Shadow deployments with explicit stop rules

Shadow mode is only useful when the team agrees on which metrics invalidate the experiment. We document thresholds in the program wiki, not in slide decks.

Latency budgets include cold-start penalties and GPU queue depth, captured every five minutes during the pilot window.

When a stop rule trips, the on-call mentor posts a single-thread incident note with timeline, data slices, and the decision to roll forward or revert.