gatekeeper

name	gatekeeper
description	Audit the current draft for craft and integrity violations. Runs `integrity-gatekeeper` and `exposition-gatekeeper` agents in parallel, synthesizes findings into Critical / Important / Minor. **Reports only — never auto-fixes.** **Does not check math.** Use after any substantive draft pass.

gatekeeper

Philosophy

Gatekeepers fail closed. They emit refusal classes, not suggestions. The user decides whether to act.

This pipeline does not check math — proofs come from a separate pipeline and are immutable input. The gatekeepers check prose integrity (citations, claim faithfulness against brief.md), structure, and craft.

Invocation

User says: "run the gatekeeper", "audit the draft". Auto-invoked after hydrate skills and motivation-pass.

What this skill does

1. Spawn integrity-gatekeeper and exposition-gatekeeper agents in parallel (one Agent call with two tool uses).

   • integrity-gatekeeper audits citation faithfulness, claim faithfulness (prose vs brief.md/proofs/), notation consistency, cross-ref integrity.
   • exposition-gatekeeper audits structure, prose, signposting, theory-paper conventions against STRUCTURE.md.

2. Synthesize findings into a single report with three bands:

   • Critical — anything in the integrity refusal classes:

     • CLAIM-UNSUPPORTED: a prose claim doesn't trace to brief.md or proofs/.
     • FABRICATED-REFERENCE: a \cite{} key not in refs.bib.
     • DANGLING-REF: a \cref{...} to a non-existent label.
     • NOTATION-MISMATCH: prose notation doesn't match model.tex / proofs/.

   • Important — craft / structural issues from the exposition-gatekeeper:

     • BOILERPLATE-INTRO, INTRO-OVERCITED, NO-CONTRIBUTION-SENTENCE, ROADMAP-OVERLOADED
     • RELATED-CHRONOLOGICAL, NO-DIFFERENTIATION
     • ORPHANED-NOTATION, NO-INTUITION
     • Section-balance deviations from STRUCTURE.md norms.

   • Minor — copy-edit issues:

     • STYLE-WORD, STYLE-HEDGE, PARAGRAPH-LENGTH
     • ABSTRACT-BOILERPLATE, ABSTRACT-LENGTH, ABSTRACT-CITES

3. Do not edit any files. Output is the report.

Output format

## Gatekeeper report — <target> — <timestamp>

### Critical (N)
- [CLAIM-UNSUPPORTED] intro.tex:42 — "robust to any prior" not in brief.md or proofs/.
- [FABRICATED-REFERENCE] related.tex:13 — \cite{bergemann2024robust} not in refs.bib.

### Important (N)
- [BOILERPLATE-INTRO] intro.tex:3 — opens with "In this paper, we…"
- [RELATED-CHRONOLOGICAL] related.tex paragraph 2

### Minor (N)
- [STYLE-WORD] intro.tex:18 — "novel"
- ...

### What I did not check
- Mathematical correctness (out of scope for this repo).
- Whether cited results say what's attributed to them (only that citations exist).

When to run it

• After every substantive writing-role pass.
• Before every latex-compile you think of as "final."
• Whenever the user asks.

What it does NOT do

• Does not auto-fix.
• Does not check math.
• Does not evaluate whether the paper is "good" — that's the user's call.