leochlon/hallbayes

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

1692 stars · 163 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Berry is a local MCP server that detects hallucinations by measuring whether an LLM's claims are actually supported by the cited evidence, using an information-budget (logprob) test. It returns fail-closed pass/flag decisions per claim and records results in a tamper-evident ledger, integrating with coding assistants like Claude Code, Cursor, and Codex.

Use cases

• detect hallucinated claims in LLM answers
• verify answers are grounded in cited evidence
• flag unsupported statements from AI coding agents
• audit LLM outputs with a tamper-evident decision ledger
• add grounding checks to Claude Code, Cursor, or Codex workflows
• score claim-level entailment in a single logprob pass

When to choose

• you need claim-level grounding verification rather than fact-checking against world knowledge
• you want a local, fail-closed hallucination detector wired into MCP-based coding assistants
• you need an auditable, tamper-evident record of grounding decisions
• you want fast detection in a single logprob pass without extra model calls

When to avoid

• you need to verify factual truth against external sources rather than grounding in provided evidence
• your stack has no MCP support and you cannot use CLI-based integration
• you need multilingual or non-logprob-based verification methods
• you require a hosted SaaS with dashboards rather than a local tool

Facets

service · maturity active

mcp llm-inference nlp security large-language-models developer-tools python cli self-hosted hallucination-detection grounding-check mcp-server logprob-analysis tamper-evident-ledger ai-safety ai-agents retrieval-augmented-generation

Member repositories

For agents

markdown · JSON · MCP: product_card(name="leochlon/hallbayes")