Automatic vs. Human-Gated Rollback Triggers

You are writing the rollback runbook for a churn-prediction model canary that feeds a retention team's outreach list. Classify each of the following observed conditions as (a) automatic rollback, (b) human-gated review, or (c) not a rollback-relevant signal at all, and justify each:

1. The canary's p99 scoring latency jumps from 40ms to 800ms.
2. The canary flags 30% more users as high-risk than the champion did on the same population, during the week a major product outage happened (which plausibly increased real churn risk broadly).
3. The champion and challenger's flagged cohorts overlap 94%, roughly in line with the champion vs. its own previous-day predictions.
4. A single 5-minute window shows the challenger's mean churn score 3% higher than the champion's; the next four 5-minute windows are back within the normal band.

For each, also state what you'd want logged or checked to make the classification defensible after the fact.