---
name: peer-review-panel
description: Use when a piece of writing or an experimental design needs to survive expert scrutiny before it goes out -- a conference abstract, manuscript, preregistration, grant, or study design -- and Mason says "peer review this", "run a panel on this", "would reviewers buy this", "referee this", "red-team my abstract", or invokes /peer-review-panel. Also fires before submission to a venue with real referees.
---

_Last updated: 2026-08-05_

# peer-review-panel

Seat a panel of simulated expert referees on a document, get independent structured
referee reports, let them argue once, then synthesize. The panel is internal red-team
material: it finds what a real reviewer would find, before a real reviewer does.

## The one thing this skill exists to prevent

An unscaffolded agent asked to "run a panel" writes all the reviewer voices itself in one
pass. It reads well. It quotes the document accurately. And it passes everything -- the
measured baseline returned **accept from 4 of 4 reviewers, mean 6.0/10, with zero
statements of what would change any reviewer's mind** (`references/red-baseline-findings.md`).

One mind wearing four name tags will not fail its user. Everything below exists to stop
that.

**Peers are dispatched, never voiced.** You do not write a referee report. You do not
write "what Reviewer 3 would say." You dispatch a seat and you parse what comes back.

### Red flags -- stop and dispatch

- Writing a reviewer's report in your own context
- "I'll draft the reports and dispatch only if they seem thin"
- "The bridge is slow, I'll simulate that seat"
- "I know what the statistician would say"
- Filling in a peer that returned nothing or broke format
- Averaging scores to reach a verdict

| Rationalization | Reality |
|---|---|
| "I can write four distinct voices" | The baseline did, convincingly, and passed a document with real defects. Distinct prose is not independent judgment. |
| "Dispatching is slow and expensive" | A panel you authored is worth nothing. Cost is the point of the exercise. |
| "One seat failed, I'll cover it" | A fabricated report launders your own opinion as independent corroboration. Report the seat failed. |
| "They'd all agree anyway" | Then the panel reports convergence as a finding. That is information; a manufactured spread is not. |
| "The doc is short, a full panel is overkill" | Use `quick screen` depth. Still dispatched. |

## Precedence of the reference files

Load `references/review-rubric.md` at runtime -- it is the **canonical instrument**. The
four dossiers are research explaining *why*; consult them for archetype voice, venue
detail, and mechanics, but where they disagree with the rubric, **the rubric wins**.
Specifically: `panel-mechanics.md` proposes criterion seats and a soundness/excitement
split; the rubric overrules both in favor of archetype seats on one anchored 1-5 scale.

| File | Use it for |
|---|---|
| `review-rubric.md` | **Canonical.** Report schema, 1-5 anchors, per-archetype checklists, per-venue blocks, validation gate, synthesis contract |
| `reviewer-archetypes.md` | Persona voice, characteristic objections, dead giveaways, seating guidance |
| `venue-norms.md` | Venue criteria, desk-reject triggers, field-specific demands |
| `panel-mechanics.md` | Bridge invocations, model parameters, independence protocol, failure recovery |
| `anti-patterns.md` | Detection tests and countermeasures behind the validation gate |
| `red-baseline-findings.md` | What failure looks like |

## Step 1 -- Intake

Call **AskUserQuestion** with three questions. Do not guess these.

1. **How many peers?** 2-6. (Default 4.)
2. **Field / venue?** consciousness research / TSC; parapsychology journal; statistics-heavy
   prereg or Registered Report; general psychology; other. This selects the venue block
   from the rubric and drives which archetypes are seated.
3. **Depth?** `quick screen` (phase 1 only, major issues, no rebuttal) or
   `full referee reports` (all three phases).

Then locate the document. If Mason named a file, read it. If a file holds many versions,
review **only** the entry specified -- confirm which before dispatching.

## Step 2 -- Seat the panel

Assign archetypes from `review-rubric.md` §3 per the venue's seating guidance.

**Composition rules, all mandatory:**

- **No two peers share both model and prompt.** This is the whole design.
- **At least one peer runs on GPT** via the gpt_brad bridge (`panel-mechanics.md` §1.1).
  If it is down, fall back to the cursor bridge and say so in the report. Bedrock was
  unreachable as of 2026-08-05 (expired sandbox SSO).
- **Panels of 4+ seat at least one skeptic archetype and one statistician archetype.**
- Spread the Claude seats across `opus` / `sonnet` / `haiku` / `fable` via the Agent tool's
  `model` parameter. `model` is ignored for fork subagents -- never seat a panel with forks.
  **Check Fable before seating it** -- it was credit-blocked as of 2026-08-05 and the seat
  fails outright. Re-seat on an unused model rather than dropping to three peers.
- Two GPT-family peers are **not** independent of each other. Observed on this machine:
  gpt_brad and cursor/gpt-5.2 returned near-identical answers to the same prompt
  (`panel-mechanics.md` §1.4). Give them different archetypes, never the same one.

Each persona is a seasoned expert **styled after the documented reviewing behavior** of
real figures in the archetype dossiers. Label every persona a simulation in the output.

## Step 3 -- Phase 1: independent reports

Extract a numbered **claim list** from the document first. Every objection must name a
claim ID -- this is what stops reviewers from critiquing the topic instead of the document.

Dispatch **all peers in a single message** so none can see another's output. Each peer
gets: the document, its own archetype checklist, the venue block, the claim list, and the
report schema from `review-rubric.md` §1.2. **Nothing else** -- no peer count, no other
archetypes, no prior findings.

Each report returns: summary of what the document claims, major issues, minor issues,
line-anchored objection blocks, an anchored 1-5 score with confidence, a verdict consistent
with that score, and **VERDICT MOVER** -- what would change their mind. All required.

Then run the **validation gate** (`review-rubric.md` §6) on every returned report before it
enters the panel.

## Step 4 -- Phase 2: moderated rebuttal (one round, skip on quick screen)

Send each peer **only the other peers' major issues** -- never their full reports, scores,
or verdicts. Each may sharpen, concede, or hold, with a reason. Peers may revise their
score; a revised score still obeys the score-to-verdict mapping.

**Rebuild the digest from a live validation pass immediately before dispatching it.** A
peer returned for re-anchoring may have complied since the last check; a stale drop list
silently excludes repaired objections from cross-examination. This happened in the
2026-08-05 shakedown and cost two objections their rebuttal round.

**Re-dispatch rather than SendMessage for phase 2.** Observed 2026-08-05: an `opus` seat
answered a SendMessage follow-up, a `haiku` seat no-opped on two of them. A fresh agent
handed its own filed report reconstitutes the position reliably.

## Step 5 -- Phase 3: meta-review

A fresh `opus` agent that has seen no peer deliberation synthesizes per the contract in
`review-rubric.md` §7: consensus verdict and its basis, ranked must-fix list, disagreements
left standing and why, and a **panel-health block** reporting convergence, degraded or
failed seats, and dropped objections.

The governing score is the **minimum**, not the mean. If every peer accepts, or all scores
land within 0.5, report that as a convergence finding rather than confidence.

## Step 6 -- Output

Write `<document-name>-panel-review.md` **next to the reviewed document**, with
`_Last updated: YYYY-MM-DD_` under the title. Then give Mason a terminal summary: consensus
verdict, the ranked must-fix list, standing disagreements, and panel health.

## Guardrails -- reproduce these in the report file

- **Internal red-team material only.** Never quote panel output publicly, or as the opinion
  of the real people the personas are styled after. These are simulations built from
  published reviewing behavior, not those people.
- **Every objection cites the line it targets.** An unanchored objection is not a weak
  objection -- it is not an objection, and it is dropped.
- **Closed citations.** Reviewers may name only works appearing in the reviewed document or
  in these reference files. A missing-literature objection may describe the *kind* of work
  and the gap it fills, naming no paper, author, or year. Citations outside the allowed set
  are dropped and the seat is flagged -- never looked up, never repaired.

## Common mistakes

| Mistake | Fix |
|---|---|
| Reviewing the field instead of the document | Claim-ID gate; objections without a valid claim ID are dropped |
| Averaging to a verdict | Governing score is the minimum |
| Peers seeing each other in phase 1 | Fan out in one message; rebuttal sees major issues only |
| Repairing a peer's bad citation | Drop the objection, flag the seat |
| Reviewing the wrong version | Confirm which entry before dispatch |
| Treating agreement as confidence | Agreement between same-family models is weak evidence |
