Summary
Key Ideas
L1: Field-level confidence — each extracted field gets its own score (not one per record). Enables granular routing.
L1: Routing threshold — fields above cutoff auto-accepted (flow downstream); below cutoff queued for human review.
L1: Record-level confidence problem — aggregate masks weak spots. Example: record scores 0.85, but date field is 0.40.
L1: Record-level routing — entire record sent to human review if any field is below threshold. Simpler but wastes reviewer effort on already-correct fields.
L1: Field-level routing — each field routed independently. Reviewer only touches fields needing help.
L1: Schema design — each field object has value (extracted content) and confidence (score 0.0–1.0). Model populates both in one call.
L1: Field selection — not every field must carry a confidence score; choose based on downstream risk.
L1: High-stakes fields — dates, monetary amounts, entity identifiers typically need confidence.
L1: Borderline zone — some systems add a secondary check for scores near the threshold.
L1: Mixed confidence example — name 0.95, total 0.88, date 0.38 → name & amount auto-accepted, date→review.
L1: Failure mode — teams trust confidence scores without calibration. Model may report 0.90 on a field where it’s only 60% correct, letting errors flow into production.
L1: Calibration — verifying whether scores match real-world accuracy on labeled data.
L2: Validation set: held-out labeled samples used to measure extraction accuracy, never used for training or prompt adjustment.
L2: Calibration: a model is calibrated when its reported confidence matches its actual accuracy (e.g., 0.90 → correct 90% of the time).
L2: Reliability diagram: plots model’s reported confidence (x-axis) vs empirical accuracy (y-axis); perfect = diagonal; above diagonal = overconfident; below = underconfident.
L2: Overconfidence is more dangerous than underconfidence — an overconfident model silently feeds errors into auto-accept (says 0.90 but only 60% correct).
L2: Underconfidence wastes reviewer capacity (says 0.60 but actually 90% correct).
L2: Four-step calibration check: (1) group predictions into 0.1-width confidence buckets; (2) compute empirical accuracy per bucket; (3) plot bucket midpoint vs accuracy with diagonal; (4) identify gaps.
L2: Routing threshold: start from policy’s accuracy target (e.g., 95%) on y-axis, trace horizontally to calibration curve, read down to x-axis — fields at/above threshold auto-accept, below go to review.
L2: Building a validation set: representative production samples (not cherry-picked) + ground-truth labels from human experts; must stay held out.
L2: Calibration is a recurring operational task — model upgrades or prompt changes require rerunning on the same validation set.
L3: Stratified sampling deliberately oversamples outputs most likely wrong — weighting by low confidence, rare categories, and high stakes.
L3: Random sampling (e.g., 1% of 10K outputs/day) dilutes signal: ~95 correct vs ~5 errors per 100 reviews, so most reviewer effort is wasted.
L3: Three stratification dimensions: (1) confidence — tunable threshold (e.g., 0.70) below which sampling rate rises; (2) category balance — per-category minimum quotas prevent high-volume categories from crowding out rare but risky ones; (3) stakes — some outputs require human review regardless of confidence because error consequences are severe.
L3: Combining dimensions: confidence × category (prioritize low confidence AND underrepresented categories); confidence × stakes (high stakes → always review); full three-way via scoring function.
L3: Reviewer burnout is a system design failure (not a people problem): processing mostly correct outputs recalibrates cognitive thresholds → reviewers approve carelessly. Stratification counters this by raising error density in the review queue.
L3: Content moderation example: 5 categories; rare category has 20% error rate vs 2% for common ones. Random sampling under-invests in the highest risk outputs. Solution: per-category sample quotas.
L4: Purpose of Review Queue: Dedicated work surface where low-confidence outputs land for human inspection before reaching downstream consumers.
L4: Routing Decision: Model confidence score + stakes determine path — low confidence or high stakes → queue; high confidence/low stakes → continue directly downstream.
L4: Queue Priority Ordering: Items sorted by lowest confidence + highest stakes first.
L4: Three Reviewer Actions (exactly three): Approve — output correct, released downstream, logged with timestamp + reviewer. Edit — output wrong, reviewer provides corrected version; original+correction pair feeds future training. Reject — output wrong, no correction; item discarded, input type flagged for analysis.
L4: Design Principles: (1) Single action surface — input, output, confidence, suggested fix on one screen; no tab switches. (2) Context first. (3) Queue priority automatic. (4) Audit trail on every action.
L4: Bypass Policy: Rule letting certain outputs skip the queue; must be logged and audited like a normal review action. Three scenarios — emergency override, high-confidence floor bypass, reviewer unavailability (fallback with warning flag).
L4: Operational Metrics: Queue depth (items waiting) and queue latency (wait time before review).
L4: Feedback Loops: Approve → positive training signal. Edit → high-value corrected pair from real production data. Reject → identifies poorly-handled input classes.
L4: Synchronous vs Asynchronous Review: Synchronous — downstream waits, pipeline stalls if no reviewer. Asynchronous — placeholder returned immediately, deferred fulfillment.
L5: Provenance tracking attaches a verifiable source pointer to every claim in synthesized output — each sentence needs a traceable origin.
L5: Claim-to-source mapping is the data structure linking each output claim to the specific source documents it came from.
L5: Inline citations expose claim-to-source mapping in readable output next to the claim, unlike end-of-paragraph bibliographies.
L5: Core problem: LLMs synthesize multiple sources into fluent prose but make sources invisible; reviewers can’t tell which sentence came from which document.
L5: Inline vs end-of-paragraph: inline = per-claim accountability; end-of-paragraph = paragraph-level at best.
L5: High-stakes synthesis (regulatory filings, clinical summaries, legal analysis) needs inline citations. End-of-paragraph is acceptable only when the cost of a wrong claim is low.
L5: Schema — each record holds: (1) claim text, (2) list of source identifiers, (3) optional confidence score. Claims become first-class data objects (queryable, auditable, independently updatable).
L5: Four reasons for provenance tracking: (1) Trust, (2) Auditability, (3) Compliance (GDPR, EU AI Act), (4) Error correction.
L5: Verification workflow: sample-based, covers 10-15% of claims. Catches fabricated citations and distorted paraphrases.
L5: Source identifier must be unique and stable — unstable identifiers break the provenance trail.
L5: Fabricated citation — the model invents a source that doesn’t match actual content or doesn’t exist. Most dangerous failure mode in synthesis.
L5: Fix: every citation must be verified against the actual source before output is finalized.
L6: Conflict Annotation: Practice of explicitly surfacing disagreement in synthesis output rather than hiding it from the reader.
L6: Resolution Strategies (three options): (1) Prefer Newer — more recent source wins (good for regulations, versions, pricing). (2) Prefer Authoritative — designated high-trust sources outrank others (requires predefined authority hierarchy). (3) Escalate to Human — when neither rule applies; expensive but necessary for high stakes.
L6: Averaging Anti-Pattern: System silently blends or picks one conflicting source without logging the choice. Reader gets a confident-sounding claim with no indication a choice was made; the discarded source may have been more accurate or authoritative.
L6: Conflict Detection (upstream step, before synthesis): (1) Semantic similarity — find passages discussing the same topic. (2) Entailment checking — test if passage contradicts, supports, or is unrelated. Semantically similar + entailment says contradict → conflict.
L6: Conflict Log: Persistent record of every detected disagreement — logs conflicting claims, source IDs, resolution strategy applied, and selected/discarded claims. Without it, a confident claim is assumed to be consensus; with it, an auditable trail exists.
L6: Presentation by Stakes: High stakes → surface inline next to claim. Medium → footnote or expandable. Full review → conflict summary at end.
L6: Key Aphorism: “Synthesis without provenance is opinion, not analysis.”