EIF

Epistemic Integrity Framework
Apply the scientific method to every assumption your AI agent makes.

The problem

AI agents compound errors. Each assumption that goes unchallenged becomes the foundation for the next one. The dominant failure is not hallucinated facts: it is unchallenged assumptions that are never tested against evidence.

The pipeline

EIF is a Python library and MCP server (24 tools) that runs an 8-phase scientific-method pipeline on every load-bearing claim. The core engine is deterministic local math: no LLM calls, no telemetry. Self-host it, or use the optional hosted alpha.

DECLAREClassify every assumption: KNOWN / ASSUMED / GUESSED

FALSIFYCollect independent evidence (host tools, code, web)

CAUSAL_GATEPearl causal levels; confounders, disjunctive bias

CALIBRATEBayesian posterior, trust-weighted by evidence tier

CHALLENGEAdversarial critic with counter-evidence

UPDATESequential Bayes; EIG stopping rule

EXPLAINHard-to-vary check on the mechanism

RECORDProvenance record; EU AI Act mapping

ACTEvidence supports; safe to proceed

REVISEAssumptions unresolved; moderate confidence

HALTEvidence contradicts a high-consequence claim

Install

git clone https://github.com/leocelis/eif && cd eif
python -m venv .venv && source .venv/bin/activate
pip install -e ".[mcp]"

# stdio MCP server (Cursor / Claude Desktop / Claude Code)
eif-mcp-server

Self-hosted, everything stays on your machine. A private-alpha hosted endpoint is also available; see the README for both paths.