Human Review Calibration
97% overall accuracy can hide catastrophic failures on specific segments. Learn to validate by document type, calibrate confidence, and prioritise reviewer capacity.
The question with human review is not *whether* to use it but *how to allocate limited reviewer capacity* to maximise accuracy while minimising cost. Get the allocation wrong and you either burn reviewer hours on easy cases or automate segments that are silently failing.
The aggregate metrics trap
A system reporting 97% overall accuracy can be hiding catastrophic failures on specific document types. The aggregate averages away the segments that matter most.
| Document type | Date accuracy |
|---|---|
| Standard invoices | 99.5% |
| Scanned PDFs | 72.4% |
| Handwritten receipts | 60.1% |
| International formats | 45.2% |
Automate everything because overall accuracy is 97%. The aggregate averages away a 45% segment that ships straight to production broken.
if overall_accuracy >= 0.95: automate_all() # hides the 45% segmentValidate accuracy per document type and field, and automate only the segments that clear the bar on their own.
for seg in segments:
if accuracy[seg] >= threshold: automate(seg)A 97% headline number can conceal a 45% segment. Always validate accuracy by document type AND field before deciding what to automate — the average is exactly what hides the failures.
Stratified random sampling
Sampling serves two purposes: ongoing per-segment accuracy measurement and novel error detection. High-confidence items are automated, so if the model develops a new error pattern that hits high-confidence extractions, only stratified sampling across all segments will catch it. Sampling only the low-confidence queue leaves the automated path blind.
The automated, high-confidence path is where undetected drift is most dangerous. Stratified sampling must cover every segment, including the ones you already trust.
Field-level confidence calibration
Raw confidence scores must be validated against labelled ground truth, because they mean different things per field. A model reporting 0.90 on date fields may be right 94% of the time, while 0.90 on amount fields is right only 82% of the time. Calibrate thresholds per field, then route accordingly.
- Above the calibrated threshold → automate (with stratified sampling).
- Below the threshold → send to human review.
- Ambiguous zone → prioritised review.
Prioritise reviewer capacity
Route the highest-uncertainty items to reviewers first using a dynamic queue, not a chronological one. Prioritise low-confidence fields, ambiguous source documents, poor-accuracy document types, and fields that express uncertainty. Spreading capacity evenly wastes it on cases that did not need a human.
Measure accuracy by document type and field, calibrate confidence against labelled data, set calibrated thresholds, add stratified sampling, and only then reduce human review. Cutting reviewers before validation is the classic mistake.
How the exam will try to trick you
The distractors below look right under time pressure — learn the tell.
- The trap
Automate all high-confidence extractions because overall accuracy is 97%.
Correct answerValidate accuracy per document type and field, and automate only segments that clear the bar.
Why: Aggregate metrics average away a catastrophic segment — 97% overall can hide a 45% document type.
- The trap
Send only low-confidence extractions to human review.
Correct answerUse stratified random sampling across all segments, including automated high-confidence items.
Why: A novel error on the automated high-confidence path is exactly what selective sampling never catches.
- The trap
Route decisions on raw model confidence scores directly.
Correct answerCalibrate confidence against a labelled validation set, per field, before thresholding.
Why:
0.90may mean 94% accuracy on dates but only 82% on amounts — the score is field-dependent. - The trap
Spread reviewer capacity evenly across all extractions.
Correct answerDynamically prioritise the highest-uncertainty items for review first.
Why: Uniform distribution wastes scarce reviewer time on easy cases that never needed a human.
Key takeaways
- Aggregate accuracy hides segment-level failures — validate by document type and field before automating anything.
- Stratified sampling must cover high-confidence automated items to catch novel error patterns.
- Raw confidence scores are field-dependent; calibrate them against labelled ground truth before using thresholds.
- Route highest-uncertainty items to reviewers first via a dynamic queue, not chronological order.
- Never distribute reviewer capacity evenly — prioritise low-confidence and poor-accuracy segments.
- Reduce human review only after the full measure-calibrate-threshold-sample sequence.
Frequently asked questions
Why is a 97% overall accuracy figure not enough to justify automation?+
Aggregate accuracy is an average that can mask catastrophic failure on high-impact segments. A system at 97% overall may extract dates from standard invoices at 99.5% but from international formats at only 45.2%. Automating on the headline number ships those broken segments to production. You must validate accuracy broken down by document type and field before deciding what is safe to automate.
Should human reviewers only check low-confidence extractions?+
No. Reviewing only low-confidence items leaves the automated high-confidence path unmonitored, which is exactly where a novel error pattern would go undetected. Use stratified random sampling across all segments — including high-confidence automated ones — for ongoing measurement, and separately route the highest-uncertainty items to reviewers first via a dynamic priority queue.