Structured-output confidence and selective review

Last updated: 5 Oct 20265 min read

tutorial

A field extractor should route uncertain or conflicting spans to review under a measured capacity budget instead of presenting every decoded path as equally certain.

Define the decision unit

The intake team may need one accepted policy ID per claim, not a probability for each token. Multiple candidate IDs, an unreadable scan or a policy mismatch can make the whole claim uncertain even when one span has a high local score. Aggregate at the field or claim level and state which mistakes the reviewer can correct.

Calibrate on held-out development claims

A decoder path score is affected by note length and label transitions. It should not be read directly as a 0–1 probability. If the model offers a field confidence, compare predicted confidence bins with observed exact-match rates on development claims, then reserve later claims for final evaluation. Thresholds must be chosen before seeing final-test outcomes.

Budget review capacity explicitly

A threshold that sends 40% of claims to people may exceed the daily queue. Report coverage, exact-match quality among auto-accepted claims and error rate among reviewed claims at several candidate thresholds. The code uses a declared confidence cutoff and a conflict flag; its scores are sample inputs, not calibrated model outputs.

Guard rare failure modes

A high average confidence can hide errors on OCR-damaged notes, unfamiliar ID prefixes or translated text. Set a hard route-to-review rule for these conditions when evidence is too sparse to trust learned confidence. Error slices show where exceptions are needed.

Keep feedback out of the final test

Reviewer corrections can form later training data after provenance and label checks. They cannot be fed back into the same evaluation window and then treated as an independent test. Version the model, review policy and adjudication rules together for rollback.

Implementation

claim_candidates = [
    {"claim": "case-47", "confidence": 0.93, "conflicting_ids": False},
    {"claim": "case-62", "confidence": 0.78, "conflicting_ids": False},
    {"claim": "case-83", "confidence": 0.96, "conflicting_ids": True},
]

def route_claims(candidates, minimum_confidence):
    decisions = {}
    for candidate in candidates:
        decisions[candidate["claim"]] = (
            "review" if candidate["conflicting_ids"] or
            candidate["confidence"] < minimum_confidence else "auto_accept")
    return decisions

routes = route_claims(claim_candidates, 0.88)
assert routes == {"case-47": "auto_accept", "case-62": "review", "case-83": "review"}

Performance and operating cost

Routing C claims costs O(C) time and O(C) output storage. Calibration and threshold selection require held-out labeled claims, while manual review imposes a capacity and turnaround cost. Budget model inference and reviewer time together; a low compute bill can hide a large queue.

Common Mistakes

Do not treat raw decoder scores as calibrated probabilities.
Do not pick a threshold from the untouched final test.
Do not ignore review capacity when proposing abstention.