← Documentation home

Canonical Markdown source · Oct 20, 2018

AI agent readiness blocker codes — technical recommendation

technical-recommendations/ai-agent-readiness-blocker-codes-2026-08-25.md · 43 lines · SHA-256 4599699f7f26

Impact summary

Unified AI trial readiness advances to version 3 with a per-surface

`blockerCodes` array. Canonical rows contain no codes. If assignment evidence fails

semantic replay, only `ASSIGNMENT_RECEIPT_INVALID` and generic operator prose are

published; provider evidence remains valid and assignment readiness remains false.

The retained-data boundary explicitly excludes parser messages and private paths.

No agent grade, deployment/publication authority, or canonical 4/2/2/0 count

changes.

Ranked coupling and failure findings

  1. High — response completion is still filename-based: readiness currently

treats any independent-review receipt path as completion without semantic replay.

  1. High — human evidence remains absent: better diagnostics cannot populate the

subjective quality or reviewer-effort gates.

  1. Medium — blocker-code evolution requires schema discipline: new codes must be

union-typed, privacy-reviewed, mutation-tested, and accompanied by a version

decision rather than arbitrary strings.

Actions

| Owner | Action and acceptance criteria | Validation |

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

| AI Reliability | Semantically parse independent-review receipts before completion credit; malformed receipts yield a bounded code and preserve earlier evidence lanes | Mutation tests cover packet, reviewer count, scores, authority, and privacy fields |

| AI Evaluation | Collect two genuine response exports per prepared packet and aggregate through the existing private packet/key workflow | Exact aggregation meets all quality, protected-dimension, agreement, and effort gates |

| Trust | Maintain an allowlisted blocker-code union and reject raw exception text or paths in public artifacts | Readiness privacy tests and staged secret scan |

| Maintainers | Document every readiness schema change and bump the version when adding required fields | Artifact fixture/test asserts version 3 and exact required fields |

Next-cycle hypothesis

Semantic review-receipt verification can replace filename-based completion without

coupling public readiness to private responses. Falsify by placing a forged receipt

at the registered path; completion must remain false, prior evidence lanes true,

and only a bounded privacy-safe code may appear.

Outcome

Confirmed. Filename-only completion is removed; forged receipts retain prior

evidence, keep completion false, and publish only `REVIEW_RECEIPT_INVALID`.