---

name: confidence-prototypes-utilities description: "Routes Alibi requests for TrustScore, LinearityMeasure, ProtoSelect, save/load round-trips, and the small shared utilities they rely on." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Confidence, Prototypes, and Utilities

Use this sub-skill when a user wants to score prediction confidence, summarize a dataset with prototypes, or save and reload an explainer.

Trigger phrases

Open this sub-skill when the user mentions:

• TrustScore or trust scores
• LinearityMeasure or linearity measure
• ProtoSelect or prototype selection
• save/load explainer, load_explainer, or explanation persistence
• category map, gen_category_map, or save_explainer

What this sub-skill owns

• TrustScore
• LinearityMeasure
• ProtoSelect
• save_explainer and load_explainer
• the shared utility helpers that support those workflows

What it does not own

• anchors
• SHAP / IntegratedGradients
• counterfactuals or similarity
• any heavier optional-backend family

Read next

• references/workflows.md for the confidence / prototype / persistence flow map
• references/api-reference.md for the verified TrustScore, LinearityMeasure, ProtoSelect, and persistence signatures
• references/saving.md for save/load behavior and version sensitivity
• references/troubleshooting.md for label, range, kernel, and persistence errors
• scripts/smoke_confidence_prototypes.py for the bundled CPU smoke

Typical flow

1. Decide whether the user wants a confidence score, a prototype summary, or a persistence check.
2. Make sure the classifier labels and input shape match the chosen helper.
3. Run the smoke script on iris-sized data to confirm the environment and the API contract.
4. If the user only needs a round-trip or a utility check, keep the example small and deterministic.

Good fits

• warning flags for classifier predictions
• model linearity around a point
• dataset summarization and prototype selection
• save/load round-trips for explainer objects

Common failure signals

• TrustScore label count does not match the class structure
• LinearityMeasure cannot infer the feature range or the model type is wrong
• ProtoSelect distance or preprocess settings are incompatible with the data
• a saved explainer cannot be reloaded without passing the original predictor