# Dispatch: frozen verification record

The original study reported 4/4 complete answers in each condition. This download lets you inspect its synthetic inputs and replay the eight retained final answers against verifier version 1. It does not rerun an agent or establish an effect of the source index.

Use Node.js 22 or later. Review the source before running `node replay.mjs > replay.json`. All four original controls must match, and all eight answers must reproduce their recorded verdicts. Files are read locally; the replay makes no network calls and leaves original records untouched. No evx account or model provider is needed.

The replay also demonstrates a known version 1 limitation: an equivalent leading-zero money cell is rejected. The current runnable example uses version 2 and more controls. Those later changes do not rewrite this result.

## Provenance

The synthetic baseline, candidate AGENTS.md, expected.json, verify.mjs and validation.mjs retain exact bytes from pre-execution commit 5255a0cc9b8989fb07e34704d827baf955baca8b. Other candidate input files were byte-identical to their baseline counterparts and are omitted here. The manifest, plan, result and summary are unchanged historical records copied from packages/framework/evidence at 11cfe4dbe485b02a4213b958693bc841d58312ca.

final-answers.json was extracted on 10 September 2026 with the framework's bounded final-answer reader. Every original trace byte length and SHA-256, and every extracted answer hash, matched the retained result. This is a later publication projection, not a pre-execution artifact. Raw provider streams are not published. replay.mjs and this README are later reproduction aids.

The study ran on a local Mac without external observation. Full model context and access outside the trial directory were not attested. Hashes identify bytes; they do not attest execution. Reported dollar costs are unknown. Four pairs at a success ceiling are inconclusive; the declared conservative interval spans -100 to +100 percentage points.