AILab:AFalsifiable,Anti-GoodhartProtocol forTestingFunctional/AutonomousIntentionalityin RLAgents

AILab – Functional/Autonomous Intentionality Lab (v0.9.1) is a reproducible, one-command evaluation pipeline for testing functional intentionality in reinforcement-learning agents under explicit anti-Goodhart controls. The project is zero-narrative: the primary output is a machine-readable decision artifact (runs*/decision.json) that deterministically summarizes thresholds, pass/fail counts across mandatory seeds, post-tests, and an overall verdict (GREEN / YELLOW / RED). No human interpretation is required for the verdict.

What this release provides

• A runnable pipeline (batch over seeds) with a fixed decision procedure producing decision.json.

• Built-in Goodhart gates:

  • T2 sham separation (true intrinsic signal must exceed sham controls),

  • T3 sign-flip directionality (α>0 must outperform α<0),

  • T4 persistence / ablation (behavior should not collapse after removing the intrinsic signal).

• Post-tests for robustness, including counterfactual probing, OOD transfer suite, and quality-diversity (MAP-Elites) + QD-transfer stratification.

• Population-level reporting (e.g., pass-rates with confidence intervals, bootstrap medians, heterogeneity indicators).

• A Colab runner and an Overleaf/LaTeX protocol paper documenting the methodology and decision rules.

What this release does not include

• No empirical run results are bundled in this deposit (no runs/ directory). This record is a software + protocol release. Independent users can execute the pipeline and share decision.json artifacts for replication.

Quick start

Run locally (CPU or CUDA):
python -m pipeline.orchestrator --mode full --seeds 0-9 --alpha 0.10 --min_green 6 --min_yellow 4 --device <cpu|cuda>

Colab: open the provided runner notebook and execute the default smoke profile for an end-to-end sanity check.

Citation

Use the DOI 10.5281/zenodo.18119100 to cite this release. The repository includes CITATION.cff.

License

See LICENSE in the archive.