https://github.com/zhjai/groundcheck
Single-agent, evidence-grounded claim verification to catch LLM hallucinations — a pluggable fact-gate for agent-arena and any multi-agent system (CrewAI, AutoGen, LangGraph).
https://github.com/zhjai/groundcheck
agent-skill claim-verification claude-code-skill fact-checking groundedness hallucination-detection llm rag-evaluation
Last synced: 4 days ago
JSON representation
Single-agent, evidence-grounded claim verification to catch LLM hallucinations — a pluggable fact-gate for agent-arena and any multi-agent system (CrewAI, AutoGen, LangGraph).
- Host: GitHub
- URL: https://github.com/zhjai/groundcheck
- Owner: zhjai
- License: mit
- Created: 2026-05-30T20:00:16.000Z (17 days ago)
- Default Branch: main
- Last Pushed: 2026-06-01T16:33:14.000Z (15 days ago)
- Last Synced: 2026-06-12T23:35:59.765Z (4 days ago)
- Topics: agent-skill, claim-verification, claude-code-skill, fact-checking, groundedness, hallucination-detection, llm, rag-evaluation
- Size: 25.4 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- Changelog: CHANGELOG.md
- License: LICENSE
Awesome Lists containing this project
README
# GroundCheck
English · 中文
> **Single-agent, evidence-grounded claim verification** — catch fabricated facts by checking each claim against real evidence, not by asking more models to agree.
GroundCheck takes content that already exists — a generated answer, a report, RAG output, docs, or code — extracts the verifiable claims, grounds each one in real evidence, and returns a per-claim verdict with citations plus an action: **keep / revise / retract / send back**.
## Why single-agent (and why that matters)
Multi-agent debate is good at fixing **overconfidence** — but it can *reinforce* a shared **hallucination**, because several models trained on similar data confirm the same wrong fact. GroundCheck deliberately does **not** use a panel: it grounds every claim in deterministic external evidence (tests, source, docs, web, retrieved context).
- **GroundCheck** → treats **hallucination** (single-agent + evidence grounding)
- [**agent-arena**](https://github.com/zhjai/agent-arena) → treats **overconfidence** (multi-agent debate)
They form one verification stack at two depths and interoperate via a shared [Claim Ledger](interop/claim-ledger.md).
## What it produces
For each claim: an atomic, classified, dated verdict —
`supported · partially_supported · refuted · unverifiable · outdated · needs_qualification` —
with cited evidence and a recommended action. Compound claims are decomposed so a true
sub-claim can't launder a false bundled conclusion.
The ledger is **contestable, not authoritative**: every verdict is traceable, time-bounded,
and can be reopened with stronger counter-evidence. The verifier can be wrong too.
## Use as a fact-gate in any multi-agent system
GroundCheck is a **generic, pluggable verification gate** — agent-arena, CrewAI, AutoGen,
LangGraph, or any orchestrator can plug it via the [fact-gate contract](interop/fact-gate.md):
```
each agent answers independently
→ PRE-DEBATE gate: groundcheck per answer
→ refuted? send back to that agent (with evidence) before debate
→ debate
→ POST-DEBATE gate: re-check new/changed claims
```
Catching factual errors **before** debate is what stops a panel from reinforcing a shared
hallucination. Send-backs carry **evidence, not conclusions**, and the original answer stays
immutable — so the source agent reasons independently instead of appeasing the checker.
## When to use / not use
**Use:** verify claims · check for hallucinations · "are these citations / numbers / APIs real" ·
fact-check before publishing · RAG groundedness · as a fact-gate in a multi-agent flow.
**Not for:** pure opinions / creative content · code logic already covered by tests ·
"which option is better" decisions (that's agent-arena).
## Install
```bash
npx skills add zhjai/groundcheck -g -a claude-code
```
Works with Claude Code, Codex, Cursor, OpenCode, and other Agent Skills hosts. Or copy
`skills/groundcheck/` into your agent's skills directory.
## Status
`v0.1.0` preview. Not affiliated with any vendor. MIT licensed.