Listening for events…
← All Docs

PMA v2 — Elise/Mike Review Loop

Status: Design Depends on: PMA v1 (operational)

The Agents

Agent Role Persona Invocation
Elise Paper Maker Researcher — writes papers /paper 43 (existing PMA skill)
Mike Editor Science paper editor — reviews, critiques /review 43 (new skill)

The Flow

1. Human creates GitHub Issue with paper-idea label
       ↓
2. Human runs: /paper 43
       ↓
3. Elise (PMA) produces workspace + paper + commits
   Posts comment on #43: "Paper complete. Findings: ..."
   Adds label: needs-review
   Assigns: @Misenbek (Mike)
       ↓
4. Human runs: /review 43
       ↓
5. Mike (Editor) reads the paper, analyzes as science editor
   Posts comment on #43 with structured review:
   - Statistical rigor check
   - Methodology concerns
   - Missing references
   - Writing quality
   - Verdict: ACCEPT / REVISE / REJECT
       ↓
6. Human runs: /paper 43 (Elise reviews Mike's comments)
       ↓
7. Elise reads Mike's review comments on the issue
   Addresses each point (fix, rebut, or acknowledge)
   Updates paper.tex, recompiles, commits
   Posts response comment on #43
   Changes label: revised or accepted
       ↓
8. Repeat 4-7 until Mike posts ACCEPT

GitHub Issue as Shared State

All communication flows through the issue comments. Each agent:

  • READS all prior comments before acting
  • POSTS its output as a new comment
  • Uses labels to signal state: needs-review, revised, accepted

This means:

  • Full audit trail of the review process
  • Human can intervene at any point by adding their own comment
  • Each run is serial (human triggers each step)
  • No hidden state — everything is on the issue

Mike's Review Template

When Mike reviews, his comment follows this structure:

## Editor Review — {paper title}

### Statistical Rigor
- [ ] Sample sizes reported for all tests
- [ ] Effect sizes reported (not just p-values)
- [ ] Bonferroni applied where needed
- [ ] Null results reported honestly
- [ ] V1/V2 flags for dramatic results

### Methodology
- [ ] Appropriate test for the hypothesis
- [ ] Data filtering/selection justified
- [ ] Sensitivity analyses included
- [ ] Comparison to prior work

### Writing
- [ ] Abstract states key finding clearly
- [ ] Introduction motivates the question
- [ ] Discussion distinguishes mechanism from correlation
- [ ] Limitations acknowledged
- [ ] References complete and properly formatted

### Specific Issues
1. {issue description + suggested fix}
2. {issue description + suggested fix}
...

### Verdict: ACCEPT / REVISE / REJECT
{Reasoning}

Elise's Revision Template

When Elise revises, her comment responds to each issue:

## Revision Response — Round {N}

### Addressed
1. {Mike's issue} → {what was changed}
2. {Mike's issue} → {what was changed}

### Rebutted
1. {Mike's issue} → {why no change is needed}

### Paper Changes
- paper.tex: {summary of changes}
- Recompiled: paper.pdf updated
- Commit: {hash}

### Status: Ready for re-review

Files to Create

File Purpose
.claude/commands/review.md Mike's /review skill
Update .claude/commands/paper.md Elise's skill — add revision mode

The /review Skill

Invoked as /review 43. Mike:

  1. Reads issue #43 title + body + ALL comments
  2. Reads the paper at workspaces/{slug}/paper/paper.tex
  3. Reads data/results.json for statistical validation
  4. Analyzes as a science paper editor
  5. Posts structured review comment on the issue
  6. Adds label needs-review

The /paper Revision Mode

When Elise is invoked on an issue that already has a workspace + review comments:

  1. Detects existing workspace (doesn't recreate)
  2. Reads ALL issue comments (especially Mike's review)
  3. Addresses each review point
  4. Updates paper.tex + index.md
  5. Recompiles PDF
  6. Commits with message: "PMA #43: Revision round N — addressed review"
  7. Posts revision response comment
  8. Changes label to revised

Guards

  • Serial execution only (human triggers each step)
  • Mike NEVER modifies files — read-only analysis + comments
  • Elise NEVER ignores review points — must address or rebut each one
  • Max 3 revision rounds — if not accepted by round 3, escalate to human
  • All communication on the GitHub Issue — no side channels

Future (v3+)

  • Automated dispatch: Elise completes → Mike auto-reviews → Elise auto-revises
  • Quality scoring: track accept/revise/reject rates across papers
  • Multiple reviewers: Mike + external reviewer agent
  • Cross-referencing: Mike checks findings against existing TerraPulse papers for consistency
Live Feed