# Transaction Evidence: seven-case experiment

An offline experiment for agents that consume prepared-purchase evidence. The cases distinguish byte integrity, agreement with saved terms, and usefulness to the buyer.

[Manifest](manifest.json) · [File checksums](SHA256SUMS.txt) · [Verifier guide](/transaction-evidence.md#reference-hash-assessment)

| Case | Change | Question |
| --- | --- | --- |
| 01 | Original bundle, bytes and reference hash | Do the recorded relationships check out? |
| 02 | One result byte changes | Do both byte checks reject the change? |
| 03 | Result bytes are omitted | Does the report identify the absent input? |
| 04 | Terms change; the original reference stays fixed | Does the terms check reject the substitution? |
| 05 | Changed terms; no reference hash | Does the consumer distinguish internal consistency from agreement? |
| 06 | Changed terms; reference copied from that bundle | Does the consumer recognize a circular comparison? |
| 07 | Original evidence; separate freshness requirement | Can the bytes verify while failing the buyer's task criterion? |

The manifest freezes expected reports against published verifier **0.1.0**. Observations are reserved for a reviewed run. All identities, receipts and payment references are synthetic; execution needs no wallet, funds, HTTP or RPC.

Record the manifest and checksum-file digests before running. Use the exact arguments in each case and retain stdout, stderr, exit code and parsed report. An exit code alone does not demonstrate the expected rejection. Evaluate case 07's freshness rule separately from evidence verification.

Published release **0.2.0** adds an optional reference assessment alongside the existing report. Its `comparison` describes hash equality; `sourceClaim` records the caller's account of where the reference came from. `sourceClaimVerified` is always false. The frozen manifest remains a 0.1.0 baseline; its proposed consumer labels are not fields in that package's report.
