---
title: "Math-proof judge: review, role & definition · Undominated.ai"
canonical: https://undominated.ai/agents/anthropic-math-proof-judge/
description: "Runs one named step of a math-proof siege: writes only the files that step names, keeps notes under 4000 words, and requires proof.md to stand alone."
---

# Math-proof judge: review, role & definition · Undominated.ai

> Runs one named step of a math-proof siege: writes only the files that step names, keeps notes under 4000 words, and requires proof.md to stand alone.

[← Explore all agents](/agents/)

MATHEMATICS / Anthropic

# Math-proof judge

Runs one named step of a math-proof siege: writes only the files that step names, keeps notes under 4000 words, and requires proof.md to stand alone.

 [Use this definition ↓](#setup)[Original source ↗](https://github.com/anthropics/claude-plugins-official/blob/b78ac49cdc6b3d7b61c4439470e311f4291265b1/plugins/math-proof/agents/math-proof-judge.md)

SOURCE REVIEW

 Reviewed 2026-10-14
 Evidence 4 linked sources
 Publisher Anthropic
 Licence [Apache-2.0 ↗](https://github.com/anthropics/claude-plugins-official/blob/b78ac49cdc6b3d7b61c4439470e311f4291265b1/plugins/math-proof/LICENSE)
 Revision b78ac49cdc6b
 [Read what was—and wasn’t—checked ↓](#review)

“The math-proof judge: it plans each round's questions, reads the answers, keeps the summary and ledger, and writes up, revises and finalizes proof.md. It does one step per launch and is launched only by the math-proof plugin's siege skill.”

 [Anthropic · upstream description ↗](https://github.com/anthropics/claude-plugins-official/blob/b78ac49cdc6b3d7b61c4439470e311f4291265b1/plugins/math-proof/agents/math-proof-judge.md) Our analysis follows below.

01 / THE REASONING

## Why this made the selection.

 - Each launch is one step with no memory except the named files, and every query file must stand alone because the worker sees only that query plus the problem statement.
- proof.md must contain the argument itself, because a referee cannot open the other files, and a marked gap is preferred to a papered-over proof.

### A good fit for

 - Provide one siege step brief naming the input files, output files and required formats; obtain question files, ledger/summary updates or a self-contained proof draft.
- Use the judge to organise separately generated arguments while preserving explicit gaps and failed approaches in persistent notes.
- Supply a siege step brief naming all permitted inputs, outputs and prior state files.
- Produce a self-contained proof document whose lemmas remain checkable without opening worker-run files.

### Weigh up before choosing

 - This is one seat in the siege protocol. A standalone copy cannot orchestrate the workflow, and ledger bookkeeping is not a mathematical proof checker.
- Read/Write/Edit/Bash/Glob/Grep are granted. Shell use is narrowed in prose to installed Python checks, text splicing and file handling; those limits still need host enforcement.
- The source has no web tool and literature claims may be recalled from memory. The resulting proof needs independent mathematical review.
- The plugin licence is Apache-2.0; its appendix retains the template copyright line. Installation and model compatibility were not tested.
- The allowlisted shell purposes are prompt instructions, not a process sandbox; proof completeness still needs independent checking.

02 / THE REVIEW RECORD

## What we actually inspected.

Source review has boundaries. A clear record is more useful than a “safe” badge.

### Material inspected

 - plugins/math-proof/agents/math-proof-judge.md
- plugins/math-proof/LICENSE
- plugins/math-proof/README.md (installation, prerequisites, use, resume and warnings)
- plugins/math-proof/skills/siege/SKILL.md (frontmatter, setup and orchestration rules; not a full protocol audit)
- Round 2: pinned definition, output contract and permission/dependency boundaries reread on 2026-10-14

### Our findings

 - The licence file is the Apache License, Version 2.0. The appendix line is Copyright [yyyy] [name of copyright owner].
- Tools are Read, Write, Edit, Bash, Glob, and Grep. model is inherit, effort is medium, and maxTurns is 120. Notes stay under 4000 words. The problem statement is not copied into a query.
- The orchestration passes the complete authoritative problem to each judge and launches a fresh agent for every step; previous state comes from named files.
- The setup checks for ledger.py and stops if it cannot locate the plugin or a working Python interpreter. Resumes retain the original problem and recorded settings.
- The README documents the installation command; this closes the earlier review gap in which only the role file had been read.
- The judge may not cite query/answer filenames as proof; it must rewrite the needed argument into proof.md.
- Its computation allowance is limited to concrete Python checks, text splicing and plain file handling, with SymPy used only if already installed.

### Not established by this review

 - The agent was not executed; model behaviour, host compatibility and task outcomes were not runtime-tested.
- No plugin installation, proof run or bookkeeping-script execution was performed.
- The remaining siege protocol and ledger.py implementation were not fully audited.

The review applies to the material and revision named here. A newer upstream release can change its behavior.

03 / PUT IT TO WORK

## Use the role in your project.

[Upstream setup instructions ↗](https://github.com/anthropics/claude-plugins-official/blob/b78ac49cdc6b3d7b61c4439470e311f4291265b1/plugins/math-proof/agents/math-proof-judge.md)
 - Copy the judge file at the reviewed commit into the location your host uses for agent prompts.
- No install command is published. Installing the whole math-proof plugin was not reviewed.

### Before you start

 - A step brief that names the files to read and write
- A host whose tool grants you are willing to live with, because the prompt's shell limit is not an allowlist

### Compatibility

Hosts that can load a Markdown agent file with a tool list

### Implementation

Markdown

### Writes the files named in the step brief. Bash is granted and then narrowed only in prose.

 - Read, write, and edit the files the step brief names
- Bash, Glob, and Grep, with the shell uses named in the body

### Cost model

Apache-2.0 instructions. Each judge and worker call uses the host’s model allowance or API budget; no cost or proof-success benchmark was measured.

04 / FOLLOW THE EVIDENCE

## The source trail.

Our notes are separate from the original resource. Check upstream before adopting a new version.

 - [Math-proof judge ↗](https://github.com/anthropics/claude-plugins-official/blob/b78ac49cdc6b3d7b61c4439470e311f4291265b1/plugins/math-proof/agents/math-proof-judge.md) Checked 2026-10-14 https://github.com/anthropics/claude-plugins-official/blob/b78ac49cdc6b3d7b61c4439470e311f4291265b1/plugins/math-proof/agents/math-proof-judge.md Supports: summary, whySelected, limitations
- [Apache-2.0 licence ↗](https://github.com/anthropics/claude-plugins-official/blob/b78ac49cdc6b3d7b61c4439470e311f4291265b1/plugins/math-proof/LICENSE) Checked 2026-10-14 https://github.com/anthropics/claude-plugins-official/blob/b78ac49cdc6b3d7b61c4439470e311f4291265b1/plugins/math-proof/LICENSE Supports: Redistribution terms, Unfilled copyright appendix
- [Math-proof plugin setup and operating limits ↗](https://github.com/anthropics/claude-plugins-official/blob/b78ac49cdc6b3d7b61c4439470e311f4291265b1/plugins/math-proof/README.md) Checked 2026-10-14 https://github.com/anthropics/claude-plugins-official/blob/b78ac49cdc6b3d7b61c4439470e311f4291265b1/plugins/math-proof/README.md Supports: Whole-plugin installation, Runtime prerequisites, Model usage and resume procedure
- [Siege orchestration: setup and rules ↗](https://github.com/anthropics/claude-plugins-official/blob/b78ac49cdc6b3d7b61c4439470e311f4291265b1/plugins/math-proof/skills/siege/SKILL.md) Checked 2026-10-14 https://github.com/anthropics/claude-plugins-official/blob/b78ac49cdc6b3d7b61c4439470e311f4291265b1/plugins/math-proof/skills/siege/SKILL.md Supports: Agent dependencies, Completion-marker handling, Fresh worker launches and interruption recovery

KEEP COMPARING

## Other approaches to consider.

Related by category or shared topics. These are alternatives to inspect, not a measured quality order.

 [### Math-proof worker ↗ Answers one self-contained mathematics question into a file, separates proved claims from sketches and open questions, and appends the task's end marker only when finished.](/agents/anthropic-math-proof-worker/)[### Math proof worker, deep ↗ Pins the higher-effort math-proof worker that answers one question per launch and writes the answer to a file the orchestrator reads.](/agents/anthropic-math-proof-worker-deep/)[### Adversarial modeler ↗ Pins Trail of Bits' differential-review agent that models high-risk changes and says not to run on low or medium risk diffs.](/agents/trailofbits-adversarial-modeler/)

 [AI Tools ↗](/tools/)[Skills ↗](/skills/)[Agents ↗](/agents/)[MCP Servers ↗](/mcp-servers/)[Workflows ↗](/workflows/)

## Continue your investigation

 - [Find reusable instructions](/skills/)
- [Inspect connections](/mcp-servers/)
- [Build a toolkit](/tools/)
- [Choose the model](/compare/)
