Skip to main content

crp.ep

Auto-generated reference for the crp.ep subpackage.

ep

crp.ep

Epistemic Profiles & Calibration (CRP-SPEC-055).

CalibrationProfile dataclass

Reliability curve for one (model, task-kind) pair.

observe(confidence, correct)

Record one verified outcome.

expected_calibration_error()

Return ECE - average gap between confidence and empirical accuracy.

overconfident_on(threshold=0.15)

True when the model is systematically overconfident.

epistemic_adjust(base_tier, risk, entropy, profile=None)

Adjust tier/risk/positioning using epistemic signals.

Parameters:

Name	Type	Description	Default
base_tier	str	Starting quality tier (S/A/B/C/D).	required
risk	str	Starting risk level.	required
entropy	float	Normalised semantic entropy in [0, 1].	required
profile	CalibrationProfile | None	Optional calibration profile for the model/task.	None

Returns:

Type	Description
dict[str, Any]	Dict with adjusted tier, risk, and optional positioning_hint.

ep.apply

crp.ep.apply

Wire epistemic signals into tier/risk/positioning (CRP-SPEC-055 §7.3.3).

epistemic_adjust(base_tier, risk, entropy, profile=None)

Adjust tier/risk/positioning using epistemic signals.

Parameters:

Name	Type	Description	Default
base_tier	str	Starting quality tier (S/A/B/C/D).	required
risk	str	Starting risk level.	required
entropy	float	Normalised semantic entropy in [0, 1].	required
profile	CalibrationProfile | None	Optional calibration profile for the model/task.	None

Returns:

Type	Description
dict[str, Any]	Dict with adjusted tier, risk, and optional positioning_hint.

ep.calibration

crp.ep.calibration

Per-model, per-task calibration curves (CRP-SPEC-055 §7.3.2).

CalibrationProfile dataclass

Reliability curve for one (model, task-kind) pair.

observe(confidence, correct)

Record one verified outcome.

expected_calibration_error()

Return ECE - average gap between confidence and empirical accuracy.

overconfident_on(threshold=0.15)

True when the model is systematically overconfident.

ep.semantic_entropy

crp.ep.semantic_entropy

Semantic entropy - uncertainty from meaning-level divergence (CRP-SPEC-055 §7.3.1).

semantic_entropy(samples, budget_ms=100.0)

Compute normalised semantic entropy over samples.

Uses a local NLI model when available, otherwise falls back to string equality. The call is budgeted; if it exceeds budget_ms the fallback path is returned.