Seven-Pass Adversarial Review
Runs seven independent reviewers, each focused on a single lens, then synthesizes their findings into one prioritized revision plan — the fan-out → reduce → judge runtime from orchestrator-protocol.md, applied with seven lenses.
Why seven passes? A single-agent review blends lenses and softens each one. Seven forked agents each approach the paper with full context budget for their own lens, then a synthesizer resolves conflicts and de-duplicates.
When to pick this over /review-paper: This skill costs roughly 7× more tokens than /review-paper (default) and ~2× more than /review-paper --adversarial. Use it when the paper is submission-ready or at R&R stage and you need maximum lens coverage. For early drafts or iterative work, /review-paper is the right tool. For journal-simulation pressure test, use /review-paper --peer <journal> instead.
Inputs
$0 — manuscript path (.tex, .qmd, .md, or .pdf). Required.
The Seven Lenses
Each lens runs as a forked subagent (context: fork) so the main conversation stays clean.
| # |
Lens |
Focus |
Agent type |
| 1 |
Abstract audit |
Does the abstract state the question, method, result, and contribution? Does it match the paper? |
general-purpose |
| 2 |
Intro structure |
Does the intro follow Cochrane / Varian framework? Literature placement? Contribution clarity? |
general-purpose |
| 3 |
Methods / identification |
Are assumptions stated? Is identification credible? Are alternatives addressed? |
domain-reviewer |
| 4 |
Results + tables |
Do tables read standalone? Is magnitude + significance discussed? Units consistent? |
general-purpose |
| 5 |
Robustness |
Are obvious threats pre-empted? Is the robustness section convincing or theatrical? |
general-purpose |
| 6 |
Prose quality |
Sentence-level clarity, hedging, passive voice, paragraph cohesion |
proofreader |
| 7 |
Citation audit |
Invokes /validate-bib --semantic; checks cite-claim direction for top-10 works |
general-purpose |
Workflow
Phase 0: Pre-flight
- Resolve manuscript path.
- Decide if
.pdf → extract text first (pdftotext -layout).
- Create output dir:
quality_reports/seven_pass_[stem]/.
Phase 1: Spawn 7 reviewers in parallel
In a single message, spawn 7 Agent tool calls (one per lens). Each subagent gets:
- The manuscript path (to re-read with its own context).
- The lens-specific prompt (below).
- Instructions to write to
quality_reports/seven_pass_[stem]/lens_[N]_[lens-name].md.
- A JSON findings array conforming to
finding-schema.json, written to quality_reports/seven_pass_[stem]/lens_[N]_[lens-name].json beside the prose report. Severities: blocker | major | minor | nit. Every finding computes its id with python3 scripts/validate-findings.py --id FILE LINE LOCUS and carries rule, evidence, and a failing_case. Phase 2 validates each array first (python3 scripts/validate-findings.py <file> — exit 0 required; a lens whose report does not validate has not reviewed), then reduces over the typed findings — it does not re-read the prose. Because ids are lens-independent, the same defect found by two lenses dedups to one finding automatically.
This is the fan-out primitive from orchestrator-protocol.md; Agent subagents are the portable mechanism (the agents that fill lenses 3/6 are in agent-fleet.md).
Lens prompt rubrics are embedded inline below — one summary paragraph per lens. Each forked subagent receives its lens's rubric plus the manuscript path.
Lens prompt summaries:
- Lens 1 (Abstract): Does the first sentence state the question? Does it name the method? Quantify the headline result? State one-sentence contribution? Cross-check: do these four things match the body?
- Lens 2 (Intro): Does the intro open with the question? Hook → context → contribution → roadmap? Lit review placed correctly (after the hook, not before)? Contribution-counted (1, 2, 3…)? Preview of findings with magnitudes?
- Lens 3 (Methods): Is every assumption stated? Are they strong or weak? Is identification one-liner clear? Are known violations (selection, measurement, reverse causality, SUTVA) addressed? Are instruments / RDD / DiD assumptions explicit and defensible?
- Lens 4 (Results): Does each table read standalone (caption, units, SEs clarified)? Is magnitude interpreted (not just significance)? Are units consistent across tables? Are figures legible at 8pt?
- Lens 5 (Robustness): Does the paper ANTICIPATE a sharp referee's objections? Are robustness checks motivated, or just listed? Power/placebo tests present? Heterogeneity explored where promised?
- Lens 6 (Prose): Sentences under 30 words? Active voice dominant? Hedging proportionate (neither overclaiming nor endless "may suggest")? Paragraph topic sentences?
- Lens 7 (Citations): Invoke
/validate-bib --semantic. For top-10 cited works, does the in-text claim match the cited paper's actual finding direction? Are contemporary / competing works cited?
Phase 2: Synthesize (reduce → judge, with the hallucination gate)
Wait for all 7 lens reports. Reduce, don't re-review: stack the seven scorecards and apply the gate predicate from orchestration-schemas.md §3 — the Executive verdict is a function of the typed findings, not a fresh eighth opinion. Then run the post-judge hallucination gate (§4): any CRITICAL the synthesis introduces that no lens raised must be re-verified in a fresh claim-verifier fork, or dropped to [JUDGE-HALLUCINATED] and the verdict recomputed. A synthesis may freely downgrade or de-duplicate lens findings; it may not invent a new blocker.
Then produce:
quality_reports/seven_pass_[stem]/_SYNTHESIS.md
# Seven-Pass Review: [Manuscript]
**Date:** YYYY-MM-DD
**Path:** [manuscript]
## Executive verdict
**Overall state:** [SUBMIT / REVISE-MINOR / REVISE-MAJOR / REJECT-AND-RESTART]
## Cross-lens CRITICAL issues
| # | Lens(es) | Issue | Recommendation |
|---|---|---|---|
## MAJOR issues (second-round)
| # | Lens(es) | Issue |
|---|---|---|
## MINOR polish
[bulleted]
## Per-lens scorecard
| Lens | Critical | Major | Minor | Score/10 |
|---|---|---|---|---|
| 1. Abstract | | | | |
| 2. Intro | | | | |
| 3. Methods | | | | |
| 4. Results | | | | |
| 5. Robustness | | | | |
| 6. Prose | | | | |
| 7. Citations | | | | |
| **Overall** | | | | |
## Revision plan (in recommended order)
1. [Highest-leverage fix — usually a lens with 2+ CRITICALs]
2. …
7. [Lowest-leverage polish]
## Contradictions between lenses
[If two lenses disagree, surface here. E.g., Lens 2 says "expand contribution" but Lens 6 says "trim intro".]
Phase 3: Token-budget report
After synthesis, print:
Seven-pass review complete.
Subagents: 7 (parallel) + 1 synthesizer.
Approx token usage: ~80–120k (vs ~15k for single-pass /review-paper).
Runtime: ~3–5 min wall-clock.
For cheaper alternatives:
- Single-pass: /review-paper
- Iterative: /review-paper --adversarial
When to use this skill
- Before first submission to a top journal.
- After a major revision when you want to catch drift.
- R&R when referees disagree — surfaces contradictions your revision must navigate.
When NOT to use
- Early drafts (use
/review-paper single-pass first).
- Short notes, comments, or replies (overkill).
- When you've already run this in the last 7 days and nothing substantive changed.
Findings are validated, not just written (v2.5)
This skill's reviewers emit findings under the machine-checked contract in
finding-schema.json. Reports are JSON arrays.
Smoke-test the harness before spending review effort — a run that fans out reviewers and
then cannot write a valid report has wasted the whole pass:
echo '[]' | python3 scripts/validate-findings.py
Then, before presenting any summary:
python3 scripts/validate-findings.py <report>.json # exit 0 required
What the contract forces, and why:
rule — the documented rule or standard violated. A finding citing no rule is an
opinion, and opinions do not gate a commit.
failing_case — a concrete configuration under which the claim breaks, or the exact
missing hypothesis. "This could be clearer" does not validate.
id = sha1("<file>:<line>:<locus>") — deterministic, so dedup across rounds is
exact and the two-strikes rule is checkable rather than eyeballed.
mechanical — true only for fixes that cannot change a result (typo, cross-reference,
formatting, label). Never for an estimand, assumption, specification, inference
procedure, sample definition, or reporting language: those return to the researcher.
Apply the per-lens evidence burdens and the "does NOT count" filters in
orchestration-schemas.md §7 before
verification, so known false alarms never reach the judge. The verifier pass is
refute-biased: only verdict: "confirmed" findings ship; anything it cannot ground is
dropped, not downgraded to a warning.
Cross-references
.claude/skills/review-paper/SKILL.md — the single-pass and --adversarial modes (cheaper, faster).
.claude/skills/validate-bib/SKILL.md — invoked by Lens 7.
.claude/skills/audit-reproducibility/SKILL.md — complementary; numeric-claims side of the audit.
- Workflow guide, Pattern 15 — the narrative explanation of why seven lenses.
Exit behavior
- Exits 0 always (review is informational). The synthesis report's "Executive verdict" is the gate.
- Any
CRITICAL at the top of the synthesis should block submission until resolved.
What this skill does NOT do
- Re-run seven lenses if the manuscript hasn't changed — check git diff against last run date in
_SYNTHESIS.md, skip unchanged lenses if requested via --incremental (future).
- Auto-apply fixes — that's
/review-paper --adversarial's job.
- Replace human judgment. A reviewer who knows your subfield still beats seven LLMs.
1---2name: seven-pass-review3description: Mechanize Pattern 15 — the seven-pass adversarial review protocol for academic manuscripts. Spawns 7 forked subagents in parallel (abstract, intro, methods, results, robustness, prose, citations), then synthesizes a prioritized revision checklist. Use for submission-ready or R&R-stage papers where single-pass review isn't enough.4argument-hint: "[manuscript path]"5allowed-tools: ["Read", "Grep", "Glob", "Write", "Bash", "Agent", "Task"]6effort: high7metadata:8 source: https://skillmd.com/skills/pedrohcgs/seven-pass-review9---1011# Seven-Pass Adversarial Review1213Runs seven independent reviewers, each focused on a single lens, then synthesizes their findings into one prioritized revision plan — the fan-out → reduce → judge runtime from `orchestrator-protocol.md`, applied with seven lenses.1415**Why seven passes?** A single-agent review blends lenses and softens each one. Seven forked agents each approach the paper with full context budget for their own lens, then a synthesizer resolves conflicts and de-duplicates.1617> **When to pick this over `/review-paper`:** This skill costs roughly 7× more tokens than `/review-paper` (default) and ~2× more than `/review-paper --adversarial`. Use it when the paper is submission-ready or at R&R stage and you need maximum lens coverage. For early drafts or iterative work, `/review-paper` is the right tool. For journal-simulation pressure test, use `/review-paper --peer <journal>` instead.1819## Inputs2021- `$0` — manuscript path (`.tex`, `.qmd`, `.md`, or `.pdf`). Required.2223## The Seven Lenses2425Each lens runs as a **forked subagent** (context: fork) so the main conversation stays clean.2627| # | Lens | Focus | Agent type |28|---|---|---|---|29| 1 | Abstract audit | Does the abstract state the question, method, result, and contribution? Does it match the paper? | general-purpose |30| 2 | Intro structure | Does the intro follow Cochrane / Varian framework? Literature placement? Contribution clarity? | general-purpose |31| 3 | Methods / identification | Are assumptions stated? Is identification credible? Are alternatives addressed? | domain-reviewer |32| 4 | Results + tables | Do tables read standalone? Is magnitude + significance discussed? Units consistent? | general-purpose |33| 5 | Robustness | Are obvious threats pre-empted? Is the robustness section convincing or theatrical? | general-purpose |34| 6 | Prose quality | Sentence-level clarity, hedging, passive voice, paragraph cohesion | proofreader |35| 7 | Citation audit | Invokes `/validate-bib --semantic`; checks cite-claim direction for top-10 works | general-purpose |3637## Workflow3839### Phase 0: Pre-flight40411. Resolve manuscript path.422. Decide if `.pdf` → extract text first (`pdftotext -layout`).433. Create output dir: `quality_reports/seven_pass_[stem]/`.4445### Phase 1: Spawn 7 reviewers in parallel4647In a single message, spawn 7 `Agent` tool calls (one per lens). Each subagent gets:4849- The manuscript path (to re-read with its own context).50- The lens-specific prompt (below).51- Instructions to write to `quality_reports/seven_pass_[stem]/lens_[N]_[lens-name].md`.52- A **JSON findings array** conforming to [`finding-schema.json`](../../references/finding-schema.json), written to `quality_reports/seven_pass_[stem]/lens_[N]_[lens-name].json` beside the prose report. Severities: `blocker | major | minor | nit`. Every finding computes its `id` with `python3 scripts/validate-findings.py --id FILE LINE LOCUS` and carries `rule`, `evidence`, and a `failing_case`. Phase 2 **validates each array first** (`python3 scripts/validate-findings.py <file>` — exit 0 required; a lens whose report does not validate has not reviewed), then reduces over the typed findings — it does not re-read the prose. Because ids are lens-independent, the same defect found by two lenses dedups to one finding automatically.5354This is the **fan-out** primitive from [`orchestrator-protocol.md`](../../rules/orchestrator-protocol.md); `Agent` subagents are the portable mechanism (the agents that fill lenses 3/6 are in [`agent-fleet.md`](../../references/agent-fleet.md)).5556Lens prompt rubrics are embedded inline below — one summary paragraph per lens. Each forked subagent receives its lens's rubric plus the manuscript path.5758**Lens prompt summaries:**5960- **Lens 1 (Abstract):** Does the first sentence state the question? Does it name the method? Quantify the headline result? State one-sentence contribution? Cross-check: do these four things match the body?61- **Lens 2 (Intro):** Does the intro open with the question? Hook → context → contribution → roadmap? Lit review placed correctly (after the hook, not before)? Contribution-counted (1, 2, 3…)? Preview of findings with magnitudes?62- **Lens 3 (Methods):** Is every assumption stated? Are they strong or weak? Is identification one-liner clear? Are known violations (selection, measurement, reverse causality, SUTVA) addressed? Are instruments / RDD / DiD assumptions explicit and defensible?63- **Lens 4 (Results):** Does each table read standalone (caption, units, SEs clarified)? Is magnitude interpreted (not just significance)? Are units consistent across tables? Are figures legible at 8pt?64- **Lens 5 (Robustness):** Does the paper ANTICIPATE a sharp referee's objections? Are robustness checks motivated, or just listed? Power/placebo tests present? Heterogeneity explored where promised?65- **Lens 6 (Prose):** Sentences under 30 words? Active voice dominant? Hedging proportionate (neither overclaiming nor endless "may suggest")? Paragraph topic sentences?66- **Lens 7 (Citations):** Invoke `/validate-bib --semantic`. For top-10 cited works, does the in-text claim match the cited paper's actual finding direction? Are contemporary / competing works cited?6768### Phase 2: Synthesize (reduce → judge, with the hallucination gate)6970Wait for all 7 lens reports. **Reduce, don't re-review:** stack the seven `scorecard`s and apply the gate predicate from [`orchestration-schemas.md` §3](../../references/orchestration-schemas.md) — the Executive verdict is a function of the typed findings, not a fresh eighth opinion. Then **run the post-judge hallucination gate** ([§4](../../references/orchestration-schemas.md)): any CRITICAL the synthesis introduces that **no lens raised** must be re-verified in a fresh `claim-verifier` fork, or dropped to `[JUDGE-HALLUCINATED]` and the verdict recomputed. A synthesis may freely downgrade or de-duplicate lens findings; it may not invent a new blocker.7172Then produce:7374`quality_reports/seven_pass_[stem]/_SYNTHESIS.md`7576```markdown77# Seven-Pass Review: [Manuscript]7879**Date:** YYYY-MM-DD80**Path:** [manuscript]8182## Executive verdict8384**Overall state:** [SUBMIT / REVISE-MINOR / REVISE-MAJOR / REJECT-AND-RESTART]8586## Cross-lens CRITICAL issues87| # | Lens(es) | Issue | Recommendation |88|---|---|---|---|8990## MAJOR issues (second-round)91| # | Lens(es) | Issue |92|---|---|---|9394## MINOR polish95[bulleted]9697## Per-lens scorecard98| Lens | Critical | Major | Minor | Score/10 |99|---|---|---|---|---|100| 1. Abstract | | | | |101| 2. Intro | | | | |102| 3. Methods | | | | |103| 4. Results | | | | |104| 5. Robustness | | | | |105| 6. Prose | | | | |106| 7. Citations | | | | |107| **Overall** | | | | |108109## Revision plan (in recommended order)1101. [Highest-leverage fix — usually a lens with 2+ CRITICALs]1112. …1127. [Lowest-leverage polish]113114## Contradictions between lenses115[If two lenses disagree, surface here. E.g., Lens 2 says "expand contribution" but Lens 6 says "trim intro".]116```117118### Phase 3: Token-budget report119120After synthesis, print:121122```123Seven-pass review complete.124Subagents: 7 (parallel) + 1 synthesizer.125Approx token usage: ~80–120k (vs ~15k for single-pass /review-paper).126Runtime: ~3–5 min wall-clock.127For cheaper alternatives:128 - Single-pass: /review-paper129 - Iterative: /review-paper --adversarial130```131132## When to use this skill133134- **Before first submission** to a top journal.135- **After a major revision** when you want to catch drift.136- **R&R when referees disagree** — surfaces contradictions your revision must navigate.137138## When NOT to use139140- Early drafts (use `/review-paper` single-pass first).141- Short notes, comments, or replies (overkill).142- When you've already run this in the last 7 days and nothing substantive changed.143144145## Findings are validated, not just written (v2.5)146147This skill's reviewers emit findings under the machine-checked contract in148[`finding-schema.json`](../../references/finding-schema.json). Reports are JSON **arrays**.149150**Smoke-test the harness before spending review effort** — a run that fans out reviewers and151then cannot write a valid report has wasted the whole pass:152153```bash154echo '[]' | python3 scripts/validate-findings.py155```156157Then, before presenting any summary:158159```bash160python3 scripts/validate-findings.py <report>.json # exit 0 required161```162163What the contract forces, and why:164165- **`rule`** — the documented rule or standard violated. A finding citing no rule is an166 opinion, and opinions do not gate a commit.167- **`failing_case`** — a concrete configuration under which the claim breaks, or the exact168 missing hypothesis. *"This could be clearer"* does not validate.169- **`id = sha1("<file>:<line>:<locus>")`** — deterministic, so dedup across rounds is170 exact and the two-strikes rule is checkable rather than eyeballed.171- **`mechanical`** — `true` only for fixes that cannot change a result (typo, cross-reference,172 formatting, label). **Never** for an estimand, assumption, specification, inference173 procedure, sample definition, or reporting language: those return to the researcher.174175Apply the **per-lens evidence burdens** and the **"does NOT count" filters** in176[`orchestration-schemas.md` §7](../../references/orchestration-schemas.md) *before*177verification, so known false alarms never reach the judge. The verifier pass is178**refute-biased**: only `verdict: "confirmed"` findings ship; anything it cannot ground is179dropped, not downgraded to a warning.180181## Cross-references182183- `.claude/skills/review-paper/SKILL.md` — the single-pass and `--adversarial` modes (cheaper, faster).184- `.claude/skills/validate-bib/SKILL.md` — invoked by Lens 7.185- `.claude/skills/audit-reproducibility/SKILL.md` — complementary; numeric-claims side of the audit.186- Workflow guide, Pattern 15 — the narrative explanation of why seven lenses.187188## Exit behavior189190- Exits 0 always (review is informational). The synthesis report's "Executive verdict" is the gate.191- Any `CRITICAL` at the top of the synthesis should block submission until resolved.192193## What this skill does NOT do194195- Re-run seven lenses if the manuscript hasn't changed — check git diff against last run date in `_SYNTHESIS.md`, skip unchanged lenses if requested via `--incremental` (future).196- Auto-apply fixes — that's `/review-paper --adversarial`'s job.197- Replace human judgment. A reviewer who knows your subfield still beats seven LLMs.