← Documentation home

Canonical Markdown source · Oct 20, 2018

AI agent review-receipt verification — technical recommendation

technical-recommendations/ai-agent-review-receipt-verification-2026-08-25.md · 40 lines · SHA-256 ef6ed929a06a

Impact summary

Independent-review completion is now based on strict semantic replay, not file

presence. Invalid receipts preserve provider and assignment evidence and emit

only `REVIEW_RECEIPT_INVALID`. Canonical state remains 4/2/0/2/0: provider

evidence, ready assignments, completed reviews, retirements, proven advantages.

Ranked findings

  1. High — genuine review is still the binding constraint: two independent

reviewers must return completed forms before an honest aggregate can exist.

  1. High — aggregate replay is not identity proof: qualification and

independence remain separately governed private evidence.

  1. Medium — public evidence must stay aggregate-only: exact-key parsing must

continue to reject fields that could leak responses or identifiers.

Actions

| Owner | Action and acceptance criteria | Validation |

|---|---|---|

| AI Evaluation | Collect two genuine returns for Calliope v10 and Clio hook v6 | Two distinct private responses per packet |

| AI Reliability | Publish only a parser-valid aggregate bound to its surface | Strict parser and verifier replay |

| Trust | Verify independence separately and explicitly accept evidence | Receipt grants no acceptance authority |

| Maintainers | Privacy-test every new diagnostic code | No raw error, path, hash, candidate, or reviewer data |

Next-cycle hypothesis

One prepared surface can obtain two genuine blinded returns whose aggregate

populates quality lift and reviewer effort without protected-dimension regression.

Until then, no agent advantage over its deterministic baseline is proven.

Binding outcome

The verifier now proves the receipt belongs to the exact registered handoff, not

merely to a named surface. A semantically valid receipt with a different packet

hash is withheld, prior provider/assignment evidence remains available, and the

public diagnostic stays `REVIEW_RECEIPT_INVALID`.