ce-reject-policy

name	ce-reject-policy
description	Configure reject and defer decision policies and interpret RejectResult behavior in prediction and explanation flows.

CE Reject Policy

You are configuring reject / defer logic for production decision pipelines. Reject policies let the framework flag, skip, or selectively process instances where the model's prediction is ambiguous, novel, or uncertain — all without modifying the underlying model.

ADR reference: ADR-029 (Reject Integration Strategy).

Load references/reject_policy_examples.md for full code examples.

---

The four policies

from calibrated_explanations.core.reject.policy import RejectPolicy

Policy	Behaviour
RejectPolicy.NONE (default)	No reject logic. Returns the same type as without any policy.
RejectPolicy.FLAG	Process ALL instances. Annotate rejected ones in RejectResult.rejected.
RejectPolicy.ONLY_REJECTED	Process and return only rejected instances.
RejectPolicy.ONLY_ACCEPTED	Process and return only accepted (non-rejected) instances.

Legacy policy strings (deprecated, emit DeprecationWarning):

Old name	Maps to
"predict_and_flag" / "explain_all"	FLAG
"explain_rejects"	ONLY_REJECTED
"explain_non_rejects" / "skip_on_reject"	ONLY_ACCEPTED

Use the enum members directly, not string values, to avoid deprecation warnings.

---

RejectResult envelope

When any non-NONE policy is active, the return type changes to a RejectResult:

from calibrated_explanations.explanations.reject import RejectResult

result.prediction   # calibrated predictions (array or None if policy skips them)
result.explanation  # CalibratedExplanations or None
result.rejected     # boolean mask: True = rejected, False = accepted
result.policy       # RejectPolicy member that generated this result
result.metadata     # dict with telemetry: error_rate, reject_rate, etc.

---

Policy selection guide

When to use	Policy
Audit mode: flag all uncertain, explain everything	FLAG
Focus investigation on uncertain instances	ONLY_REJECTED
Production only-confident mode (skip uncertain)	ONLY_ACCEPTED
Legacy behavior / benchmarking	NONE

---

Out of Scope

• Mondrian group calibration (see ce-mondrian-conditional).
• Calibrated predictions without reject logic (see ce-calibrated-predict).
• Guarded factual explanations (explain_guarded_factual — legacy interface, use explain_factual + reject policy instead).

Evaluation Checklist

• RejectPolicy enum member used (not deprecated string value).
• Return type checked: RejectResult when policy != NONE, plain type otherwise.
• result.rejected mask inspected for actual rejection counts.
• Initialization failure path tested (metadata["init_error"]).
• Per-call override tested alongside explainer-level default.
• Regression: initialize_reject_learner(threshold=t) called before predict_reject.
• Tests use pytest.warns(UserWarning) when fallback warning is expected.