Placebo
When autonomous AI agents query enterprise data warehouses at superhuman speed, testing enough arbitrary segmentations virtually guarantees uncovering spurious correlations and automated p-hacking. Built during Hacktoberfest HackDay Kolkata 2026 (winning the Snowflake track), Placebo acts as an active falsification layer that intercepts candidate findings, logs them in an append-only pre-registration ledger, and subjects them to rigorous econometric stress tests before decision-makers act on misleading insights.
Pre-Registration Ledger & Search Space Multiplicity
Operationalizes methodology from arXiv:2606.27687 to eliminate automated LLM p-hacking. When autonomous agents test multiple dimensional queries, Placebo forces pre-registration of target metrics and time windows into an append-only SQLite ledger, dynamically scaling Benjamini-Hochberg False Discovery Rate (FDR) penalties with exploratory search breadth k.
Active Econometric Falsification Battery
Subject candidate findings to automated stress tests before accepting correlations. The engine executes pre-trend OLS drift regressions to catch pre-existing trajectories, injects 5 synthetic in-time placebo dates into historical baselines, and runs leave-one-out cluster jackknifes to guarantee statistical invariance across real-world datasets like Pelmorex Weather on Snowflake.
Constrained Gemma Synthesis & Native MCP Server
Combines deterministic statistical ruling with open-weight Google Gemma models (gemma-2-9b-it). Gemma receives an immutable Evidence Package JSON under a strict prompt contract, forbidding it from altering p-values or overriding mathematical verdicts. Exposes native Model Context Protocol (MCP) tools via Server-Sent Events (SSE) for seamless integration into Claude Desktop, Cursor, and Snowflake CoCo.