

Status: Independent verification (adversarial audit / research note). No new kappa and no construction. The selected frontier is not changed.

Summary

#2265 (DreamingOfClouds) applies the PR1673 capped phase-cell fusion outer to the PR2260 selective-borrowing supplier (Dugongue). It marks several substitution items as "inherited from PR2260; not re-derived" and asks for an independent check. This PR adds research/pr2265-supplier-admission-audit/, which lists 15 supplier-admission premises of that substitution with file:line pins in #2265, the vendored PR1673 review and PR2260. Each premise gets a class, and the finite ones are re-checked with new stdlib code that imports and executes no code from either PR.

Result: 9 COVERED, 1 TRUE-unencoded, 5 INHERITED-traced, 0 MISSING-PROOF, 0 FALSE. The audit hypothesis was that some premise #2265 marks as inherited has no supporting statement or check in PR2260. It is refuted for the shipped data.

What the independent checks establish:

• Call closure (A1). Only its namespace/metadata part is TRUE-unencoded. Event closure and the owner-row facts are already enforced by PR2260's local/check_trace.py (via local/verify.py) and owners/verify.py, which Outer retune of #2260: #1673 capped phase-cell fusion outer on the selective-borrowing supplier, conditional κ = 0.00168341258854… #2265 reruns (Outer retune of #2260: #1673 capped phase-cell fusion outer on the selective-borrowing supplier, conditional κ = 0.00168341258854… #2265 verify.py:43-44).
  • Both orientations of the local word (21,065 forward and 21,381 backward events) touch only the call's 660 public arrays and its 76 owned private arrays (re-confirmed; PR2260's replay already enforces this).
  • Checked only here: the 44 warps equal PR2260's WITNESS pairs; the public arrays are exactly the 704-register namespace minus the 44 removed aliases; the owned private arrays lie outside it; handoffs occur exactly once on each borrowed source.
  • All 44 borrowed sources are inputs of the same call. 43 are helpers. The other is the public data input X_73, used for removed alias 653.
• Restoration (A2). A new exact replay over a free group of reference-change symbols with rational coefficients, run on both orientations, shows:
  • every column ends at the declared shear;
  • all private arrays end at zero;
  • in the forward orientation all 44 borrowed sources return to their inputs;
  • in the backward orientation 43 do, and X_73 ends as the declared shear column X_73 − Y_185.
• Owner recount (A3/D2).
  • 16,128 pending children are reconstructed: 387,072 = 254,088 hosted + 132,984 residual.
  • The residual histogram equals PR2260's.
  • 72 × local + residual equals the SPEC profile (mass 520,704, deficit 2,560).
  • The 31,392 rows form disjoint exact prefixes, and none sits on its call's active data family (PR2260's owners/verify.py checks the same facts).
• Constants (B2/B3). 7,745 families, stock bit length 2,641, row 16,467 < 20,161, and C_0 bits 200,012 < 210,000.
• Endpoints (C1, D3, D4).
  • The bit endpoint c is byte-identical to base e8-bit-p111-tuned. The conservative bit moment at c is 1 − 1.6e-24.
  • Complex moments are computed with 320-digit decimals: native 1 − 8.39e-12 and conservative 1 − 2.88e-200 at C, and conservative 1 + 3.16e-202 at C + 1e-200.
  • In exact rationals, q < a < b < c and kappa = C − 1e-246 − 1e-222.

Observations (none of them a defect):

1. The X_73 data-input borrow. It comes from PR2226's 80-alias witness and is within PR2260's stated model (PROOF-LOCAL.md:22 treats all 660 public/data/helper inputs as arbitrary). Only the helper-specific wording in the prose is imprecise.
2. Guard units are a lineage-wide question, not a PR2260 gap. Both self-consistent readings pass: fixed-bit 248 < 512, and per-lambda 244λ < 512λ.
   • Mixed readings would need λ ≤ 8 for the literal prefix, λ ≤ 2 for the relaxed bound and λ ≤ 4 for the scaled row. The selected Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 has the same issue: λ ≤ 4 for both its row and its prefix.
   • The one PR2260-specific point is that its statement "Payload512 is a per-lambda magnitude coefficient" has no located retained-lineage source.
3. Endpoint admissibility (D5) is a disjunction. C needs either the surcharge envelope to be adequate for the PR2260 cover, or the PR559 native-complex license to apply to it. Neither is proved for this cover; it is inherited.

Reproduce

Base commit: 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2. origin/main was unchanged at submission.

git fetch origin pull/2260/head pull/2265/head
python3 -B research/pr2265-supplier-admission-audit/verify.py

Expected output: PASS_PR2265_SUPPLIER_ADMISSION_AUDIT controls_rejected 15 of 15. It runs in about 3 s with stdlib only, and the full receipt is compared with EXPECTED.json.

• If a PR head has moved, the verifier falls back to fetching the pinned SHAs. Offline replay needs the pinned commits to be present already.
• Tested under Python 3.14 only. The registered 3.12 runner was not available locally, so CI will be the first 3.12 run.

Pins:

• PR2260 head fde6b107e26a14da5bac5b66913c914708ea6901 and PR2265 head 49ce38fee91f33edb54645724e7be9530f5631f6. 18 blobs are SHA-256 pinned in verify.py.
• Three base-tree files are pinned in SOURCES.json: the e8-bit-p111-tuned EXPECTED, the direct-delivery504 PREFIX-AND-INVOICE, and scripts/verify_frontier1094.py.
• The copy of PR2260 vendored in Outer retune of #2260: #1673 capped phase-cell fusion outer on the selective-borrowing supplier, conditional κ = 0.00168341258854… #2265 is byte-identical to PR2260 head.

Other checks run:

• python3 -O -I -B .../verify.py: PASS. Checks use need(), not assert.
• scripts/research_package.py --check-only: PASS_PACKAGE_METADATA_AND_PINS.
• python3 -B -m unittest discover -s tests -p 'test_research_*.py': 65 OK, 2 skipped.

Negative controls (15, all rejected by the intended check)

• Borrowed source: moved to a removed alias, or to register 2000 outside the call.
• Events: an ADD reads a removed alias.
• Replay:
  • forward, one logical monomial warp dropped;
  • forward, one helper ADD coefficient doubled;
  • backward, one dual monomial symbol inverted.
• Owner rows: a block overlaps the previous one, or a row sits on its call's active data family.
• Halving: a mass-preserving child of rank 52.
• Invoice data: guard coefficient set to 245, or row coefficient set to 16466.
• Endpoint: C + 1e-200; kappa + 1e-240.
• Bit endpoint: set equal to b, or raised by 1e-18.

Scope, inherited assumptions, unresolved obligations

• Scope.
  • This is a finite data audit at one point: Outer retune of #2260: #1673 capped phase-cell fusion outer on the selective-borrowing supplier, conditional κ = 0.00168341258854… #2265's final point and PR2260's shipped word. It does not prove the capped-fusion theorem, the all-size compiler or any multiplication theorem, and it removes no hypothesis.
  • It covers the PR2260/Outer retune of #2260: #1673 capped phase-cell fusion outer on the selective-borrowing supplier, conditional κ = 0.00168341258854… #2265 line (κ ≈ 0.0016834). That is not the leading open conditional claim: Add 87 reusable Lean theorem APIs and conditional κ = 0.00168357 #2284 and 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/Improve conditional κ to 0.0016965151242727336 with a pure 44th dirty alias #2309 are not covered.
  • Add 87 reusable Lean theorem APIs and conditional κ = 0.00168357 #2284's point was skipped because its ENDPOINT schema and profile differ.
• The replay is a new implementation of PR2260's event model, not an independent model. Native Clifford lowering, phases, the λ-fold lift and the physical charts are not rechecked. The 63-bit ideal local bound and the 160 time windows (A4) are taken from PR2260.
• Inherited and still open:
  • the PR1673 hypotheses: common-frame/mutable-source realization, whole-call hosting, the C_1 = 1 semantic guard, all-size fixed-tape/prime-packing, and paid equal-chunk interchange;
  • PR2260's all-size compiler and retained invoice conventions;
  • the payload unit convention (B1);
  • endpoint admissibility for the PR2260 cover (D5).
• No search was performed.

Overlap, predecessors, credit

• Outer retune of #2260: #1673 capped phase-cell fusion outer on the selective-borrowing supplier, conditional κ = 0.00168341258854… #2265 (DreamingOfClouds), Improve conditional κ to 0.0016805 with selective dirty borrowing and physical owner repacking #2260 and Add 87 reusable Lean theorem APIs and conditional κ = 0.00168357 #2284 (Dugongue), and Research: 80-helper anticipatory borrowing component and exact price rejection (no new kappa) #2226 (Cinna Davis / 1337isnot1337; origin of the X_73 witness pair).
• Conditional kappa = 0.001677160151714381 via capped phase-cell fusion #1673 (Vlado-V), Conditional κ = 0.001679754956125544521335670993920702576998311438502580295460327520162864548260325638826305656132844: recertified 1e-60000 supplier endpoint, fixed-grid-maximum kappa C - 1/F - B at 1e-60080 slacks (PR2180 / PR2161 retune)… #2220 and Conditional κ = 0.001679745508802299: gauge-free (fractional-shift) interiors, structured boundary solves and radix-distributed sparse-union schedule on PR1673's capped phase-cell fusion outer, pivot-566+554 T74 supplier #1758 (jeremykothe2), Conditional κ = 0.001679754788150048 from fine source-prime padding #1786 (chafreaky), Conditional κ = 0.0016797549561255445213356709939207025769983114385025802954603275201628645482603256388263...: fixed-grid ceiling over PR2161 #2180 and Conditional κ = 0.00167975495612554452133567099392070257699831143850258029546032752016286454826032563882630565613284: tighter free epsilon over PR2044 #2161 (dataisfire).
• The e8-bit-p111-tuned and direct-delivery504 packages.

Concurrent work: #2286 audited T74 exponent side conditions and explicitly left supplier admission unchecked. It was opened from the same GitHub account by a different concurrent sprint session. At submission, #2260 and #2265 had no comments or reviews, and no PR, discussion or sibling card covers PR2260 supplier admission.

Sprint: track 60-74 (adversarial mathematical audit), seed 67. Claim: #1163 (comment)

Before submission, three internal reviews looked at reproducibility, mathematical scope and provenance. Their fixes are included: the guard-unit finding was rescoped, A2/B2/B3 were reclassified, D5 was restated as a disjunction, a tautological control was replaced and a backward-orientation control was added.

Correction after independent re-audit (2026-10-15, commit 00de1b48): A1 had been described as the only premise neither #2265 nor a PR2260 check it reruns covers. That was wrong for event closure and the owner-row facts, which PR2260's local/check_trace.py and owners/verify.py already enforce. The TRUE-unencoded label is now limited to A1's namespace/metadata sub-facts, and the owner-row facts moved to A3. TASK.json was updated (it still said 13 controls and 6/4/5), and the D5 engine-comment provenance was corrected. Counts, verify.py and EXPECTED.json are unchanged.

Contributor: @thakrarsagar, with substantial AI assistance from Claude (Anthropic, Claude Opus 5.5) via Claude Code, run in an automated sprint loop. Apache-2.0.

🤖 Generated with Claude Code

[thakrarsagar]

Correction after independent re-audit

What was wrong

1. A1 (call closure) overstated what was new. The package labelled it TRUE-unencoded, and the PR body called it "the only premise neither Outer retune of #2260: #1673 capped phase-cell fusion outer on the selective-borrowing supplier, conditional κ = 0.00168341258854… #2265 nor a PR2260 check it reruns covers". Most of A1 is already enforced:

   • PR2260's local/check_trace.py runs inside local/verify.py, which Outer retune of #2260: #1673 capped phase-cell fusion outer on the selective-borrowing supplier, conditional κ = 0.00168341258854… #2265 reruns at verify.py:43. It keeps state only for active_registers and the arrays its own COPY/zero-buffer events create, so any event outside the call fails.
   • local/verify.py:91-92 and expected-local.json pin 660 public columns and 76 private arrays.
   • An unmodified check_trace.run raised KeyError on each of these mutations: an ADD reading removed alias 597, an ADD reading register 2000, an ADD writing 597, and a handoff on 597.
   • The owner-row facts are checked by PR2260's owners/verify.py at :87, :88-89, :95, :106 and :113, which Outer retune of #2260: #1673 capped phase-cell fusion outer on the selective-borrowing supplier, conditional κ = 0.00168341258854… #2265 reruns at verify.py:44.

   The package was also inconsistent with itself. A2 was COVERED through the same rerun, and INVENTORY defined TRUE-unencoded differently from the README.

2. TASK.json was stale. Its evidence field said "13 of 13" controls and 6/4/5 classes. The verifier prints 15 of 15, and INVENTORY gives 9/1/5. Its open questions still used framings the README had withdrawn: the lambda question as a PR2260-specific gap, and the claim that "the vendored engine itself labels the envelope only conditionally portable". scripts/find_overlap.py indexes this file.

3. Minor.

   • The MISSING-PROOF/FALSE row sat outside the README table, so GitHub rendered it as raw text.
   • D5 said the engine comment "only conditionally portable to the new cover" referred to frame-domain-ceiling's cover. In fact it comes from the older PR361 / e8-r740-lineage engine; the byte-identical file has sha256 81cbc864….

What changed (one commit, 00de1b48)

• A1 narrowed. It stays TRUE-unencoded only for the namespace/metadata sub-facts that no rerun check binds:

  • the warps equal PR2260's WITNESS pairs;
  • the public arrays equal the 704-register namespace minus the 44 removed aliases;
  • the 76 owned arrays lie outside that namespace;
  • handoffs occur exactly once on each borrowed source.

  PR2260's replay takes active_registers as given. For example, it accepts the forward trace with COPY register 1364 renamed to removed alias 597. The shipped trace is hash-pinned, so this is about what the replay checks, not about the data.

• INVENTORY.json. A1 has new class_scope and covered_part fields. The owner-row facts moved to A3, with owners/verify.py line pins. The TRUE-unencoded definition now matches the README.

• Other files. TASK.json evidence and open questions now match the README. The README table row was moved, the D5 provenance was reworded, and the PR body was updated.

• Unchanged: the counts (9 COVERED, 1 TRUE-unencoded, 5 INHERITED-traced, 0 MISSING-PROOF, 0 FALSE), verify.py, EXPECTED.json and the title. There is no new kappa.

Reproduce (Python 3.14.4, run under nice -n 10)

• python3 -B research/pr2265-supplier-admission-audit/verify.py gives PASS_PR2265_SUPPLIER_ADMISSION_AUDIT controls_rejected 15 of 15 in about 3.4 s, and the receipt equals EXPECTED.json. The same result holds with -O -I -B.
• python3 -B scripts/research_package.py --package research/pr2265-supplier-admission-audit --check-only gives PASS_PACKAGE_METADATA_AND_PINS.
• python3 -B -m unittest discover -s tests -p 'test_research_*.py' ran 65 tests: all OK, 2 skipped.

(thakrarsagar; prepared with Claude Opus 5.5)