papercheck

A reproducible audit harness for mathematical LaTeX papers.

papercheck extracts a paper's structure, records findings in schema-validated ledgers that are checked against the exact source text, and runs a final gate that reports whether the paper is ready. It runs as a command-line tool, or as an MCP server that an agent (Claude Code, Codex, Cursor, or similar) drives while you supply the mathematical judgement.

The point is discipline. Instead of asking one model to "review the paper," papercheck segments the manuscript, inventories the claims, runs narrow hostile audits, verifies each finding against the source, adjudicates, patches only what was accepted, and gates. The gates are enforced in code: an agent cannot patch before adjudication, and a finding whose quote does not appear in the source never enters the ledger.

The staged pipeline — segment, budget, specialist review, synthesize, adjudicate — follows the approach described in Google Research's Paper Assistant Tool (PAT)[](#references)<<<[1], adapted into a small, open, model-agnostic harness that runs no models of its own.

A deterministic Python core with two frontends — a CLI and an MCP server. papercheck itself never calls a model; the reasoning lives in the agent or in your hands at the CLI. What the core provides:

- LaTeX-AST structure extraction (theorems, labels, references, citations, equations, draft markers) intostructure.json.
- Schema-validated issue ledgers, each finding tied to a file and line.
- Quote verification at intake: if a finding's quoted text is not in the source, it is rejected asREJECTED_SOURCE_TARGET_INVALIDbefore it reaches the proposed ledger.
- A stage-gated state machine (INIT → SCANNED → … → ADJUDICATED → … → GATED) that refuses out-of-order operations.
- A CLI, an MCP server (29 tools), and a local web UI.

A single "review my paper" prompt tends to fail in three ways. papercheck handles each in code rather than in prompting:

pipx install papercheck # or: pip install papercheck papercheck scan path/to/paper # extract structure.json papercheck segments path/to/paper # propose audit segments and budgets papercheck gate path/to/paper --mechanical-only papercheck report path/to/paper # self-contained HTML report papercheck serve path/to/paper # local web UI

Run it on the bundled example (prints==== READY ====):

papercheck gate examples/toy_clean_paper --mechanical-only

papercheck ships a FastMCP server exposing 29 tools plus the audit prompt pack. Register it once:

claude mcp add papercheck -- papercheck-mcp

Then ask the agent to audit a paper. It walks the workflow through the MCP tools —init_audit,run_scan,propose_segments,submit_issue,adjudicate_issue,run_gate— under the same code-enforced gates as the CLI.

papercheck does not call a model, so tailoring the audit to a paper's field is a two-step split: the agent reads the paper and drafts the pack, papercheck validates and stores it.

papercheck packs scaffold --paper-root ./paper # deterministic draft from the scan papercheck packs create draft.json --paper-root ./paper # validated -> Paper_Audit/domain_pack.json

The same is available through thescaffold_domain_packandcreate_domain_packMCP tools. Generic packs ship for stochastic analysis, PDE, numerical analysis, optimization, machine-learning theory, and a general fallback.

flowchart LR A[INIT] --> B[SCANNED] --> C[SEGMENTED] --> D[INVENTORIED] --> E[AUDITING] E --> F[SYNTHESIZED] --> G[ADJUDICATED] --> H[PATCH_PLANNED] H --> I[PATCHING] --> J[REGRESSED] --> K[GATED] E -. submit_issue .-> V{quote and label verified?} V -- no --> X[REJECTED_SOURCE_TARGET_INVALID] V -- yes --> L[PROPOSED ledger] style X fill:#7f1d1d,stroke:#ef4444,color:#fff style K fill:#14532d,stroke:#22c55e,color:#fff

One deterministic core, two frontends, no LLM calls inside the harness. papercheck stays strictly an MCP server and CLI — no provider adapters, no self-orchestration — so it works with any model and makes no network calls of its own.

For a paper repository,docs/ci.mddescribes a mechanical, LLM-free GitHub Action that runsscanandgate --mechanical-onlyon every pull request and sends nothing to any model provider.

Seedocs/limitations.mdfor the full version.

- papercheck is not a theorem prover and not a replacement for peer review.
- Catching a semantic error depends entirely on the driving model. papercheck keeps the process disciplined and traceable, but AI findings must be checked independently.
- papercheck makes no network calls. The agent you drive it with may transmit your manuscript to its model provider, so review the relevant data terms before auditing unpublished work (docs/privacy.md).

Contributions welcome — seeCONTRIBUTING.md. The architecture is fixed for the 0.x line: strictly MCP and CLI, JSON as the source of truth, gates enforced in code. Prompt changes are guarded by the eval fixtures described indocs/agent_eval.md.
- Rajesh Jayaram, Drew Tyler, David Woodruff, Corinna Cortes, Yossi Matias, Vahab Mirrokni, and Vincent Cohen-Addad.Towards Automating Scientific Review with Google's Paper Assistant Tool.arXiv:2606.28277, 2026.https://arxiv.org/abs/2606.28277

BibTeX for the above is indocs/references.bib.

MIT — seeLICENSE. If papercheck is useful in your work, cite it viaCITATION.cff.

Sync Calendars, Scheduling Links, AI Executive Scheduling Assistant, Unified Calendar