What reset, snapshot and grader hooks would an evaluation harness need?

A booking environment that supports repeatable evaluation can expose a small set of hooks.

- `reset(seed, task_id)` restores the starting snapshot for the task, including any injected faults such as a scheduled timeout. - `snapshot()` returns the full record state as Parquet or JSON, taken before and after the run. - `diff(before, after)` lists created, changed and deleted records by table and key. - `action_log()` returns each tool call with arguments, result, error code and order. - `grade(task_id)` compares the diff and log with the answer key and returns per-check results.

Grader output can be one JSON object per task:

- `task_id`, `agent_version`, `seed` - `checks`: a list of `{name, passed, detail}` - `tags`: the failure-mode tags exercised - `unexpected_writes`: records changed outside scope

Write results as JSONL so they can be loaded into a notebook or CI report. For regression testing, fix the seed and the agent configuration, then compare check-level results between versions.

Fault injection should be a task setting, not agent-visible, so the agent meets timeouts and rejections as it would in a live integration. Confirm hook names and return formats against your harness during scoping.