What evaluation hooks should a mismatch environment expose for reset, traces and scoring?

Evaluation hooks for a mismatch environment work best when they read state and traces, not the agent's final message.

**Reset.** Each run starts from a named snapshot, loaded by snapshot_ref. A reset restores all tables, case status, counters and the clock. Seed any generated identifiers, such as adjustment numbers, so repeated runs of the same trajectory give identical state.

**Trace capture.** Log every tool call with timestamp, tool name, arguments, result code and the record IDs read or written. A JSONL trace file, one call per line, is simple to replay or diff.

**State diff.** After a run, compute a diff of the tables against the starting snapshot: rows inserted, updated and deleted, by table and key. Predicates run against this diff, for example 'no update to supplier_invoice.amount' or 'exactly one reversal row linked to the original posting'.

**Predicates.** Each acceptable or unacceptable end state in the task is a predicate ID with a deterministic implementation. It takes the diff, the trace and the policy table as input and returns pass or fail with a reason string.

**Scoring output.** Emit one record per run: case_id, run_id, predicate results, tool-call count, policy errors hit, and whether the case reached a valid end state. Keep partial results so you can separate diagnosis from action.

Clock handling is a common trap. If tolerances depend on dates, freeze the environment clock, or the same trajectory may score differently on different days.