Ranking

• open: a person opens the report. action: a person or their agent acts on it. fix: someone fixes the problem it describes. Call inbox-ranking-outcomes-list for the current meaning, horizon and base rate of each outcome, because they follow the served model.
• A prediction says what may happen, never what did. A high fix probability does not mean the report is fixed: read status and the PR fields for that.
• A lift is the probability divided by the outcome’s base rate. 1.0 is an average report, 2.7 is 2.7x as likely. Compare lifts across outcomes. Raw probabilities do not compare, because each outcome has its own base rate.
• readable: false means the training run could not measure the holdout AUC of that outcome. Rank and filter on that outcome with caution.
• status is scored, stale (the title or summary changed after scoring, so the predictions describe the old text), unscored (no score yet, which is not a low score), or not_available (the user cannot see the ranking model, so ordering and filtering by it return 400).
• inbox-reports-retrieve also names the model that made the predictions. Predictions from different models do not compare.

Recipes:

• Likely to get a fix: view=actionable with ordering=-ranking_fix.
• Hidden gems: priority=P3,P4 with ranking_min_lift=fix:2.
• Probably noise: reports with a low lift on all three outcomes are candidates to dedupe or dismiss. Read them before you act: a low lift is a hint, not a verdict.
• Add search or source_product to any recipe to narrow it to a topic.

ranking_min_lift and ranking_min_probability take comma-separated <outcome>:<value> pairs, and a report must meet every pair. Unscored and stale reports never match.

Caveats:

• The predictions are model estimates, and the top end over-predicts. Use them to rank and filter, not as literal odds.
• Most fix positives are also action positives, so do not add the two together.