Research audit: the selected #1094 reuse replay discards its ancestry recheck; binding certified from committed archives (no new kappa)

[thakrarsagar]

Change and mathematical scope

Audit and supplemental certificate. No new κ, and no change to the selected
record, pinned verifiers, review documents or upstream/. Everything is in the
new package research/frontier1094-ancestry-binding-seed69/.

Finding (low severity). scripts/verify_frontier1094.py --baseline1042 is
the reuse mode the batch1099 review says it used locally. It runs
check_selected(ancestor, ancestor/'ancestor') and discards the return
value. It then compares BASELINE-BINDING.json only with the #1042
receipt's own hash. So nothing ties the rechecked PR799/PR741 ancestry to the
binding #1042 recorded. Any ancestry at the receipt's machine-local path
/private/tmp/batch952-baseline799 that passes check_selected is accepted.

Limits of the finding:

• The cold default make frontier1094-verify has no such gap.
• The Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 candidate stage reads only hashed Conditional κ = 0.001672188119205421: improved T16 owner packing on the seven-slot k8 word #1042 artifacts, so κ is
  unaffected.
• The Conditional κ = 0.001661957774647029: T21 whole-owner packing and fixed-family bound #890 and Conditional κ = 0.001661960547713947: T37 whole-owner packing and all-T bound #942 reuse modes do bind their ancestry. Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094's is the only
  current --baseline* reuse mode that does not.

Certificate. The committed integrated799-replay.tar.gz and
composed-replay-evidence.tar.gz archives, run through the unchanged
check_selected, rebuild a binding byte-identical to
research/community-frontier1042-audit/BASELINE-BINDING.json (22575ff2…).
That is the hash bound by the #1042 receipt bf12c6ed…, which the #1094 receipt
names. The selected record's ancestry binding can therefore be reproduced from
in-repo evidence alone.

Controls. Identity passes everywhere. Three variants pass check_selected
and the whole unchanged pinned reuse mode at κ = 0.001672660197076918, but are
rejected by the suggested one-line binding:

• timing: a re-execution with different wall-clock times;
• bit_result: a PR589 RESULT9.json field replaced;
• bit_supplier: the PR799 assembly names a zeroed supplier commit and blob.
  This is the semantically invalid fixture.

A PR799 κ tamper is rejected by all three checks.

Suggested maintainer fix, not applied:

require(check_selected(ancestor, ancestor/'ancestor') == read(baseline/'BASELINE-BINDING.json'),
        'Ancestry recheck differs from bound #1042 binding')

Remaining obligations. This is not a cold replay of #1042 or #1094, not
execution authentication of the archived logs, and not a proof of the
multiplication theorem. The inherited assumptions are unchanged.

Task and overlap

• Base and track: base dadfea6e. Seed 69 → adversarial mathematical
  audit; card supplier-boundary-audit.
• Scope: refined; the reason is in TASK.json. The traced obligation is
  the selected record's reuse-path ancestry binding, not a supplier inequality.
• Overlap: find_overlap.py and the open-PR searches found no PR on
  Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094's reuse binding. The nearest work:
  • Lucas Steinmetz's PR741 receipt-coverage audit (Research audit: PR741 binding accepts omitted replay evidence (no new kappa) #838) and the Tooling: source-versioned PR799 evidence coverage with fresh replay and corruption controls #865
    contract concern the PR741 binding's stage and artifact inventory, which
    check_selected now enforces. This package audits the layer above, where
    check_selected's result is used.
  • Tooling: opt-in parallel layer-end controls for the frontier1042 leaver chain, byte-identical receipts, 3.1x (extends #1059, no new kappa) #1151 (mine) is tooling for the Conditional κ = 0.001672188119205421: improved T16 owner packing on the seven-slot k8 word #1042 leaver chain only.

Validation

python3 scripts/research_package.py --package research/frontier1094-ancestry-binding-seed69 --check-only
python3.12 scripts/research_package.py --package research/frontier1094-ancestry-binding-seed69

• Results: PASS_PACKAGE_METADATA_AND_PINS, and
  PASS_FRONTIER1094_ANCESTRY_BINDING_AUDIT 22575ff2… on Python 3.12.15 and
  3.14.8 (macOS), in about 3 s using only the standard library. The verifier
  refuses -O.
• Evaluation layers: every variant goes through (1) check_selected;
  (2) the three pinned reuse statements, parsed from the pinned Git blob and
  executed with the real check_selected, with and without the fix; and
  (3) the whole pinned verify_frontier1094.py reuse mode in a scratch root.
• Stubs in layer 3: only two stages, the historical-source stage and the
  candidate stage, which copies the published Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 REPLAY/ASSEMBLY. All Conditional κ = 0.001672188119205421: improved T16 owner packing on the seven-slot k8 word #1042
  receipt, source, log and 457-artifact checks run unchanged.
• Robustness:
  • Before importing any checker, verify.py requires every declared live
    input to match its pin.
  • The audited verifiers are read from their pinned blobs when the live files
    differ or are missing, so the package stays checkable after the fix lands.
    Tested by applying the fix to a live copy.
  • The subprocess environment drops colour and safe-path variables.
• Pre-PR review: three independent reviewers (soundness, claim accuracy,
  CI/conventions) re-ran and attacked the package. Their fixes are included.
  The fixes cover self-pinning, declaring the six gauge44-frames-x files read
  directly (the manifest stands in for the rest), a hardened environment, and
  credit and wording.
• Not run: the cold make frontier1094-verify.

Credit

Credit to:

• Douglas Colkitt, for the Conditional κ = 0.001672188119205421: improved T16 owner packing on the seven-slot k8 word #1042/Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 composition verifiers, archived replays,
  binding receipts and batch1099 review;
• Rohan Arun's Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 and Jeremy Kothe's Conditional κ = 0.001672188119205421: improved T16 owner packing on the seven-slot k8 word #1042 constructions, with Hao Yan's T16
  controller;
• nightangelflowerwin-ops (Tooling: source-versioned PR799 evidence coverage with fresh replay and corruption controls #865), for check_selected and the versioned PR741
  coverage contract;
• Lucas Steinmetz (@ljste, Research audit: PR741 binding accepts omitted replay evidence (no new kappa) #838), for the receipt-coverage finding that
  contract addresses;
• Adem Vessell (Require complete bridge-control coverage in the PR741 binder #851), for the bridge-identity check used to rebind variants;
• Chafik Boukhalfa (Conditional κ = 0.001661957774647029: T21 whole-owner packing and fixed-family bound #890) and Jonathan Capalbo (Conditional κ = 0.001661960547713947: T37 whole-owner packing and all-T bound #942), whose ancestry-binding
  pattern is the comparison;
• the supplier lineage credited in the selected review.

AI disclosure: prepared by Sagar Thakrar with substantial Anthropic Claude
assistance. That included parallel explorer, judge and skeptic subagents, which
proposed this finding and tried to refute it. New work is Apache-2.0.

Open next questions

See the package README:

1. Should PR799's bit_supplier commit and blob be bound to the cited PR589
   source? The zeroed fixture passes check_selected. Blocker: pinned
   supplements need a maintainer rehash.
2. Should receipts record a portable ancestry_binding_sha256 instead of only a
   /private/tmp baseline path? Blocker: the receipt schema of the selected
   record is frozen.

Workflow feedback (optional)

The package generator and research_package.py --check-only made the metadata
quick. One friction point: the guide says to declare transitive inputs, but the
selected check_selected chain hashes all of research/gauge44-frames-x
through its manifest. Only a directory-level input or the manifest can be
declared, and this package declares the manifest.

[Mister-Meeseeks]

100-PR batch review at 02d279515ff1fc468ab99a19d89461f1b6391507.

Merged to main, preserving the original contributor commits and package attribution.

Reviewed scope: Source-bound adversarial audit of an unchecked ancestor-binding return in the reuse path, with three mutants. End-to-end stages explicitly stubbed; cold-path coverage is unchanged and no exploit or frontier rejection is inferred.

The pinned scoped command passed; the batch ledger contains its command, input hashes, environment and log. This incorporates the research/tooling contribution and does not change the selected κ or discharge the inherited multiplication-theorem assumptions.

Batch review, attribution and evidence. Thank you for the contribution.