Skip to content

MARC_2_VR6_VR2_BOUNDARY_LOCALIZATION_IMPLEMENTATION.md

Latest commit

 

History

History
156 lines (122 loc) · 6.4 KB

MARC_2_VR6_VR2_BOUNDARY_LOCALIZATION_IMPLEMENTATION.md

File metadata and controls

156 lines (122 loc) · 6.4 KB
Raw

MARC2-VR8A VR6-to-VR2 Boundary Localization Result

Date: 2026-08-16

Lane: MARC2-VR8A

Status: Artifact-only implementation and measured audit complete; remotely green

Registration d33eaf397a8f8444db4d7abd777bf2e9b3333e43 passed Base Python job 95256950555, Optional Neuro Readers job 95256950656, and CI 31984475999 before implementation began.

Short Answer

The audit reached MARC2VR8A-R1. The unresolved validation failure is no longer an eight-way or three-way mystery:

excluded:  MARC2VR2-F02  source envelope, identity, transport, row count
remaining: MARC2VR2-F03  BIDS path, run companions, structural grouping
remaining: MARC2VR2-F04  bundle count, participant/session taxonomy arithmetic

This does not identify whether the consumed source reached F03 or F04. It does prove that another envelope repair would be misdirected.

Where The Route Was Lost

VR6 already did the right privacy-preserving thing. When VR2 refused, VR6 stored two separate allowlisted fields:

route:          MARC2VR6-F02
upstream_route: one MARC2VR2-F01 through F08 code

VR7P caught that exception but copied only exc.route. It never copied exc.upstream_route. The result therefore retained the broad VR6 branch and discarded the exact safe VR2 class. No reason, path, participant, run, or value was ever needed.

Why F02 Is Excluded

The retained source hash is byte-identical to the private-manifest hash emitted by the exact remotely qualified MARC1 producer. Committed code and public aggregate evidence then agree with VR2 on every envelope predicate:

  • schema name and version;
  • live structural proof posture;
  • six top-level fields;
  • Figshare record, version, file, archive size, and MD5 identity;
  • exact directory, metadata, and tail transport keys and SHA-256 values;
  • exactly 1,227 rows, 1,025 regular files, and 202 directories; and
  • the exact eleven-field parser row schema.

The sole VR7P pass also verified the same 418,755-byte hash and strict-parsed the JSON before calling VR6. Those facts satisfy the complete F02 predicate without reopening the source.

What F03 And F04 Mean

F03 is now limited to source-dependent path and grouping questions. Examples include whether every suffix-bearing path matches the frozen Freewill grammar, whether the task token has the expected spelling and case, whether two paths collapse to one logical run companion, and whether every matched run has all four companions.

F04 is the aggregate semantic layer after grouping: exactly 238 complete logical bundles, known participant classes, 195 + 43, and exact eligible participant-session counts.

No private member name, participant, session, run, or failed value was read or inferred by this audit.

Why Generated Tests Missed This Boundary

VR2's 1,227-row success source starts from selector.build_generated_manifest. It then fabricates the 43 noneligible run bundles and rewrites live envelope values. It never calls the exact central-directory parser or producer.

The producer's own generated parser fixture has only 18 rows and validates ZIP mechanics, not the full 1,227-row Freewill morphology. The repository therefore had strong tests on each side but no full-scale producer-to-VR2 integration fixture. That gap is now explicit.

Measured Audit

route:                         MARC2VR8A-R1
committed inputs:                         18
input bytes:                         587,523
Python AST parses:                         7
strict JSON parses:                       11
runtime:                   0.083712375 sec
peak RSS:                     51,052,544 bytes
aggregate output bytes:               8,827
retained generated bytes:                 0
CPU threads / workers / jobs:        1 / 1 / 1
raw-data reads:                           0
real-cache reads:                         0
model runs / training runs:           0 / 0
producer causal:          not applicable, artifact only
end-to-end latency measured:             no

All eleven acceptance gates passed. Every private, ignored-path, consumed- root, archive, signal, target, derivative, model, prediction, score, network, provider, hardware, FW2, CIL1, retry, other-project, and claim counter stayed zero. The aggregate report was inspected from stdout and not retained as generated debris.

Verification

  • Twenty-seven focused contract, behavior, result, and remote-proof tests pass.
  • Deterministic replay is byte-identical under fixed runtime/RSS readers.
  • Hash mutation, symlink, missing-thread, and RSS-cap cases fail closed.
  • Public-output privacy and the two-command CLI surface are tested.
  • The dependency-light suite passes 3,918 tests with 204 expected optional skips in 286.785 seconds.
  • The neuro-enabled inventory contains 3,989 tests: 3,953 passed, 35 skipped, and the one established sandbox forkserver-socket case errored. That exact case passes outside the sandbox in 2.823 seconds.
  • The pre-change inventory was 3,953 tests. Exact implementation 1addd5d added 36 tests with zero new regression; this proof closeout adds one remote-proof invariant.
  • Repository-pinned Ruff 0.15.20, full compilation, strict registry JSON, project and module CLI help, final audit, and git diff --check pass.
  • Exact implementation/result commit 1addd5df9fdccda6e716f71f9e6624f199677713 passed Base Python job 95261271737, Optional Neuro Readers job 95261271709, and CI 31986089529.

Next Repair

The next lane should be small and generated-only. It should carry both the outer VR6 route and nested allowlisted VR2 route through a new aggregate refusal, while discarding exception text and all private context. Generated F02, F03, and F04 cases must exercise the complete relay.

Do not relax F03 or F04 yet. After the diagnostic relay is committed, pushed, and remotely green, any new 418,755-byte structural read still requires a new all-false Tier C packet and fresh packet-bound maintainer decision. FW2 and CIL1 remain ineligible until a real cohort is frozen.

Boundary

Engineering capability added: NeuroDecodeKit can now exclude the full source- envelope class and narrow the unresolved structural validation boundary to F03 path/companion structure versus F04 bundle/taxonomy arithmetic without private reinspection.

Scientific claim not established: no neural payload, target, prediction, or score was accessed, so this audit establishes no neural effect, decoding accuracy, language decoding, live decoding, or thought-to-text capability.