A review is a ledger of concerns

rs_review records a review; it does not perform one. Reading the manuscript and deciding what is wrong with it is your work, in your own reading of the document (rs_read what=file gives you the prose; what=outline names the sections). What the tool gives you is that no concern can be lost, no concern can be closed by wishing, and the two questions a revision round turns on stay separate.

Record what you found, with where you found it

rs_review slug=sleep-memory-review action=panel
  items=[{ "role": "methodology", "concern": "the inclusion criteria are stated but
      no search date is given, so the review cannot be reproduced",
    "severity": "major", "criterion": "reproducibility",
    "evidence": "method", "suggestion": "state the search date and the databases with
      their dates of access" },
   { "role": "domain", "concern": "the synthesis treats slow-wave sleep as one
      mechanism across paediatric and adult samples",
    "severity": "major", "criterion": "evidence", "evidence": "c0148",
    "suggestion": "separate the two samples or say why they are pooled" }]

Recorded round 1 for sleep-memory-review: 2 concern(s)
  f0001  [major, methodology, reproducibility]
  f0002  [major, domain, evidence]

Next: answer each one with `rs_review action=answer id=… status=addressed reply="…"`,
then have a fresh reader check it with `rs_review action=rereview`.

Always give evidence — a section slug, or a claim id, so the two tools meet: the reviewer names c0148 and the author finds the sentence, its status and the source it rests on with rs_read what=claims. A concern with no location is one the author cannot find, and that is the commonest way a review wastes a revision round.

Roles: journal-fit, methodology, domain, devils-advocate, editor, author. Criteria: fit, method, evidence, novelty, clarity, ethics, reproducibility, significance. Severities: major, minor, note. Criterion is what groups the report, severity is what decides the round, and neither is worth arguing about — pick the honest one.

Reviewers read; they do not edit

A reviewer's output is findings, not a redraft. When you dispatch a reviewer subagent, tell it to use read-only tools, to read the manuscript as it stands, and to report concerns with their evidence — not to rewrite a sentence, not to edit a claim, not to run rs_write. Say this in the subagent's own prompt, because nothing here enforces it: the tool has no way to know whether the words it was handed came from a reader or from an editor, so this discipline is yours. A reviewer that silently repairs the document destroys the finding that would have justified the repair.

Two columns, and two different statements

action=answer records the author's reply: { id, status: addressed|rejected, reply }. reply is required — "addressed" with no sentence is not something a reviewer can check.

action=rereview records what a fresh reader found after the revision: { id, verified: true|false, reason, by }. It is a separate column because "I fixed it" and "it is fixed" are different statements, and a letter that merges them is how a revision passes review without the change existing.

rs_review slug=sleep-memory-review action=answer id=f0001 status=addressed
  reply="the search date, the four databases and the access dates are now in the
    first paragraph of Methods"
rs_review slug=sleep-memory-review action=rereview id=f0001
  verified=false by="a fresh reader subagent, given the revised method section only"
  reason="it names the databases and their access dates but still no search date"

Finding f0001 is now addressed (your words, not a verification)
- your answer: the search date, the four databases and the access dates are now in
  the first paragraph of Methods

A re-review decides whether it is actually settled:
`rs_review action=rereview id=f0001 verified=true|false reason="…"`. Until then the
matrix says `not re-reviewed`.

verified=false is a normal, useful answer: the revision did not do what the reply said it did. Record it, with a reason that says what did not settle, and tell the user in plain words what is still outstanding. A re-review of a concern that has no answer yet is refused: there is nothing to verify.

The devil's advocate, and the protocol the tool enforces

action=attack raises one. action=concede records a concession and requires the rebuttal's score — how directly and how evidential the reply was, not how firmly it was made:

• 1 — it repeated itself.
• 3 — it answered a different question.
• 4 — it answered this attack with evidence.
• 5 — that, plus a demonstration.

Below 4 the concession is refused, the advocate holds position and restates the attack; and two concessions in a row are refused outright, because a sequence of concessions that looks like a debate and is actually agreement with extra steps is the sycophancy pattern this protocol exists to slow. Treat a refusal as the protocol working, not as an error to route around: score honestly, record the score you actually mean, and if the reply really was answered, say what evidence answered it. action=attack is where a held position goes when it is worth keeping.

The sycophancy warning is a signal, not a statistic

action=read prints the advocate's attack count, concession count and rate, and when the rate is at or above half over enough exchanges the report says so in a sentence. That warning is the only place the pattern is visible — every individual concession looked reasonable. Raise it with the user when it appears, and re-read the attacks the advocate held: conceding is not the same as being answered.

Writing it up

action=read prints the traceability matrix: one row per concern, with the author's answer and the re-review's finding in separate columns and still open in the answer column where there is none. action=write puts reviews/panel-r<n>.md (the round's concerns, grouped and counted) and reviews/traceability.md on disk — that pair is what a response-to-reviewers letter is written from, and the matrix is pasted into it.

A response letter with an open concern in it concedes the point. Before writing one, check the open count action=read reports and either answer and re-review what is open, or say plainly in the letter what remains unresolved and why. Writing the letter early, while the matrix still says still open, is how a revision looks finished while the concern stands.