A reason-coded review order

1. Inspect high-volatility items first.

   Position volatility was the second-largest positive permutation contribution in the fitted forest. Use it to order diagnosis of query mix, tracking continuity, and intent fit before editing (Figures 2 and 4).

   POS-VOL-HIGH

2. Read CTR relative to position.

   The position-bucket curve shows sharp click sparsity outside the leading buckets, while CTR gap also appears among the largest Ridge coefficients. Review snippet and intent alignment as a testable hypothesis; raw CTR alone is a poor quality verdict (Figures 1 and 3).

   CTR-GAP-LOW

3. Require a sustained movement pattern.

   Trend and volatility add temporal context to current position, but held-out absolute fit was weak. Escalate only when a pattern persists with adequate impressions, and keep the reason code attached to the review (Figures 2 and 4).

   POS-TREND-DOWN

4. Keep query breadth descriptive.

   The query correlations use a window that overlaps the outcome. Use breadth to frame an analyst question about dependence on a narrow query set; do not include it in a forward score (capstone notebook, cell 9 output).

   QUERY-BREADTH-DESCRIPTIVE

5. Measure interventions prospectively.

   June R² was near zero even when rank ordering improved. Record eligibility, reason code, intervention date, and a predeclared follow-up period; compare reviewed items with an eligible holdout before attributing impact (Figure 2).

   TEST-BEFORE-CLAIM