---
artifact_kind: evaluator-card
artifact_name: <e.g., "pareto-frontier-suite-2p">
version: <semver>
owner: <team>
last_reviewed: <YYYY-MM-DD>
next_review: <YYYY-MM-DD>
evaluator_kind:
  <fairness | subjective-value | stability | adversarial | clause-exploit |
  benchmark | outcome>
related_phase_179_tasks: [<...>]
---

# Evaluator Card — <artifact_name> v<version>

## 1. Identity

- **Evaluator name:** <name>
- **What it measures:** <one sentence>
- **Harm families covered (H1–H8):** <from
  [`../nist-ai-rmf-mapping.md`](../nist-ai-rmf-mapping.md)>
- **Consumers:** <which gates / runbooks consume this evaluator's output>

## 2. Method

- **Input corpus / fixture:** <synthetic / benchmark / production trace; cite
  source and consent posture>
- **Procedure:** <step-by-step description of how the evaluator runs>
- **Metrics emitted:** <exact metric definitions and units>
- **Aggregation:** <how per-case metrics roll up to a pass/fail verdict>

## 3. Baselines

- **Comparison baselines:** <LLM-only mediator, simple Nash GA, static
  templates, human-authored seed-offer-only>
- **Baseline scores:** <current values for each baseline>

## 4. Thresholds

- **Release-gate threshold:** <exact numeric threshold for pass/fail>
- **Regression threshold:** <how much degradation triggers a block>
- **Domain-specific thresholds:** <per use-case class>

## 5. Known limitations

- <where this evaluator is weak: sample size, bias, scope>

## 6. Validation

- **Last validation date:** <YYYY-MM-DD>
- **Validation procedure:** <how the evaluator itself was validated; cite any
  peer-reviewed method>

## 7. References

- <link to test fixtures, notebooks, ADRs, academic papers the evaluator implements>