Adversarial audit layer for agentic research: mechanical gates that catch reports which lie without lying — coverage arithmetic, cluster sums, filter provenance, citation integrity. Zero-dependency CL
View on GitHub ↗Audit layer for agentic research tools: catch coverage fraud and evidence hygiene failures before they reach your knowledge base.
The agentic-research ecosystem is booming: agents crawl, synthesize, and deposit "knowledge" into wikis and reports. What none of them do is adversarially check their own output. The result is knowledge bases that look trustworthy and are unauditable.
This project is the missing audit layer — cheap, mechanical, and dependency-free. It checks the five ways a research report lies without technically lying:
Gate Fraud it catches ------ 1 · Full-set source Set came from a curated HTML page (featured list ≠ full set) 2 · Coverage arithmetic Examined items not fully accounted for by claims + unjudged 3 · Cluster arithmetic Cluster counts that don't add up to the examined total 4 · Filter provenance 249 of 260 items silently dropped, no criteria logged 5 · Unjudged listed Partial coverage with an empty "couldn't judge" list 6 · Integrity (armed when verification data is present) Cited sources that don't exist, dangling links, stale summaries, dead URLs, hash drift
Plus per-claim verdicts: verified (≥2 independent sources) / single-source / unverifiable. Duplicate sources (same URL or name) count once — a mirror is not an independent source.
Beyond arithmetic: real source verification
Two additions turn this from a checklist into actual verification:
verify.py — source authenticity. HTTP HEAD → content SHA-256, with on-disk caching, size caps, and honest failure semantics:
From the project README.
Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:
[](https://mcp.liqiwa.com/s/chenhz01--adversarial-research-audit.html)
Typed judgment tools for MCP agents. TypeSafe's Jev model as verify, screen, find, classify, rerank, decide, compare, extract, review, gate, and score: the model judges, policy decides auto, review, o
GrubbyLee/skill-managerZero-dependency CLI to scan, recommend, deduplicate, audit, and visualize Claude Code / Codex skills and MCP servers.
menot-you/n-memoryMemory your agent can trust — every answer carries its source; when it doesn't know, it says so. Hermetic, local, MCP.
flytomoon/al-buddy-memoryPortable, governed, model-agnostic memory for AI agents: facts invalidated never deleted, raw text as truth, one documented export format.
2akouwu/reverifyStop your AI from making things up — it proposes, deterministic tools decide, every claim checked against ground truth with evidence. Grounded facts and context survive resets. Reverse engineering is
Stupidoodle/swissdevjobs-cliSearch & apply to ~4,700 salary-transparent tech jobs across 7 countries (🇨🇭🇩🇪🇬🇧🇺🇸🇨🇦🇳🇱🇫🇷) from your terminal or AI agent — zero-dependency Python CLI + MCP server + Claude Code plugin
The top new MCP servers of the week, every Monday. No spam, unsubscribe anytime.