Coverage

Coverage asks what an evaluation actually reaches: languages, domains, classes, source documents, risk categories, thresholds, fallbacks, and user segments. It is different from quality. A slice can be covered and fail badly; an uncovered slice makes the aggregate score silent. This is why golden datasets need metadata, not just inputs and labels.

Measuring slice coverage

For categorical slices, coverage is the fraction of required cells with at least one valid example:

$$\text{coverage}=\frac{|\{c\in C:\text{count}(c)>0\}|}{|C|}.$$

For numeric regions, define bins before evaluation: score thresholds, latency bands, document age, or confidence ranges. Repeated sampling can quantify metric variability, but it cannot create evidence for cells that have zero examples.

For evidence-grounded systems, source coverage uses the same idea over required source handles:

$$source\; coverage=\frac{|S_{\mathrm{tested}}\cap S_{\mathrm{required}}|}{|S_{\mathrm{required}}|}.$$

Worked calculation

Suppose the required launch claim spans four languages and two domains. The evaluation set contains English billing, English legal, German billing, German legal, and Spanish billing examples:

language	billing	legal
English	covered	covered
German	covered	covered
Spanish	covered	missing
French	missing	missing

The set covers 5 of 8 required language-domain cells, so slice coverage is $5/8=0.625$. Aggregate accuracy from this set says nothing about French, and it only tests Spanish billing. That gap should block claims about broad launch readiness even if offline evaluation looks strong.

Caveats

Counting rows can exaggerate coverage when examples are near-duplicates. Source coverage is not answer coverage in RAG: a question may retrieve a policy document yet never exercise the refusal path. Coverage plans should include abstention and severe-error categories, not only happy-path inputs.

References

• NIST AI Risk Management Framework
• scikit-learn documentation: Metrics and scoring

Nav

Section — Experimentation and Evaluation

← Comparing Generative AI and Classical ML Systems Abstention →