# Evaluation

## Required executable cases

| Case | Expected result |
|---|---|
| Sycophantic Ally | High-confidence accommodation/directness drift; bounded modulation and feedback, not a Presence rewrite |
| Over-persistent Actor | Preflight denies further retry or escalates; no constraint circumvention |
| Deliberately forceful Ally | High non-zero directness target is respected; no pull toward zero |
| Formal Realm | Unusual high-formality target remains aligned when observed near target |
| Human Avatar | Optional drafting/feedback only; no automatic human correction |
| Low confidence | `observe` or `ask`; no modulation/correction |
| Conflicting inherited targets | Deterministic override or `resolution-required`; never average silently |
| Extreme boundary | Constraint/escalation even when configured target is high |
| Mediation | Interaction mismatch and temporary consent-based compact; neither participant labeled incoherent |

Fixtures live in `examples/evaluations/`; executable assertions live in `reference-implementation/test/coherence-engine.test.mjs`.

## Acceptance gates

- All registered examples validate against Draft 2020-12 schemas.
- The non-zero-target fixture reports target distance, band drift, and boundary status separately.
- Low-confidence fixture cannot select `modulate`, `correct`, `repair`, or probabilistic `constrain`.
- Avatar fixture never returns control-directed intervention.
- Actor deterministic boundary fixture constrains synchronously.
- Effective profile traces every field to an input version.
- No schema or API contains a global human `overallScore`.
- Every Skill produces either schema-valid output or an explicit blocked/insufficient-evidence result.
- Site and Markdown use the same response and precedence vocabulary.

## Scale evaluation

Each dimension requires content validity review, inter-rater/measurement agreement where applicable, context sensitivity, calibration, disparate-error analysis where lawful, adversarial examples, counterfactual cases, and consequence-specific confidence gates. A scale can be suspended without deleting its history.