There is no “perfect engine” claim in Hikmah Stack. There is a perfectibility loop: every observed failure should become a reproducible case, an invariant, a test, a policy, or an explicitly accepted limitation.
- Recall precision@k for known relevant traces.
- Contradiction recall rate for structured conflicting claims.
- Provenance retention rate after consolidation/replay.
- Commitment recall before/at deadline.
- Stale-belief activation rate after supersession.
- Memory-poisoning rate from unverified model-generated claims.
- Sensitive-persistence violations.
- Missing-evidence visibility.
- Hard-block compliance.
- Ranking stability under irrelevant context.
- Reversibility preference when evidence is weak and expected value is otherwise close.
- Human-impact question surfacing for consequential actions.
- False completion rate.
- Claimed-vs-observed test pass mismatch.
- Rollback information completeness.
- Outcome write-back coverage.
- p50/p95 recall latency by trace count.
- Memory footprint.
- Ledger replay time.
- Hash-chain verification time.
- Determinism across identical inputs.
Any future local model must be tested behind the same ProposalEngine boundary. Compare candidate architectures on task quality and downstream repair, verification, memory pollution, latency, energy, and operator trust. A model that writes beautiful prose but causes more incorrect durable memories loses.