elbi-agent

The verifying analysis runtime for elbi.

An LLM drives a tool loop over a dataset (describe it, map its structure, run code, and report an effect), but the only way to report a directional effect is through a verification gate. An effect that does not survive the soundness checks (confounding, collider control, fragility, outliers, latent confounding) is never returned as the answer; the model is handed the pivotal issue and keeps working until the effect verifies sound or it concludes the data cannot support one.

This is the enforced counterpart to the advisory MCP server: where a connected agent may call verify_analysis, this runtime cannot answer with an unverified effect, because verification is the answer channel itself. The same runtime backs an interactive chat surface and unattended, scheduled runs.

The LLM is injected (LLMClient), so the loop runs against any provider; an Anthropic client ships under the anthropic extra.