← Documentation home

Canonical Markdown source · Oct 20, 2018

Calliope provider replay — technical recommendation

technical-recommendations/calliope-provider-replay-2026-08-25.md · 34 lines · SHA-256 b151d40db9f0

Impact summary

Calliope's reviewer workflow now depends on semantic replay of the v10 provider

receipt, not file presence. The parser binds exact fixtures, recomputes aggregate

Anthropic pricing, validates ordered latency and refusal/grounding gates, rejects

schema drift, and preserves false consequential authority. Model behavior and the

pending independent-review requirement are unchanged.

Ranked coupling and failure findings

  1. Critical — filename authority: any JSON at the registered v10 path was

previously counted as provider evidence and could expose the aggregation route.

  1. High — economic drift: aggregate token and cost fields were not recomputed

when readiness consumed the receipt.

  1. High — fixture/safety drift: readiness did not independently bind the two

fixture hashes or replay the five pre-provider malformed-rights refusals.

Actions

| Owner | Action | Acceptance criteria | Validation |

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

| AI reliability | Preserve strict v10 replay | Extra/missing keys and fixture, count, price, latency, grounding, safety, blocker, or authority drift fail closed | `node --import tsx --test --test-concurrency=1 tests/services/calliope-curatorial-trial.test.ts` |

| Evaluation operations | Collect the two prepared independent responses | Exact v10 packet receives two distinct timed private responses; public aggregate contains no identities or candidates | `pnpm audit:ai-agent-value:readiness` |

| Agent governance | Extend semantic replay to retired Clio and reconciliation evidence | Provider and retirement totals require parsed source receipts rather than paths | `node --import tsx --test --test-concurrency=1 tests/services/ai-agent-trial-readiness.test.ts` |

Next-cycle hypothesis

Mutating any authoritative Calliope provider field prevents readiness from routing

to independent-review aggregation, while the exact historical v10 receipt and

assignment receipt still replay. The focused parser and readiness tests falsify

this hypothesis on any accepted drift or rejected canonical receipt.