Independent reading of trial 06

The retained trial 06 results pass the bounded archive, source-conservation and relationship checks described here. This is acceptance of measured evidence delivery for the 40 staff question occurrences, not legal answerability or acceptance of the separate original imprisonment demonstration.

What was checked

An independently written audit read all 328 retained packages from trials 05 and 06. It made no assembly, source-acquisition, network or model calls. It did not import the evaluator's path or location-scoring helper. The pure, hash-bound concept resolver was used to check which declared public aliases activated each requirement.

The receipt records:

• 164 complete, unique question × stage × budget cells in each trial, including the separate unknown-question control;
• all 82 logical-baseline packages byte-identical between the two trials;
• all 616 earlier and 623 final report input bindings verified;
• the exact five engine modules checked against their admitted Git commit;
• every archive's canonical bytes, context identifier, byte budget and hash;
• all 5,454 returned assertion occurrences equal to their full declared source assertion, including conjunctive guards, and 7,067 returned selection paths with valid directed hops and selected endpoints;
• 625 distinct complete source records, 2,436 distinct source spans and 189 frozen extraction files checked; source spans reconstruct the whole selected unit using its declared joiner;
• all 203 original staff obligations retained and never supplied as evidence;
• all eight unknown packages across both trials contain no source evidence, relationship or active requirement.

The supplement confirms that repeated source identifiers retain exactly the same full record across 4,209 selected source record occurrences. It also checks the generated comparison against the independent totals and binds the final protocol, runner, helper and manifests to source-freeze commit 350c60d8ba7bc83297aa85cdcfd7ddd2d5d789c7.

Results and trade-offs

Path and route counts are case incidences, not distinct rules or measures of legal completeness. All larger staff packages report both retrieval and assembly truncation. All 40 structured staff packages at 32 KiB contain no source evidence. The audit preserves this failure; a smaller response must not be described as containing the omitted source.

The three questions without a source-read selection remain Staff 001, Staff 008 and Staff 018. The frozen question wording and engine are unchanged; the improvement is an additive source-selection increment, not an answer-quality or runtime-speed experiment.

Separate acceptance boundary

The original imprisonment and hospital questions are a different denominator. Their first separate acceptance run found incomplete delivery of the authored custody selection. This staff audit does not override that result or authorise claiming complete custody routing. Its failure and any later repair must stay visible in the separate original-acceptance record.

This review checks source extraction bytes and retained PDF hash identifiers. It does not re-render those PDFs, verify every boundary's meaning, certify current law, close legal applicability questions or establish specialist acceptance, model-answer accuracy, affordability or public deployment.

Reproduce without overwriting this receipt

Use an isolated scratch checkout containing the frozen files. Keep this checkout and its receipts unchanged. In the scratch checkout only, move independent-review.json to a unique temporary archive before running:

node --experimental-strip-types \
  evaluation/manual-structure/context-probe/analysis-06/independent-audit.mjs \
  --explorer-root /path/to/okf-explorer-at-the-admitted-engine

The script refuses to overwrite an earlier receipt. It reads historical trial 05 files from the retained Git commit when their current projection differs. It does not execute the context assembler.