Skip to content

feat(shipping): observe legacy Gate A checksum identity (S2-1a) - #572

Merged
raylei50653 merged 5 commits into
mainfrom
feat/549-legacy-native-identity
Oct 11, 2026
Merged

raylei50653 merged 5 commits into
mainfrom
feat/549-legacy-native-identity

Conversation

@raylei50653

@raylei50653 raylei50653 commented Oct 10, 2026 •

Copy link
Copy Markdown
Owner

Implements #549 S2-1a legacy native N-T3 observability under owner authorization and accepted contract clarification, from main@15e819502e4d5ea581792a3a16f97857435fc585.

Gate A records the six actual bindings and caller metadata locations. Metadata observations hash the same buffer parsed by Gate A. Missing expected hashes remain null/unchecked. Only all existing Gate A checks passing promote checksum_matched; RunCompletion seals that record through later completion/failure writes. Report v4 and journal v2 share the identity, while the reader preserves historical v3/v1 null identity and rejects malformed or inconsistent evidence.

Implemented

  • Legacy checksum identity and per-binding diagnostics; unchanged Gate A refusal policy and Gate B loader.
  • Journal v2, report v4 (including existing model-root context), reader cross-checks and historical compatibility.
  • CPU/CLI positive and negative controls with actual rejection reasons, no-CUDA observations, exact parsed metadata buffers, and real SIGKILL windows.
  • Canonical CC-536-01-02 clarification linked to the accepted owner decision.

Verified

  • CPU standalone CTest 8/8; true frozen-model Gate A 3515 checks/0 failures; completion 267 checks/0 failures; 13 actual writer/reader cases.

  • Reader 174 passed, zero skips, including current default-build writer integration. Independent source review: 226 passed, zero skips across reader, config-loader and measurement-surface suites.

  • New local binaries: 23 CLI controls passed; native GPU controls 6/6 including serial and double-buffer.

  • Complete local pre_push at 5b5e3b8d789c0d18845738f7660d155b5907f176: exit 0; 5575 passed, 52 skipped, 105 deselected, 5 xfailed. The final follow-up changes three documentation lines only; runtime/reader/tests are byte-identical to the GPU replay source. Final doc links, strict doc structure and attested static coordinate check passed.

  • Independent Codex review after submission at final head 2b2b45aa4a995fa2d1047fd1142b9445c57f3a81: no findings. It independently checked the 11 actual native run identities/run IDs/report hashes and source/binary/coordinate evidence.

  • Shipping measurement-surface and link checks passed; the measurement binary was refused for the intended measurement symbols.

  • Final-head PR CI 8/8 SUCCESS, head 2b2b45aa4a995fa2d1047fd1142b9445c57f3a81: build, lockfile, shipping-config-loader, ruff, mypy, contracts, pytest and lint-typecheck. CI run, C++ build run; both workflow head SHA values match.

  • Runtime-coordinate republication was independently reviewed at #573 head fe4b83f1fd08842b361e918f5ac9fb5e23f2851c, with no findings, and locally incorporated into this branch. The five-axis candidate and actual probe witnesses were independently recomputed/checked; old canonical archived unchanged, equivalence stays unproven. chore(provenance): review S2-1a runtime coordinate republication #573 has no CI checks and grants no remote-CI claim. Both PRs remain unmerged.

Unverified

  • Fresh GPU replay completed on clean source df475bb0b633577e26a3b8e39b6a2e47e50525de: four positive modes and seven negative controls each executed seven sequences / 5316 frames. Frozen V5 qualification refuses NVIDIA driver 617.42 versus required 616.92; formal qualified parity stays UNRESOLVED, controls standard caught=false. Fresh raw positive detector/MOT/graph comparisons equal where run; raw controls show the expected section DIFFERS. Positive native runs completed with identical journal v2/report v4 checksum identities. This is local diagnostic evidence, without a qualified EXACT/CAUGHT or package claim. Historical EXACT is not verification of these binaries.
  • Installed package is unchanged and unverified for S2-1a; no re-pin or public distribution.

Deferred

S2-1 full model-bundle mode, S2-2, S2-3, #537 completion/state work, #541 architecture/check coverage and #547 distribution obligations remain open. No allowlist, dual roots, require-identity policy, Gate B load_verification/loaded_buffer, TOCTOU repair or package split. This PR does not close CC-536-01-02 or authorize S2-1.

Keep this PR unmerged. Independent Codex review of the exact final head is complete; owner authorization to merge that head is still required. CI success grants no merge authorization. No #536/#549 implementation status or #550 checkpoint is updated before authorized merge and green main CI.

Versioned evidence summary: native_identity_observability_549.md. Local raw GPU evidence results/549_s2_1a/gpu_df475bb0b633/summary.json; SHA256SUMS digest 2b0ddce3fefcc53a3591d4485d74c0f80831c1ef8f7c1e373826e03514c92dfa; these gitignored paths do not promise evidence availability in clones or CI.

Current review head: 2b2b45aa4a995fa2d1047fd1142b9445c57f3a81. No expected_source_verified, publisher authentication, load-verification implementation or behavioral equivalence is claimed. The original model, lineage, attestation, thresholds and presets are unchanged.

Fresh capture at clean source df475bb
used --run-probe --build-dir build/runtime_identity --require-complete.
All five axes and the bounded MOT17-09 identity probe are complete.

Only implementation changes: 294 files remain 294 files, with seven
shipping blobs changed for Gate A identity, RunCompletion, v4 reports
and test-only CMake wiring. The implementation coordinate moves from
9a0fe01a4645616a01034c90ffc8c89389d42eb5072c469f24fee688021f62f5 to
0fcbdb09c745c06a677ffad47b02bdcb9d7cf7eeaee044f3089c82c0e9dd1094.
decision_surface, environment, identity_semantics (27 files), and
runtime_inputs remain unchanged. The fresh probe build witness binds
the newly built Release extension and scan plugin in build/runtime_identity.

The rerun probe digest equals the archived reading
2dabed0bc05e3bc75ec2115b3213f5c0b1aed3e837c22dd2325109339e4719b5.
It is a fixture change detector and does not prove semantic equivalence;
equivalence.state remains unproven. It also does not qualify the separate
legacy V5 NVIDIA-driver mismatch (617.42 vs required 616.92).

Archive the old complete canonical byte-for-byte (SHA256
171017c356d2e71cf4104a36f5334286de87670793a17c274c0c341271c572a2).
Keep the two-file coordinate change on a separate exact-head review
surface before incorporating it into the unmerged implementation PR.

Copy link
Copy Markdown
Owner Author

#549 S2-1a — separate owner merge authorization (2026-10-10, Asia/Taipei)

Decision: APPROVED FOR BOUNDED SOURCE-LEVEL MERGE of PR #572 into main by a merge commit, strictly pinned to reviewed head 2b2b45aa4a995fa2d1047fd1142b9445c57f3a81. This is separate from the S2-1a implementation authorization and legacy binding clarification acceptance. The PR is currently DRAFT, so transition it to Ready only after confirming the head and checks are unchanged; this comment authorizes the merge, but does not execute it. Reject a moved head without renewed review/authorization.

Reviewed evidence and exact boundaries

  • Verified current main and PR base 15e819502e4d5ea581792a3a16f97857435fc585; feat(shipping): observe legacy Gate A checksum identity (S2-1a) #572 head 2b2b45aa4a995fa2d1047fd1142b9445c57f3a81 is 5 commits ahead, mergeable. PR head CI 8/8 SUCCESS: CI + C++ build; this is not a post-merge main CI claim.
  • Scope accepted: legacy Gate-A-only N-T3 checksum_matched / nullable six-binding diagnostics with actual caller expected-source locations, immutable identity; versioned journal v2 / report v4 and reader checks including historical compatibility and corrupt-report rejection; original Gate A rejection policy and Gate B loader unchanged. Neither expected_source_verified nor load_verification is implemented.
  • Local checks reported in PR #572 and bounded evidence record: CPU protocol/reader checks, 23 real-binary CLI controls, 6 native GPU controls; local pre-push PASS. Independent Codex read-only review at the final head reports no findings. This is not a human/GitHub APPROVED review, and locally retained raw evidence is not portable CI evidence.
  • #573 coordinate-only review, head fe4b83f1fd08842b361e918f5ac9fb5e23f2851c, is an ancestor of feat(shipping): observe legacy Gate A checksum identity (S2-1a) #572; both canonical publication and append-only outgoing archive Git blobs match feat(shipping): observe legacy Gate A checksum identity (S2-1a) #572 exactly. Its review was independent and read-only; chore(provenance): review S2-1a runtime coordinate republication #573 has no CI checks and must not be merged separately. The candidate implementation coordinate changed; the other four axes are unchanged; equivalence.state=unproven.
  • Explicit qualification risk acceptance for SOURCE-LEVEL MERGE ONLY: seven original-model sequences and negative controls were executed as fresh local diagnostics, but frozen V5 required NVIDIA driver 616.92, whereas the host used 617.42. Therefore qualified GPU parity remains UNRESOLVED, formal EXACT/CAUGHT is not established, and caught=false must remain as recorded. Raw detector/MOT/graph equality and raw negative differences do not override frozen V5 qualification. Historical GPU parity does not requalify the new binary. This risk is knowingly accepted for merging the observability change at source level, not for claiming behavioral equivalence, release qualification, installed package correctness, or closing the GPU evidence obligation.

Merge and post-merge obligations

  1. Confirm PR feat(shipping): observe legacy Gate A checksum identity (S2-1a) #572 has been marked Ready, the exact head is still 2b2b45aa4a995fa2d1047fd1142b9445c57f3a81, base is still 15e819502e4d5ea581792a3a16f97857435fc585, required PR CI remains 8/8 success and mergeable CLEAN. Use a merge commit, preserving the reviewed coordinate history; refuse to merge a moved head. Do not merge chore(provenance): review S2-1a runtime coordinate republication #573 on its own.
  2. After the actual merge, verify merge parents and run the current main required CI; count only checks actually reported (do not infer a C++ job or GPU PASS). Only after successful main CI write a bounded [Architecture umbrella] Separate model artifacts from Saccade runtime and release lifecycle #549 S2-1a checkpoint and a architecture(B): design a supported architecture with traceable diagrams and behavior contracts #536 CC-536-01-02 source-level implementation status, with GPU V5=UNRESOLVED and installed package=UNVERIFIED; update [Hub] SACCADE 工作總入口:進度、優先序、依賴與 checkpoint #550's single cross-Issue hub. architecture(B): design a supported architecture with traceable diagrams and behavior contracts #536 and [Architecture umbrella] Separate model artifacts from Saccade runtime and release lifecycle #549 remain open.
  3. Preserve a named later follow-up for qualified V5 recheck on the stipulated 616.92 host/driver, or a separately owner-approved versioned change to that qualification protocol. Do not silently relax V5 or promote raw comparison results to qualified EXACT.
  4. Not authorized: re-pin/rebuild/distribute the installed package, public release, TR-1b allowlist, --model-bundle, S2-1/S2-2/S2-3 implementation, Gate B loader or TOCTOU redesign, new model bytes, performance/benchmark claims, CI-gate weakening, or release rights rulings. [465 follow-up] Public distribution readiness: licence decisions, release key, publish decision #547 stays distribution.status=local-only / owner_confirmation=null.

This approval is an explicit exception only to the qualified GPU evidence needed for a source-level S2-1a landing, not an exception to truthful evidence status, any existing mandatory CI checks, or later package/release gates.

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.

1 participant