Canonical flow

Document routing

Non-generative classification pipeline for routing documents.

Abstract pattern

Non-generative classification pipeline

The substrate-agnostic proof that PAA can govern a small classifier and still rely on live gating, monitoring, and promotion evidence.

Worker: A routing classifier with no generative model in the loop.
Boundary: Input is the document plus metadata. Output is a label and a routing destination.
Evidence log: Document features, predicted label, evaluator verdict, downstream correction when a human reroutes the item, and the handling outcome.
Evaluator: Inline checks and monitoring gate each routing decision.
Promotion rule: Held-out precision and recall meet the bar and live reroute rate stays low, so the pipeline can route directly.
Demotion rule: Reroute rate climbs or drift fires, which escalates uncertain documents to human triage.
Fallback: Out-of-distribution or low-confidence documents go to human handling.
Lives: Full autonomy for the in-distribution bulk, with an out-of-distribution escalation path always open.

Evaluator detail

What the gate actually checks

Target: Input and output
Technique: Schema validation, out-of-distribution detection, and low-confidence escalation in flight, with downstream reroute monitoring behind the scenes.
Oracle: Invariants for the inline checks and human reroutes as the live signal.
Position: autonomous

Teaching point

The live evaluator and the promotion evidence are different things, and both are required if the automation is going to stay safe.

Six questions

Without PAA: Routing errors surface only as downstream complaints; there is no live signal for drift or confidence degradation; a bad batch ships in full before anyone knows.
What gets gated: Out-of-distribution and low-confidence routing decisions are held for human triage; in-distribution high-confidence decisions run to full autonomy.
What is logged: Document features, predicted label, evaluator verdict, downstream correction when a human reroutes the item, and the handling outcome.
Earns promotion: Held-out precision and recall meet the bar and the live reroute rate stays low enough to confirm that live performance matches offline evaluation.
Triggers demotion: Reroute rate climbs above threshold or a drift monitor fires, expanding the escalation scope and returning uncertain documents to human handling.
Never full-auto: Out-of-distribution documents — the classifier declares uncertainty on these and routes them to human handling as a permanent design constraint, not a temporary state.

This page is linked from the canonical card set on the flows index.