Domain 5 · Lesson 3 · Context & Reliability (15%)
Human Review, Confidence & Provenance
Task Statements 5.5 & 5.6 — calibrating who reviews what, and not losing the sources.
Course progress: Domain 5 ▸ lesson 3 of ~3
Human review & confidence calibration (5.5)
Aggregate metrics lie
"97% overall accuracy" can mask poor performance on a specific document type or field. Before automating high-confidence extractions, segment accuracy by document type and field — a great average can hide a 60%-accurate segment.
- Stratified random sampling of high-confidence extractions → measure ongoing error rate and catch novel error patterns.
- Field-level confidence scores, calibrated on a labeled validation set, to route review attention.
- Route low-confidence or ambiguous/contradictory-source extractions to humans, prioritizing limited reviewer capacity.
Note the contrast with Domain 5.2: self-reported confidence for escalation is unreliable, but field-level confidence calibrated against labeled data is a legitimate routing signal for extraction review. Calibration against ground truth is the difference.
Provenance & uncertainty in synthesis (5.6)
Preserve claim→source mappings
Attribution is lost when findings are summarized without preserving the claim-source mapping. Require subagents to output structured mappings (source URL, doc name, relevant excerpt) that downstream agents preserve and merge through synthesis — don't compress them away.
- Conflicting statistics from credible sources → annotate the conflict with attribution; don't arbitrarily pick one value. Complete document analysis with both values included and annotated, letting the coordinator reconcile.
- Temporal data → require publication/collection dates so a time difference isn't misread as a contradiction.
- Structure reports to distinguish well-established from contested findings, preserving original source characterizations and methodology.
- Render by content type — financial data as tables, news as prose, technical findings as structured lists — not one uniform format.
Check yourself
Your extraction system reports 97% aggregate accuracy and you want to reduce human review. Before automating, you should:
- Automate now, since 97% comfortably clears the accuracy bar
- Segment accuracy by document type and field to find weak spots
- Raise the confidence threshold uniformly across all document types
- Ask the model to self-report confidence and trust its high scores
Correct: option 2. Aggregate accuracy can hide a poorly-performing segment; analyze by document type and field first. Automating on the average (1) risks that hidden segment; a uniform threshold (3) ignores segmentation; self-reported confidence (4) isn't calibrated.
Two credible sources report different statistics for the same metric. The synthesis output should:
- Choose the value from the source judged most authoritative
- Average the two figures into a single reconciled number
- Present both values, annotated with their source attribution
- Drop both as unreliable since the sources disagree here
Correct: option 3. Preserve both with attribution rather than arbitrarily selecting or averaging; also check publication dates in case it's a temporal difference, not a true contradiction. Dropping the data loses real information.
Decision rules
Before automating → segment accuracy by type+field (aggregate hides gaps). Route review → field-level confidence calibrated on labeled data + stratified sampling. Synthesis → preserve claim-source mappings; conflicts annotated w/ attribution; dates for temporal data; render by content type.