Prediction Review Guide

Guide Maps

graph LR
  evaluate["evaluate.py"] --> predictions["predictions.csv"]
  predictions --> queue["review-queue"]
  predictions --> threshold["threshold-review"]
  queue --> release["Release review"]
  threshold --> release

flowchart LR
  question["Which incidents deserve closer review?"] --> predictions["Read the prediction review surfaces"]
  predictions --> compare["Compare misclassifications and borderline rows"]
  compare --> release["Return to the release decision with concrete records"]

Use this guide when aggregate metrics look acceptable but you still need to know which records deserve human attention. The goal is to make predictions.csv, review-queue, and threshold-review work together as one honest review surface.

Review layers

Surface	Best question
predictions.csv	what happened on each promoted eval row
make review-queue	which false positives and false negatives most need immediate inspection
make threshold-review	which promoted predictions are closest to the current decision line
report.md	which rows are worth mentioning in the human release summary

Review rules

• use review-queue when you need known mistakes first
• use threshold-review when the decision line itself is under pressure
• return to raw predictions.csv when one team or incident pattern needs closer inspection
• do not treat aggregate metrics as a substitute for record-level review when the release question is operational

Best companion guides

• read CONTROL_SURFACE_GUIDE.md when the next question is whether threshold changes are still comparable
• read RELEASE_REVIEW_GUIDE.md when the next question is whether record-level evidence changes downstream trust
• read MODEL_GUIDE.md when the next question is whether a record pattern points back to the promoted scoring behavior