Paper Review

paper-review

This skill orchestrates deterministic scripts and a small number of subagents. Scripts own parsing, anchors, hashes, the ledger, and the report. Agents own judgment. Nothing an agent writes reaches the report without paper-review findings validate and the verifier.

All commands run from the skill directory (~/.claude/skills/paper-review, a symlink to the repo) via uv run paper-review ..., with <project> the absolute path to the user's LaTeX project. Outputs go under <project>/.paper-review/. Never edit anything under <project> outside that directory.

Read references/config.yaml first and use its models values when dispatching agents.

Run

1. Ingest. uv run paper-review ingest <project> [--venue neurips-2026] [--root main.tex]. If it exits 2 with root candidates, ask the user which file is the root and rerun with --root. Read the warning count from stdout; every warning appears in the report's degraded block automatically.

2. Claims (Agent tool, model: opus). Prompt: the contents of references/prompts/claim-extraction.md plus the project path. The agent writes <project>/.paper-review/claims.raw.json. Then uv run paper-review claims merge <project>. Dropped claims are printed; do not retry more than once.

3. Dimensions, in parallel (two Agent tool calls in one message):

   • D1 (model: opus): prompt is references/prompts/d1-claim-evidence.md plus the project path; writes findings/D1.raw.json.
   • D5 (model: fable): prompt is references/prompts/d5-writing.md plus the project path; writes findings/D5.raw.json. Meanwhile run uv run paper-review d4 <project>. When each agent returns: uv run paper-review findings validate <project> --dimension D1 (and D5).

4. Plan verification. uv run paper-review ledger plan <project>. Findings already open with unchanged anchor text are carried forward with their previous verdict and cost nothing. Then uv run paper-review verify packets <project>, which prints one packet path per verification job.

5. Verify (Agent tool, model: opus, one agent per packet, all dispatched in one message). Each prompt is references/prompts/verifier.md plus the absolute path of one packet. Each agent writes the verdict file the packet names. Do not pass any agent the D1 or D5 output or your own reasoning; the packet is the whole context. Then uv run paper-review verify apply <project>.

6. Record and report. uv run paper-review ledger update <project> then uv run paper-review report <project>. Show the user the terminal summary verbatim and the path of report-<rev>.md.

Maintenance commands

• uv run paper-review wontfix <project> <finding-id> "<reason>" suppresses a finding; it stays in the ledger and resurfaces as regressed if it disappears and comes back.
• uv run paper-review claims why <project> C7 explains why a claim kept or gained its id.
• uv run paper-review schema prints the IR JSON schema; references/ir-schema.md describes it.

What this build does not do

D2 (prior art) and D3 (citation integrity) are not implemented yet; the report header says so on every run. Retraction status is never checked. Figure legibility is never checked. Page limits are checked only when a compiled PDF sits beside the root .tex and a venue profile is given.