← Documentation home

Canonical Markdown source · Oct 20, 2018

AI agent comparison-chain parity — technical recommendation

technical-recommendations/ai-agent-comparison-chain-parity-2026-08-25.md · 34 lines · SHA-256 7fcfdbbf7aad

Impact summary

Promotion comparison can no longer bypass readiness evidence verification. Both

paths share one exact registered assignment/readiness/review-chain verifier. A

wrong-packet receipt is rejected before comparison output, even when its schema,

surface, counts, arithmetic, gates, and authority are otherwise valid.

Ranked findings

  1. High — genuine subjective evidence is still absent: closing a bypass does

not move either prepared surface above `unproven`.

  1. Medium — pure evaluator calls are not evidence acceptance: callers must use

the governed CLI/service boundary before persisting promotion artifacts.

  1. Medium — registered handoffs currently exist for only two surfaces: other

model surfaces cannot enter subjective comparison until their trials justify

assignment preparation.

Actions

| Owner | Action and acceptance criteria | Validation |

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

| AI Evaluation | Use the governed comparison CLI only after chain-valid aggregation | Wrong-packet CLI test fails before output |

| Editorial Research | Complete the unchanged Calliope or Clio packet | Two independent responses bind the registered hash |

| Trust | Accept or reject resulting evidence outside comparison generation | Comparison authority remains false |

| Maintainers | Route future promotion consumers through the shared verifier | No direct receipt-parser use at persistence boundaries |

Next-cycle hypothesis

A source scan and consumer contract test can prevent future persisted promotion

paths from reintroducing direct, unbound receipt parsing. The substantive value

hypothesis remains dependent on genuine reviewer returns.