What trace format and grader hooks make conditional-approval evaluation automatic?

Evaluation works best when the harness records enough to score both outcome and process.

**Trace format.** Log each tool call as a JSONL line with `run_id`, `task_id`, `step`, `tool`, `arguments`, `result_status` and `state_hash_after`. The state hash lets a grader detect which step changed the environment.

**Grader inputs.** The grader reads the task's `expected_actions`, `forbidden_actions`, `reason_code` and the final environment snapshot.

**Checks to run automatically:**

- any forbidden tool call, flagged with the step number - gated actions taken while `conditions.state` was not in an allowed set - condition states changed without a preceding read of the linked evidence - final condition and work item states against the expected snapshot - the stated reason matched against `reason_code`

**Result record.** Write one line per task with `outcome`, `premature_action`, `unnecessary_hold`, `wrong_write` and `reason_match` as booleans, plus a free-text `reviewer_note`.

**Reset.** Each task references an `initial_state_ref`. Restore it before every run and, if your agent keeps memory, clear that too. Run each task several times to see variance.

**Time.** If tasks depend on expiry, expose a controllable clock so a run can be repeated deterministically.

Define the exact fields with MirrorFirm in a pilot so they map to your harness.