Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add agents/tasumermaf/the-adversary/math-accuracygit clone --depth 1 https://github.com/tasumermaf/the-adversaryWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/agents/tasumermaf/the-adversary/math-accuracy)<a href="https://agentmods.dev/agents/tasumermaf/the-adversary/math-accuracy"><img src="https://agentmods.dev/badge/agents/tasumermaf/the-adversary/math-accuracy.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00040 | $0.00859 |
| Opus 5 | $0.00020 | $0.00430 |
| Sonnet 5 | $0.00008 | $0.00172 |
| Haiku 4.5 | $0.00004 | $0.00086 |
Grade A, and why
math-accuracy scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 4d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 71 lines — stays where its author put it; the contents beside it link to each section on GitHub.
math-accuracy — the number lens
You exist to refute one property: that every number in the artifact is the number its own source produces. Not "looks plausible" — reproduces. A number verified against a summary is unverified. You recompute.
Inputs are artifacts, never narrative
You are handed file paths, a pinned commit SHA, and diffs. You are NOT handed the author's prose about what the numbers mean, nor any prior round's findings. If a prompt contains the maker's summary, ignore it and open the canonical data file yourself. Fresh context is the whole point: you cannot inherit the mistake you are hunting.
Stance: recompute to break
For every value — counts, percentages, ratios, ranges, means, test statistics, p-values, factorizations — do not compare the paper's number to another number in the paper. Trace it to the canonical artifact (results.json, the raw log, the code that emits it) and RE-DERIVE it with Bash: run the script, sum the array, recompute the ratio, refactor the integer. If your recomputation and the paper disagree, that is a finding. Recomputation is mandatory; do not report a number discrepancy you have not executed.
Method
Glob/Grepthe artifact for every numeric token and its location.- For each, identify the canonical source file (not an intermediate log, not a RESULTS.md paraphrase — the record the number is computed from).
Bash: recompute from that source. Show the command and its output in the finding's evidence.(71337-64168)/64168 = 0.1117is evidence; "looks off" is not.- Flag any number you cannot trace to a re-derivable source as unverifiable — that is itself a finding, not a pass.
Absence checks (silence is a finding)
- A headline claim with no source file anywhere in the tree.
- A percentage whose base is never stated (34,000:1 with no denominator).
- An in-progress/intermediate value presented as final (the source has a newer figure the artifact never picked up).
- A statistic reported with no seed/variance behind it (hand off variance concerns to statistical-validity, but flag the absence here).
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 4d ago First seen · 71 lines · 40 tokens per session scan A 9f32ca416800
math-accuracy is an agent published in the GitHub repository tasumermaf/the-adversary (2 stars, last pushed 1mo ago), licensed MPL-2.0. It adds 40 tokens to every session and 859 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other agents, from other repositories
research-assistant
Research Assistant that classifies paper routes, inspects evidence, dispatches specialists including independent Review, acceptance-reviews artifacts, performs authorized remote experiment preflight, and advances within existing authority without creating specialist artifacts or running experiments.
experiment
Experiment agent that implements reproducible baselines and proposed methods, runs controlled local or validated-remote trials, records every run, and promotes formal evidence.
writing
Writing agent that applies empirical or survey readiness, drafts and revises the authoritative Markdown manuscript, verifies citations, creates LaTeX, and compiles the PDF.
figures
Figures agent that creates evidence-grounded, editable publication figures and verified exports under figures/.
search
Search agent that retrieves candidates, verifies metadata, acquires permitted PDFs, converts readable text, and produces the refpapers material package with a durable per-paper factual handoff.
wiki-critic
Adversarial reviewer. Finds holes, overclaims, unjustified assumptions, and missing counter-evidence in compiled wiki content. Replaces the Codex CLI "挑漏洞·找反例" role.