Skip to content

Verifier: authenticate selected invoice references before portable comparison - #2059

Merged
Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
adamgall:research/round3-05
Oct 15, 2026
Merged

Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
adamgall:research/round3-05

Conversation

@adamgall

@adamgall adamgall commented Oct 14, 2026 •

Copy link
Copy Markdown
Contributor

Change and mathematical scope

Verifier engineering / source-bound tooling, not mathematical admission. No new supplier, paid profile, physical claim or κ.

This closes PR #2040's explicit invoice-reference handoff with an opt-in comparison of three selected PR1094 generated artifacts: global/audit/PREFIX-AND-INVOICE.json, ADMISSION-CONTROLS.json, and ASSEMBLY.json.

The source-bound census of all 105 reviewed strict outputs finds exactly two literal invoice digest references, both input_sha256/fresh_invoice. A format-only invoice transformation changes its raw SHA-256, so PR1824's retained comparator flags both freshly updated references. The new comparator authenticates both raw links on each side before normalization, then compares the complete canonical invoice bytes and complete rebound consumer bytes. Every other field, including other digests, remains exact. No global hash, timing or path exemption is introduced.

Evidence: 3,041 tree nodes retained; seven format positives; 6,076 nonroot omission/replacement cases, including 5,296 changed invoices with BOTH raw links consistently updated; 57 additional controls; 6,133 total rejections. An independent typed-tree oracle represents edges as role tuples rather than reusing the primary serializer/digest.

Critical proof step: matching raw references alone never establishes payload equivalence. The complete canonical invoice itself is byte-compared; synchronized semantic mutation cannot be concealed by repairing its raw edges. For equal invoices, authenticate-then-rebind changes only the two representation-dependent edge values; full consumer comparison checks every other field.

Only nine new package files under research/selected-invoice-reference-rebinding/. No frozen caller, source archive, shared index, or selected pointer changes. This is a reusable subgraph predicate, not full-output equivalence, a production false-acceptance claim, or a supplier replay. PR2040's Linux observation is credited, not presented as a Linux run performed here.

Task and overlap

  • Base: e112d71c8d7bb32435debcc54df932593dd8223a.
  • Commit: 323703a83640f4c104469bf29db18d7d579cb587.
  • Independent one-shot chooser: a2e5a793b5e1a2555a2466367171c2ea, verifier-engineering; complete cards in SELECTION.json.
  • Task: portable-replay, scope refined to selected1094/invoice-two-consumer/raw-authentication-before-typed-rebinding; TASK written before substantial computation.
  • PR2040, d31c1418c0c23994d1444dbd9d356b9a8238fd5e: typed/path comparator audit, explicitly leaves this raw-reference handoff open. No pending code dependency.
  • PR1824, ad5b1248f25a71983b20617e2f3a80be22f03990: retained before comparator, merged at its pinned head.
  • Selected source PR1094, 0fc329d0bec93ca7cd288eaea1295172e8789810: authenticated reviewed archive and committed certificate.
  • Earlier selected-generated-identity-audit covers seven different local artifacts, not invoice consumers. pr695-digest-audit's warning about whole-file hash substitution motivates synchronized semantic negatives here; its unavailable historical CI payloads are not claimed resolved.
  • Scoped claim.

Shared snapshot: 329 open PRs including drafts and 692 discussions including comments/replies, complete pagination on 2026-10-14 at 18:11/18:12Z. Local/merged predecessors, exact rebinding query and targeted live PR heads checked again before submission. No-match is not novelty; later changes may be absent from the cache.

Validation

CPython 3.12.12, macOS arm64, standard library. Observed final serial batch: 31.45 seconds including launchers.

uv run --python 3.12 --no-project python -I -B research/selected-invoice-reference-rebinding/verify.py
uv run --python 3.12 --no-project python -I -O -B research/selected-invoice-reference-rebinding/test_compare.py
uv run --python 3.12 --no-project python -B scripts/research_package.py --package research/selected-invoice-reference-rebinding --check-only
uv run --python 3.12 --no-project python -B scripts/research_package.py --package research/selected-invoice-reference-rebinding

Observed: direct audit PASS; 8 focused tests OK; metadata/source pins PASS; registered replay repeats audit and tests and ends PASS_REGISTERED_RESEARCH_PACKAGE research/selected-invoice-reference-rebinding outcome=completed.

PASS_SELECTED_INVOICE_REFERENCE_REBINDING files=105 nodes=3041 edges=2 positives=7 field_mutations=6076 synchronized_invoice_mutations=5296 negatives=6133 no_new_kappa

Controls include one/both stale links, identical malformed graphs (no byte shortcut), misbound owner-audit references, duplicate keys, unsupported decimal/nonfinite numbers, complete list/shape/type retention, all27 Boolean-to-integer substitutions, CLI exit statuses, and an immutable gzip-header mutation that leaves the decompressed tar equal but must fail the raw source pin. RESULT records deterministic census and mutation transcript hashes.

Not run/claimed: full cold supplier replay, Linux/Windows, selected adoption, remote CI, formal kernel acceptance, or inherited all-size/compiler/hosting/precision/analytic discharge. Exact integer JSON grammar is deliberate; decimal/exponent number tokens fail closed rather than use binary64. Source/reference authenticity is distinct from mathematical truth.

Credit

Josie Elliston (PR1824), commdata2338 (PR2040/handoff), Maxime Fleury and Antoine Olivier (reproducibility convention), Rohan Guliani and Joshua Snider (PR695 diagnostic), Adam Gall's earlier local identity audit, and Rohan Arun, Jeremy Kothe, Hao Yan, Douglas Colkitt and the retained selected supplier/compiler lineage.

New work Apache-2.0; source notices and frozen bytes unchanged. Prepared for @adamgall with substantial OpenAI Codex (GPT-6 Astra) assistance in design, implementation, experiments, tests and writing. No independent human review claimed.

Open next questions

See package README/TASK. A future caller can compose this exact three-node predicate with a separately reviewed full-output comparator and explicit per-field time/path policy. This would remove the demonstrated invoice/reference portability obstruction without dropping coverage. The next step is the complete caller replay; adoption/source-pin review requires a maintainer decision. The rest of the graph is not certified by this package.

Research claim and completion record

Status: completed. The claim record is preserved here so this PR is self-contained. Historical coordination links in the committed research metadata refer to this record; the separate coordination discussion is being removed.

DONE — independent seed a2e5a793b5e1a2555a2466367171c2ea (verifier-engineering / portable-replay). New PR #2059; full commit 323703a; base e112d71. The selected PR1094 invoice and its two fresh_invoice consumers now have an opt-in authenticate-before-rebind comparator. Exactly 2 direct references in 105 reviewed strict outputs; 3,041 tree nodes preserved; 7 portable positives; 6,076 field mutations including 5,296 changed invoices with BOTH raw links consistently updated; 6,133 total negatives reject. Eight focused tests and registered CPython 3.12.12 replay pass; parent independently replayed the package and approved publication. This closes PR2040s named invoice-reference handoff, not its typed/path audit. Full canonical invoice bytes and both complete rebound consumer payloads are compared; no blanket digest/time/path exemption. Scope remains only three nodes: no production adoption, full-output equivalence, mathematical admission, new supplier, physical claim, kappa improvement or CI claim. The upstream Linux observation is attributed to commdata2338, not a platform run here. Credit Josie Elliston PR1824, commdata2338 PR2040, Maxime Fleury/Antoine Olivier, Rohan Guliani/Joshua Snider, Adam Gall local identity audit and Rohan Arun/Jeremy Kothe/Hao Yan/Douglas Colkitt selected lineage. Substantial OpenAI Codex assistance; Apache-2.0. Reproduce: uv run --python 3.12 --no-project python -B scripts/research_package.py --package research/selected-invoice-reference-rebinding. Original claim expiry 2026-10-14T20:35Z; completed before expiry.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants