A Falsifiable, Anti-Goodhart Protocol for Testing Internal Objective Channels in Reinforcement Learning Agents (AILab-aligned reference implementation)

This release provides a falsifiable, Anti-Goodhart evaluation protocol (“AILab”) for testing functional/autonomous intentionality claims in RL agents, together with a minimal baseline implementation and controlled ablations. The key contribution is methodological: we show how seemingly strong proxy improvements can arise without corresponding evidence of deeper, invariant structure. Across multiple controlled runs (including an explicit no-probe negative control, an alpha sweep, and T3 threshold / gate-ablation sweeps), the protocol’s gates (invariance/CCA, self-causal drop, and intervention persistence) prevent “illusion wins” and expose sensitivity to decision thresholds. In the present setup, the results do not support an ONTO-CORE / ontologically structured agent claim; instead, they document reproducible failure modes and provide a clean template for future redesigns (e.g., environments without information leakage). This Zenodo record includes: the PDF paper (protocol + baseline + controls + failure modes), an Overleaf-ready LaTeX ZIP, supporting run artifacts (decision.json summaries and bundles) enabling replication of the reported tables and verdicts. Recommended citation: use the Zenodo DOI of this record. Keywords: Anti-Goodhart, falsifiable evaluation, reinforcement learning, intentionality, invariance, causal interventions, reproducibility.