Impact summary
The retired reconciliation behavior is now justified by replayable source
evidence rather than registered filenames. Readiness parses both provider receipts
against the exact held-out cases, recomputes the retirement aggregate and strict
deterministic-dominance predicates, and validates false authority. Runtime remains
deterministic and future spend still requires the separately adjudicated v3 design.
Ranked coupling and failure findings
- Critical — retirement filename authority: an arbitrary file could mark a
behavior retired and alter the next-action route.
- High — unreplayed source aggregates: cost, latency, outcomes, and selection
deltas were copied into retirement evidence without consumer-side reproduction.
- High — provider receipt authority: the historical provider path was counted
without checking its manifest binding, telemetry, safety, tools, or authority.
Actions
| Owner | Action | Acceptance criteria | Validation |
|---|---|---|---|
| AI reliability | Preserve strict source and retirement replay | Any schema, manifest, count, token, cost, latency, safety, tool, aggregate, dominance, disposition, or authority drift fails closed | `node --import tsx --test --test-concurrency=1 tests/services/reconciliation-tiebreaker-trial.test.ts` |
| Reconciliation governance | Keep the dominated behavior absent from runtime | Public server path imports no model adapter; legacy configuration cannot reactivate it | `node --import tsx --test --test-concurrency=1 tests/pages/curator-reconciliation-page.test.ts` |
| Evaluation operations | Register only a materially different future hypothesis | Two-reviewer adjudicated v3 representative remains deterministically unresolved and adversarial case requires equivalence abstention | `node --import tsx --test --test-concurrency=1 tests/scripts/reconciliation-tiebreaker-trial-script.test.ts` |
Next-cycle hypothesis
Only the exact two historical receipts reproduce a retirement decision, and any
source or aggregate mutation prevents `revise-model-hypothesis` routing. The
focused trial/readiness tests falsify this on any accepted drift or rejected
canonical evidence.