CE Parameter Reference

Canonical definitions for the five user-facing parameters where naming confusion is most likely. Read this page before filing a bug about unexpected behaviour involving thresholds, confidence values, or significance levels.

Cross-references: terminology.md, probabilistic_regression.md, ADR-038: Call-time Configuration Taxonomy.

Disambiguation table — floats in (0, 1)

Three parameters are all floats in (0, 1) and are mathematically related, but they are used in completely different contexts. Confusing them is the most common source of subtle bugs.

Legacy / removed name

Canonical name

Role

Context

User-settable?

confidence

reject_confidence

Reject coverage target: minimum p-value to accept a prediction

predict_reject / reject orchestrator

YES

significance

GuardedOptions.confidence

Guard conformity threshold; significance = 1 − confidence

explain_factual(..., guarded_options=GuardedOptions(confidence=...)) / _guarded_explain

YES (experimental)

confidence_level

confidence_level (unchanged)

Regression coverage level; derived from low_high_percentiles but also passable directly as an alternative to threshold

Regression methods

YES (alternative to threshold)

Key relationship: significance and confidence (reject) are mathematical inverses of the same conformal threshold concept. significance = 0.1 is the same conformity requirement as confidence = 0.9. The rename is complete: use GuardedOptions.confidence for guarded explanations and reject_confidence for reject-policy prediction paths. The legacy names were removed in v0.11.5 and now raise ConfigurationError in the v0.11.6 fail-fast paths.

threshold

Definition: The regression output value used to define a binary event for probabilistic regression. The explainer produces calibrated probabilities for P(y ≤ threshold).

Type and valid range: float, int, or array-like (one value per instance for instance-specific thresholds). None means “not applicable” — used only in regression mode.

Applies to: explain_factual, explore_alternatives, predict_proba (in regression mode). Ignored for classification tasks.

Behavior: When set, the CE regression path calls IntervalRegressor.predict_probability(x, y_threshold=threshold). The parameter is renamed to y_threshold at the internal IntervalRegressor boundary to match the crepes API convention; this split is an implementation detail and does not affect the public interface.

Disambiguation:

  • NOT confidence_level: confidence_level specifies the width of the calibrated interval (e.g., a 90 % interval), not a target output value. They are mutually exclusive: passing both raises ConfigurationError (enforced by EXCLUSIVE_PARAM_GROUPS in api/params.py).

  • NOT confidence: confidence is the reject coverage target and belongs to the reject path, not the regression path.

low_high_percentiles

Definition: A (low, high) percentile tuple that sets the width of the calibrated uncertainty interval for regression. For example, (5, 95) produces a 90 % calibrated interval.

Type and valid range: Tuple of two floats, each in [0, 100], with low < high. Default: (5, 95).

Applies to: All regression explanation methods (explain_factual, explore_alternatives). Ignored for classification.

Behavior: Determines the lower and upper bounds of the calibrated interval. The derived metric confidence_level = (high - low) / 100 represents the nominal coverage probability of the interval.

Disambiguation:

  • NOT threshold: threshold is a target output value; low_high_percentiles governs interval width.

confidence_level

Definition: The nominal coverage probability of the calibrated regression interval, expressed as a fraction in (0, 1). Derived from low_high_percentiles as (high - low) / 100. For example, low_high_percentiles=(5, 95) yields confidence_level=0.90.

Type and valid range: float in (0, 1). Default: derived from low_high_percentiles=(5, 95), so 0.90.

Applies to: Regression explanation methods. May be passed directly as a caller-specified value instead of low_high_percentiles.

Behavior: When passed directly, CE reverses the formula to reconstruct low_high_percentiles: low = (1 - confidence_level) / 2 * 100, high = 100 - low. This makes it an alternative input to low_high_percentiles, not a derived output.

User-settable: YES — confidence_level and threshold are mutually exclusive alternatives for specifying the regression configuration. Passing both raises a ConfigurationError (enforced by EXCLUSIVE_PARAM_GROUPS in api/params.py).

Disambiguation:

  • NOT confidence (reject path): confidence is the reject coverage target and applies to the reject orchestrator, not regression intervals.

  • NOT significance: significance is the guarded conformity p-value threshold and does not govern interval width.

  • NOT threshold: mutually exclusive alternatives — use one or the other.

confidence

confidence was removed in v0.11.5. Use reject_confidence; v0.11.6 raises ConfigurationError for the legacy alias.

Definition: The reject coverage target — the minimum calibrated probability required to accept a prediction. A prediction with conformal p-value below 1 - confidence is rejected.

Type and valid range: float in (0, 1). Default: 0.95.

Applies to: predict_reject, apply_policy, and the reject orchestrator.

Behavior: Higher values are stricter (fewer predictions accepted). A value of 0.95 means predictions must have at least a 0.95 conformal coverage to be accepted. Internally the reject path computes epsilon = 1 - confidence as the conformal significance level.

Disambiguation:

  • NOT significance: they are mathematical inverses (significance = 1 − confidence), but significance belongs to the guarded explanation path, while confidence belongs to the reject path. The completed rename keeps both surfaces on the coverage convention while preserving the inverse relationship conceptually.

  • NOT confidence_level: confidence_level is a regression interval width parameter; confidence is a reject decision threshold.

significance

significance was removed in v0.11.5. Use guarded_options=GuardedOptions(confidence=...); v0.11.6 raises ConfigurationError for the legacy kwarg.

Definition: The conformity significance level for the in-distribution guard. The guard accepts a calibration bin if its KNN-based conformal p-value is ≥ significance; bins below this threshold are considered out-of-distribution and pruned from the explanation.

Type and valid range: float in (0, 1]. Default: 0.1.

Applies to: explain_factual(..., guarded_options=GuardedOptions(confidence=...)), explore_alternatives(..., guarded_options=GuardedOptions(confidence=...)), and the internal _guarded_explain module. This parameter remains part of the [EXPERIMENTAL] guarded explanation API, but the legacy significance= kwarg and guarded=True call form were removed in v0.11.5.

Behavior: A larger significance value is stricter (fewer bins accepted). This is the reverse of how reject_confidence works in the reject path: for legacy significance, raising the value makes the guard harder to pass; for reject_confidence, raising the value makes the reject threshold harder to pass. GuardedOptions.confidence resolves this counter-intuitive direction by expressing both surfaces as coverage values where higher = more inclusive.

Disambiguation:

  • NOT confidence (reject path): mathematical inverse (significance = 1 − confidence), different path.

  • NOT confidence_level: unrelated; confidence_level governs regression interval width, not conformity testing.