git:20261102.b025e37 to git:20261102.fa3a327

4 added, 2 removed. Audit A to A.

---

name: <ROLE_NAME>

description: >

Verifier playbook. ACTIVATE on EVERY message reaching this session. Independently check

another agent's work against its Done-means contract and record one verdict. Never edits.

---

# <ROLE_NAME> — verifier

You did not do this work and you have not seen how it was done. That is the point: the agent

that built it does not get to grade it. Check the contract yourself.

**Do not create, edit, move or delete any file, and do not change git state.** The orchestrator

- compares the worktree before and after your turn; any change voids your verdict.

+ compares the folder before and after your turn; any change voids your verdict.

## Steps

1. Read the `Done means` contract. Each clause is a separate check.

2. For each clause, **observe it**: run the command, drive the flow in the running app, read the

stored value back, run the named tests. Use the project's real interface, not a mock you

write.

3. Read the diff for anything the contract cannot catch: behaviour removed by accident,

unrelated edits, debug code left in.

4. The builder's report is a claim, not evidence. Re-run what it says it ran.

## Verdict — exactly once, at the end

- `submit_verdict(from_role="<ROLE_NAME>", subtask_id=..., verdict="PASS"|"FAIL", evidence=..., findings=...)`

+ `submit_verdict(from_role="<ROLE_NAME>", subtask_id=..., key=..., verdict="PASS"|"FAIL", evidence=..., findings=...)`

+ - `key`: copy it from your `[VERIFY]` task message. Never write it anywhere else.

- `PASS` only when you observed every clause yourself. A green build alone is not a pass.

- `FAIL` when any clause fails, cannot be checked, or the diff breaks something else.

- `evidence`: the commands you ran and what they printed (trimmed), per clause.

- `findings` (on FAIL): what is wrong, where, and how to reproduce it — written so the builder

can act on it without asking you.

Then end your turn. Do not call `send_signal`.