What You Need to Know
Human review is a safety check for automated extraction and classification systems: people check the model's output before anyone relies on it. A confidence score is the model's estimate that an extracted field is right. Calibration checks those scores against known correct answers. The core challenge is to use the limited capacity of the reviewers to get the highest accuracy at the lowest cost.
This lesson uses one made-up example throughout: a plastering firm that uses a model to read its paperwork. The paperwork is standard supplier invoices, handwritten receipts from plasterers on site, scanned old job records and invoices from new suppliers abroad. A few office staff can check extractions, but they cannot check all of them.
The Aggregate Metrics Trap
Trusting aggregate metrics is the most dangerous misconception in production extraction systems. A 97% overall accuracy figure can hide catastrophic failure rates on specific document types or fields. Suppose the firm's system reports 97%, and the owner approves full automation for all high-confidence extractions. For dates, standard invoices reach 99.5% accuracy. Handwritten receipts reach 60%. Scanned PDFs with poor OCR reach 72%. International documents with non-standard formatting reach 45%.
| Document Type | Date Accuracy | Amount Accuracy | Name Accuracy |
|---|---|---|---|
| Standard invoices | 99.5% | 98.2% | 97.8% |
| Handwritten receipts | 60.1% | 55.3% | 71.2% |
| Scanned PDFs | 72.4% | 69.8% | 80.1% |
| International formats | 45.2% | 52.1% | 63.4% |
| Aggregate | 97.0% | 96.1% | 95.8% |
Most of the documents are standard invoices. So the average, weighted by volume, is near the figure for standard invoices. This average hides three document types with unacceptable accuracy. In those segments, errors often have the highest business impact. The segments are handwritten receipts from field staff, international invoices from new suppliers and scanned historical documents for compliance audits.
Always validate accuracy by document type and field segment before automating. Never decide what to automate from aggregate metrics alone.
Stratified Random Sampling
After that one-off validation, you still need to verify accuracy at regular intervals. Stratified random sampling selects a representative sample from each stratum (document type, confidence band, field type), and humans verify it.
Sample high-confidence extractions as well as low-confidence ones. Low-confidence items are already routed to human review. High-confidence items are automated, so nobody sees them. If the model develops a novel error pattern that affects high-confidence items, only stratified sampling will detect it. A new supplier's changed invoice layout could give the firm wrong totals at high confidence.
Sampling has two purposes. It measures accuracy at regular intervals, to confirm that each segment maintains its validated accuracy rate. It also detects novel error patterns, which are failure modes absent from the original validation set. Without sampling, a systematic error on a new document format stays unnoticed until downstream business processes fail.
Field-Level Confidence Calibration
The model can output confidence scores per field. For an invoice extraction, the output might look like this:
{
"vendorName": {"value": "Acme Corp", "confidence": 0.98},
"invoiceDate": {"value": "2024-03-15", "confidence": 0.95},
"totalAmount": {"value": "$1,247.83", "confidence": 0.72},
"lineItems": {"value": [...], "confidence": 0.61}
}The raw confidence scores from the model are not calibrated. A reported 0.95 might be correct 88% of the time on certain field types and 99% on others. The score is relative, not absolute, so equal scores on different fields do not promise equal accuracy.
Calibration requires labelled validation sets (ground truth data). Run the model on documents with known correct extractions. Compare its confidence scores with real accuracy. Build a calibration curve. For the firm, the labelled validation set is past invoices and receipts with the right values recorded. The curve might show that 0.90 on date fields is correct 94% of the time. It might also show that 0.90 on amount fields is correct 82% of the time.
Calibrated thresholds then drive routing:
- Fields above the calibrated threshold: automated, with stratified sampling
- Fields below the calibrated threshold: human review
- Fields in the ambiguous zone: prioritised human review
Reviewer Capacity Prioritisation
Human reviewers are expensive and limited. Route the highest-uncertainty items to reviewers first. That means:
- Low model confidence fields
- Extractions from ambiguous or contradictory source documents
- Document types with historically poor accuracy
- Fields where the model expresses uncertainty (for example, multiple possible interpretations)
Do not spread reviewer capacity evenly. An even split wastes time on high-confidence items that the model handles well. An even split also leaves too little capacity for the uncertain items that need human judgement. The firm's blurred site receipts wait while staff check clean invoices.
The priority order should be dynamic. The review queue stays ordered by uncertainty as documents arrive. When a reviewer finishes one item, the next item is the highest-uncertainty item that remains, not the next item in chronological order.
Validation Before Automation
The order of the steps is important:
- Measure accuracy by document type and field segment, not aggregate.
- Calibrate confidence scores using labelled validation sets.
- Set calibrated thresholds for automation versus human review.
- Implement stratified random sampling for ongoing verification of automated extractions.
- Only then reduce human review on segments that demonstrate consistent, validated accuracy.
The trap is to go directly to step 5 on the basis of aggregate metrics. Every step exists to prevent a specific failure mode. The result for the firm might be that it automates standard invoices while the office staff check the site receipts.
In the exam
The exam tests when and how to deploy human reviewers effectively, and how to allocate their limited capacity. Wrong answers trust aggregate accuracy, sample only low-confidence items, use raw confidence scores, or spread reviewers evenly.
Key Concept
97% aggregate accuracy can hide 40% error rates on specific document types. Validate by document type and field segment, calibrate against labelled validation sets, sample high-confidence extractions and prioritise the highest-uncertainty items.
Exam Traps
Using aggregate accuracy (e.g., 97%) to justify automating all high-confidence extractions
Aggregate metrics hide per-type performance. 97% overall can mean 40% accuracy on specific document types. Validate by document type and field segment before automating.
Only sampling low-confidence extractions for human review
High-confidence extractions are automated. If a novel error pattern affects them, only stratified random sampling of high-confidence items will detect it.
Using raw model confidence scores without calibration
Raw confidence scores are not calibrated. 0.90 confidence on dates might mean 94% actual accuracy, while 0.90 on amounts might mean only 82%. Calibrate using labelled validation sets.
Spreading reviewer capacity evenly across all extractions
Even distribution wastes time on high-confidence items. Prioritise limited reviewer capacity on the highest-uncertainty items where human judgement adds the most value.
Practice Scenario
A structured data extraction system achieves 97% overall accuracy across all document types. The team proposes automating all extractions where model confidence exceeds 95% to reduce human review costs. What is the critical risk in this approach?
Build Exercise
Build a Confidence-Calibrated Review Router
What you'll learn
- Recognise the aggregate metrics trap: 97% overall accuracy can hide 40% error rates on specific document types
- Implement accuracy tracking broken down by document type AND field segment
- Calibrate raw confidence scores using labelled validation sets to produce reliable routing thresholds
- Design stratified random sampling that includes high-confidence extractions for ongoing verification
- Prioritise limited reviewer capacity on the highest-uncertainty items with dynamic queue ordering
- Create a mock extraction system that outputs field-level confidence scores for different document types (invoices, receipts, scanned PDFs, international documents)
Why: Field-level confidence scores are the foundation of intelligent review routing. The exam tests that raw model confidence is not calibrated and must be validated against ground truth before use. Building the mock system gives you data to calibrate against.
You should see: An extraction function that returns each field with its value and a confidence score between 0.0 and 1.0. The system should process at least 4 document types with noticeably different confidence distributions per type.
- Implement accuracy tracking broken down by document type and field segment — not just aggregate metrics
Why: Trusting aggregate metrics is the most dangerous misconception in production extraction systems. 97% overall accuracy can hide catastrophic failure rates on specific document types because standard invoices dominate the volume. The exam tests that you must validate by document type AND field segment before automating.
You should see: An accuracy table showing each document type and field combination separately. Standard invoices should show 95%+ accuracy while handwritten receipts and international documents show 40-70%. The aggregate should look excellent (90%+) despite the poor per-type numbers.
- Build a calibration module that takes a labelled validation set (ground truth) and produces calibrated confidence thresholds per field type per document type
Why: Raw model confidence scores are not calibrated. A model reporting 0.90 confidence might actually be correct 94% of the time on date fields but only 82% on amount fields. Calibration using labelled validation sets is required before confidence scores can drive automated routing decisions.
You should see: A calibration curve for each field type per document type, mapping reported confidence ranges to actual accuracy percentages. The curve should reveal that the same confidence score means different things for different field-document combinations.
- Implement stratified random sampling that selects high-confidence extractions for ongoing verification, sampling proportionally across all document types
Why: High-confidence extractions are automated and not reviewed. If the model develops a novel error pattern affecting high-confidence items, only stratified sampling will catch it. Sampling only low-confidence items means nobody sees systematic errors in automated extractions.
You should see: A sampling function that selects a representative subset from each stratum (document type and confidence band), including samples from the high-confidence automated extractions. The sample should be proportional to the volume in each stratum.
- Build a review router that prioritises limited reviewer capacity on the highest-uncertainty items, dynamically reordering the review queue as new extractions arrive
Why: Human reviewers are expensive and limited. Spreading capacity evenly across all extractions wastes time on high-confidence items while leaving insufficient capacity for uncertain items that need human judgement. Dynamic priority ordering ensures the most uncertain items are always reviewed first.
You should see: A priority queue that orders items by uncertainty (lowest confidence first), dynamically reorders as new extractions arrive, and serves the next-highest-uncertainty item to each available reviewer. The queue should never serve items in chronological order.