Human Review, Confidence Calibration and Source Provenance
👋 In plain words: '97% accurate' sounds safe enough to stop checking by hand, but it can hide a type of document where accuracy is much worse. This lesson shows how to prove it's safe to reduce human review, and how to keep track of where every claim in a report came from.
“97% accurate” can still be unsafe to automate. The final two task statements cover how to prove accuracy before removing humans, and how to keep track of where every claim came from (Task Statements 5.5 and 5.6).
🔑 Key terms
- Aggregate accuracy: One overall accuracy number across all documents.
- Segment: A slice of the data, such as handwritten invoices or a single field.
- Stratified sampling: Checking a random sample from every segment, not just from the whole pile.
- Calibrated confidence: Confidence scores tested against known correct answers, so that 0.9 really means about 90% right.
- Provenance: A record of where each claim came from: source, quote and date.
📖 The story in this lesson
ShopNest's invoice extraction is 97% accurate overall, and finance wants to stop human checks. Ravi is asked to prove whether that is safe.
💡 Think of it like this
A 97% class average can hide one student who is failing. Before you stop marking homework, check each student's grades, not just the average.
🚧 Aggregate accuracy hides weak segments
97% overall can mask 70% on handwritten invoices or on one particular field. Before reducing human review, analyse accuracy by document type and by field.
🌍 Real-world example: Handwritten invoices at 70%
Ravi splits accuracy by document type. Typed invoices are 99% accurate, but handwritten ones from small suppliers are only 70%, and the total field on them is often wrong. Human review stays on for handwritten invoices and is reduced for typed ones.
📏 Measuring and routing
- Stratified random sampling of high-confidence extractions measures ongoing error rates and catches new failure patterns.
- Field-level confidence scores must be calibrated against a labelled validation set before you set review thresholds.
- With limited reviewer capacity, route low-confidence and ambiguous or contradictory documents to humans first.
🌍 Real-world example: Spot checks that catch new problems
Every week a random sample from each invoice type, including high-confidence ones, goes to a human. When a supplier changes its layout, the sample catches the new errors within days. Low-confidence and contradictory invoices always go to reviewers first.
🔗 Provenance in multi-source synthesis
- Summaries drop attribution — require claim → source mappings (URL or document, excerpt, date) that synthesis must preserve and merge.
- Conflicting figures from credible sources: keep both, attributed, and annotate the conflict. Never pick one arbitrarily or average them.
- Publication and collection dates stop a change over time from being misread as a contradiction.
- Separate well-established from contested findings in the report.
- Render content by type: financial data as tables, news as prose, technical findings as structured lists.
🌍 Real-world example: Two sources, two numbers
One report says a carrier's on-time rate is 94% and another says 88%. Instead of picking one or averaging them, ShopNest's research summary shows both, with source and date. The dates reveal the 88% figure is a year older.
🛠️ Try it yourself (about 8 minutes)
- Take any accuracy number from your work, such as test pass rate or data quality.
- Split it by two categories, such as browser and platform.
- Find the weakest segment and write down its number.
✅ You are done when: you found at least one segment that is clearly worse than the overall figure, or showed that none is.
🧪 Exam-Style Practice Questions
Try these before opening the answers. They use the same scenario style as the real exam.
Question 1. How should you measure error rates among high-confidence extractions on an ongoing basis?
- A. Run monthly full re-extractions
- B. Review only low-confidence items
- C. Assume they're correct
- D. Use stratified random sampling for human verification
Show answer
Answer: D. It detects drift and new error patterns. C is blind trust. B misses confident errors. A is expensive and doesn't measure accuracy.
Question 2. How should field-level confidence thresholds for routing to human review be set?
- A. Ask the model for the threshold
- B. Calibrate them against a labelled validation set
- C. Pick 0.8 by intuition
- D. Copy another team's value
Show answer
Answer: B. Calibration needs ground truth. A, C and D are arbitrary.
Question 3. A synthesis report converts quarterly revenue figures, news events and technical findings into uniform paragraphs, and readers struggle. What's better?
- A. Use JSON for everything
- B. Render each type appropriately: financial data as tables, news as prose, technical findings as structured lists
- C. Use bullet points for everything
- D. Keep everything as prose
Show answer
Answer: B. Match the format to the content type. A, C and D force one uniform format.
📝 Assignment
Do these tasks in order. Each one has a Done when check, so you know when it is complete. The stretch task is optional but close to what the exam expects from an architect. Keep your answers in one notes file or Git repository, and revisit them in your final week of revision together with the cheat sheet in the last lesson of this course.
- Task 1: Break a sample accuracy figure down by document type and field and find the weakest segment.
✅ Done when: Your breakdown shows at least two segments and identifies the weakest by name and accuracy. - Task 2: Design a routing rule that sends low-confidence and contradictory documents to reviewers first.
✅ Done when: Your rule uses a calibrated threshold and sends contradictory documents to review regardless of confidence. - Stretch: write a synthesis output that keeps two conflicting statistics with sources and dates.
✅ Done when: Both statistics appear with their source and date, and the text says plainly that they conflict.
❓ Frequently Asked Questions
🤔 Is 97% accuracy enough to remove human review?
Not until you check accuracy by document type and field, because the overall figure can hide weak segments.
🤔 What is stratified sampling used for?
Measuring error rates in high-confidence outputs over time and spotting new error patterns.
🤔 How should confidence thresholds be set?
Calibrate them against a labelled validation set, not by intuition.
🤔 What if two sources disagree?
Keep both values with attribution and annotate the conflict rather than picking one.
🏁 Summary
- Segment accuracy by document type and field before automating.
- Stratified sampling plus calibrated confidence drive review routing.
- Keep claim-source-date mappings; preserve and annotate conflicts.
Next, continue with Claude Architect Hands-on Labs: 8 Practical Exercises.
Independent study material by SQA Tools. Not affiliated with, endorsed by or sponsored by Anthropic. Claude is a trademark of Anthropic. Mock questions are original practice items, not real exam content. Exam facts checked against Anthropic's official exam guides and certification FAQ on 11 November 2026.