ce-reject-policy

Reject policies for uncertain model predictions are configured using NONE, FLAG, ONLY_REJECTED, or ONLY_ACCEPTED modes. You can apply these as explainer-level defaults or per-call overrides within your prediction and explanation workflows to govern live ML systems.

A RejectResult envelope is a data structure that exposes the prediction, explanation, rejection mask, applied policy, and telemetry metadata. It wraps the output of explanation flows to standardize how deferred or flagged uncertainty cases are handled in production pipelines.

To flag uncertain predictions instead of dropping them, apply the FLAG policy to your prediction workflow. This configuration marks uncertain cases within the RejectResult envelope while still returning the prediction and explanation data for downstream review.

Yes, you can override the default rejection policy for individual prediction calls. The system supports per-call policy overrides, allowing you to switch between NONE, FLAG, ONLY_REJECTED, and ONLY_ACCEPTED modes on the fly without changing explainer-level defaults.

Yes, defined policy enums, per-call overrides, and calibration outputs are required to function correctly. The reject and defer decision policies integrate directly with calibration outcomes to assess uncertainty and determine whether to skip, flag, or defer cases.

You inspect rejection rates and init failure handling by examining the telemetry metadata within the RejectResult envelope. This metadata tracks rejection metrics and exposes init_error states, allowing you to monitor policy effectiveness and initialization issues.