# Window Review Policy

This file defines how `tikv-slow` evaluator reviews repeated spikes inside one
incident without changing the top-level mismatch taxonomy.

## Goal

Use `window_reviews` to preserve maintainer-facing detail when:
- one case contains multiple duration rises
- the same root-cause family repeats across multiple windows
- the replay got the overall route right but still flattened repeated spikes
  into one undifferentiated conclusion

## Rules

- keep the existing top-level `match_level` and `mismatch_type` contract
- use `window_reviews` as structured evidence, not as a second verdict system
- when all windows point to the same path and same bottleneck family, say so
  explicitly instead of repeating independent root-cause wording for each
  window
- when different windows point to different mechanisms, record that split in
  `window_reviews` and reflect it in `comparison_notes`
- do not invent exact timestamps unless they were present in the ticket,
  comments, or maintainer-supplied evidence

## First-Version Scope

The first version only needs to support:
- window labels such as `window-1`, `window-2`
- optional `start_time` / `end_time`
- a short `summary`
- an `inferred_path`
- an `inferred_bottleneck`
- evidence bullets

This keeps the representation simple enough to evolve without breaking the
existing evaluator contract.