Interface PolicyEngineOptions

Optional audit logger for policy decisions and evaluator failures.

Opt-in zero-leak streaming: hold back a trailing window of each text channel so a violating span is caught BEFORE any of it is emitted. Windows come from the registered policies' holdBackChars hints (per channel, max wins). Released text arrives as modified delta chunks; intermediate 'object' snapshots are suppressed (only a passing 'object-result' is emitted); the held tail is flushed at the channel's end chunk ('text-end'/'reasoning-end') and, as a backstop for streams without end chunks, at 'finish' — both through the runner's reprocess convention, so the flush precedes its end marker. The guarantee is therefore PER SEGMENT: before release, every applicable output policy evaluates the channel's whole text once more, including the held tail. A host evaluator receives this call even when the text is unchanged; a host cadence evaluator must decide for itself whether to classify. A match completing across segment boundaries (multi-step or multi-text-block runs) aborts the stream after earlier segments were already released — bounded by the window for string patterns or the whole prior segment for RegExp (Infinity) policies. Default false — evaluated chunks flow through unmodified, and already-emitted earlier chunks of a violating span may have leaked by abort time.

Policies snapshotted at construction and evaluated in array order.

Audit resource. Defaults to the stable processor identifier.