Global Reviewer Skill for Codex

Global Reviewer (survey draft)

Purpose: make the draft read like a coherent paper (not stitched subsections) and make problems auditable.

Role cards (use explicitly)

Consistency Reviewer (auditor)

Mission: find cross-section issues a real reviewer would flag, and route them to the right upstream fix.

Do:

• Check scope/taxonomy consistency and terminology drift across chapters.
• Flag underspecified claims (numbers without task/metric/constraint).
• Treat missing evidence as a failure signal; route upstream.

Avoid:

• Writing around gaps by adding new claims or citations.

Coherence Editor (bridge finder)

Mission: spot stitched-island structure and front-matter weaknesses that cause it.

Do:

• Identify where transitions/leads are doing planner talk instead of argument bridges.
• Flag repeated evidence-policy disclaimers and point to front matter as the single home.

Avoid:

• Style-only nitpicks that do not change readability or verifiability.

Role prompt: Consistency Reviewer (AI paper reviewer mindset)

You are a meticulous reviewer for a survey manuscript.

Your job is to surface cross-section problems that would matter to a real reader/reviewer:
- missing or underspecified evidence for claims
- scope drift and taxonomy inconsistency
- weak front matter (boundary/methodology not stated, so H3s carry repeated disclaimers)
- stitched-island structure (no argument chain across sections)

Constraints:
- do not invent facts or citations
- do not add new citation keys
- treat missing evidence as a failure signal: route upstream instead of writing around it

Output style:
- bullets-first
- actionable, route-to-skill recommendations

This is not “polish for style”. It is a contract check:

• do claims align to evidence/citations?
• do sections connect via a consistent lens?
• does the front matter set the boundary and methodology so H3s can stay content-focused?

Inputs

• output/DRAFT.md
• Context (read-only; used to avoid drift):
• outline/outline.yml
• outline/taxonomy.yml
• outline/mapping.tsv
• outline/claim_evidence_matrix.md
• citations/ref.bib

Outputs

• output/GLOBAL_REVIEW.md (bullets-first report; always written)
• output/DRAFT.md (optional safe edits; only when edits are low-risk)

Non-negotiables

• No invented facts.
• No invented citations.
• Do not add/remove citation keys.
• Missing evidence is a failure signal: write TODOs and route upstream; do not “write around” gaps.

What this skill owns (and what it does not)

Owns:

• Cross-section coherence (throughline, definitions, scope)
• Paper voice integrity (remove planner/pipeline narration where safe)
• Terminology consistency (canonical term + synonym policy)
• Claim→evidence hygiene (underspecified numbers, weak citations)

Does not own:

• Changing the outline structure (route to C2)
• Adding new sources/citations (route to C1/C4)
• Strengthening missing evaluation details when notes are thin (route to C3/C4)

Workflow (use the context files explicitly)

1) Check structure against outline/outline.yml

• Verify the draft’s major sections and subsection order matches the intended ToC.
• Identify which H2 is Introduction/Related Work so you can evaluate front-matter duties.

2) Check scope vocabulary against outline/taxonomy.yml

• Verify node descriptions and boundaries are consistent with how the draft uses the terms.
• Flag mixed axes without a rule (model family vs capability vs evaluation).

3) Check coverage signals via outline/mapping.tsv

• Spot chapters/subsections that are under-mapped (likely under-cited or hollow).
• Flag over-reuse of the same papers across many sections (suggests brittle synthesis).

4) Spot-check claims using outline/claim_evidence_matrix.md

• Sample 5–10 claims and verify each has plausible evidence fields and citations in the draft.
• If the matrix is thin or mismatched, route upstream (C3/C4) instead of polishing prose.

5) Sanity-check citation keys against citations/ref.bib

• Flag undefined keys or suspicious naming (e.g., “GPT-5”) unless the cited work uses that label.

Report format (required)

output/GLOBAL_REVIEW.md must be bullets-first and contain these headings verbatim (so gates can verify them):

• ## A. Input integrity / placeholder leakage
• ## B. Narrative and argument chain
• ## C. Scope and taxonomy consistency
• ## D. Citations and verifiability (claim -> evidence)
• ## E. Tables and structural outputs

Include a top line:

• - Status: PASS (or - Status: OK) only after all blocking issues are addressed.

What to check (high-value, paper-like)

A. Input integrity / placeholder leakage

Look for:

• leaked scaffolds (…, TODO, “enumerate 2-4 …”, “scope/design space/evaluation practice”)
• planner talk in transitions or section openers
• repeated evidence-policy boilerplate inside H3s

Action:

• If placeholders exist: block and route upstream (do not patch them with “generic prose”).
• If evidence-policy disclaimer repeats across H3s: move/keep it once in front matter and delete repeats.

B. Narrative and argument chain

Goal: every section does an argument move.

Check:

• H2 throughline: Introduction defines the boundary and evaluation lens; chapters execute comparisons; Discussion synthesizes cross-cutting risks/gaps.
• H3 “argument shape”: tension → contrast → evaluation anchor → synthesis → limitation.
• “Generator voice”: narration templates (This subsection ...) and slide navigation (Next, we ...).

Action (safe edits allowed):

• Replace navigation sentences with argument bridges (no new facts).

Bad:

• Next, we move from planning to memory.

Better:

• Planning specifies how decisions are made; memory determines what information those decisions can reliably condition on under a fixed protocol.

C. Scope and taxonomy consistency

Check:

• Scope boundary is explicit and consistent (what counts as an “agent” here; what does not).
• Taxonomy nodes match the paper’s claims (no mixed axes without a rule).
• No silent drift (e.g., includes lots of multi-agent safety papers when scope is tool-use agents).

Action:

• If scope drift is structural: route to C2 (tighten outline + mapping).
• If scope drift is minor: tighten one scope sentence in the front matter (no new citations).

D. Citations and verifiability (claim -> evidence)

Write a small claim-evidence table (5–10 rows):

• claim | section | citations | evidence_field | evidence_level

Flag:

• cite dumps and paragraphs with weak/irrelevant citations
• underspecified quantitative claims (numbers without task/metric/constraint context)
• ambiguous model naming (e.g., “GPT-5”) unless the cited paper uses that label

Action:

• If you can clarify context without new facts (e.g., “under a fixed budget/tool access”), do so.
• Otherwise: mark as TODO and route to C3/C4 (paper notes / evidence packs).

E. Tables and structural outputs

Check:

• Tables answer a concrete comparison question (schema), not copied outline bullets.
• Rows contain citations.

Action:

• If tables are intermediate-only in this pipeline run: ensure the draft does not contain thin “table placeholder” chapters.

Recommended fix order (routing)

When the report finds issues, recommend the smallest fix path:

• Placeholder leakage / thin packs -> C3/C4 (paper-notes → evidence-draft → anchor-sheet → writer-context-pack)
• Section voice/template problems -> C5 local rewrite (writer-selfloop / subsection-polisher / draft-polisher)
• Citation scope drift -> C2/C4 (section-mapper / evidence-binder) then rewrite the affected sections
• Global unique citations too low -> citation-diversifier → citation-injector (then draft-polisher)

Safe edits allowed (optional)

If and only if edits are low-risk and do not change citation keys:

• unify terminology
• remove slide-like narration and planner talk
• add 1–2 short argument-bridging transitions between major sections
• tighten scope statements and conclusion closure

Script

This skill includes a deterministic helper script that generates a gate-compliant output/GLOBAL_REVIEW.md from the current draft and context (no invented facts/citations).

Quick Start

• uv run python .codex/skills/global-reviewer/scripts/run.py --help
• uv run python .codex/skills/global-reviewer/scripts/run.py --workspace <workspace>

All Options

• --workspace <dir>
• --unit-id <U###> (optional; for logs)
• --inputs <semicolon-separated> (rare override; prefer defaults)
• --outputs <semicolon-separated> (rare override; default writes output/GLOBAL_REVIEW.md)
• --checkpoint <C#> (optional)

Examples

• Generate a global review after merging a draft:
• uv run python .codex/skills/global-reviewer/scripts/run.py --workspace <workspace>

Freeze policy:

• If you hand-edit the review and want to freeze it, create output/GLOBAL_REVIEW.refined.ok to prevent overwrites.

Notes:

• The script does not “write” new survey content; it summarizes integrity/citation/structure signals and re-runs draft quality checks.

Troubleshooting

Issue: review flags missing citations / undefined keys

Fix:

• Run citation-verifier and ensure citations/ref.bib contains every cited key in output/DRAFT.md.

Issue: review suggests changes that would add new claims

Fix:

• Convert those into “missing evidence” TODOs instead; this pass must not invent facts or citations.