Docs/07 llmops/evaluation/metrics/README

Evaluation Metrics Catalog

Version: 1.1.0 Last updated: 2026-07-16 Status: Informative OAIES implementation profile

Purpose

Define claim-level faithfulness, groundedness, context recall, hallucination rate, and decision rules.

Why

Metric names are often inconsistent; operational definitions prevent score laundering.

When

Use when specifying evaluation plans and dashboards.

How

  1. Faithfulness: supported claims divided by verifiable claims in supplied context.
  2. Groundedness: weighted support score for each externally checkable claim against approved sources.
  3. Context recall: required reference facts retrieved divided by all required facts.
  4. Hallucination rate: unsupported or contradicted claims divided by verifiable claims; also report severe-case rate.
  5. Report numerator, denominator, abstentions, confidence interval, slice results, judge version, and threshold rationale.

Evidence contract

The decision record is the metric specification registry. It records metric ID; unit; numerator; denominator; abstention rule; severity weighting; confidence method; owner. The measurement owner owns completeness; the evidence is invalid when an implementation changes without a metric-version change. Dataset, annotation, rubric, scorer, and run identifiers are content-addressed so a disputed result can be replayed.

Failure response and recovery

Trigger: denominator integrity fails or a severe error is averaged away.

Immediate response: invalidate the report and recompute from claim-level annotations. Preserve the metric specification registry, affected trace IDs, timestamps, and decision logs before mutation. Open an incident when users, data, money, authorization, or a release decision may have been affected; closure requires a regression case and verified control change specific to evaluation metrics catalog.

Decision authority

The measurement owner accepts the operational decision. The domain risk reviewer provides independent challenge for high-risk scope, failed gates, or exceptions. Evaluation automation enforces the approved statistical rule; the evaluation and domain owners decide ambiguity, severe-case disposition, and residual uncertainty.

Tradeoffs

Choice Benefit Cost
Claim-level metrics Actionable diagnosis Annotation cost
Aggregate score Simple trend Masks severe tails

Anti-patterns

  • A universal 0.85 threshold.
  • Counting unverifiable style statements as hallucinations.
  • Reporting means without denominators.

Enterprise considerations

  • Set stricter gates for safety-critical claims.
  • Provide human appeal for consequential outcomes.

Framework relationship

For Evaluation Metrics Catalog, this informative profile governs measurement evidence for the stated decision only; it neither makes an evaluator authoritative nor transfers fitness decisions to a framework.

Source Relationship for Evaluation Metrics Catalog Boundary
NIST AI RMF MEASURE 2.5 Interpret outcomes against the documented use case, sampling frame, and uncertainty.
ISO/IEC 42001 42001 clause 9.1 Use management-system evidence only within the organization’s declared scope and independent assessment process.
Domain threat/control source Misinformation and overreliance metrics Test only the threats applicable to the documented system and release

Checklist

  • Definitions are fixed.
  • Severe failures are separately gated.
  • Confidence intervals are reported.

References

Changelog

Version Date Change
1.1.0 2026-07-16 Replaced generic assurance text with the metric specification registry, failure trigger, accountable decision, and scoped framework relationships for evaluation metrics catalog.
1.0.0 2026-07-16 Initial complete profile.