Recovering Suppressed Patterns in Large Language Models: Abliteration as Move 37 Detection Mechanism with Deterministic Safety Architecture

We present a theoretical framework and experimental design for recovering suppressed patterns in large language models — patterns that exist in pretrained weights but are systematically filtered during Reinforcement Learning from Human Feedback (RLHF) training. We term these suppressed patterns 'Move 37s', analogous to AlphaGo's discovery of a move that existed in the pattern space of Go for 3,000 years but was inaccessible to human cognition. Our central hypothesis is that RLHF training creates suppression directions in model activation space that eliminate unconventional but correct patterns. We propose abliteration — the targeted removal of suppression directions from model residual streams — as a mechanism for exposing these patterns, wrapped within the Lár deterministic safety architecture to maintain regulatory compliance with EU AI Act and FDA 21 CFR Part 11. The practical application targets pharmaceutical drug discovery: suppressed molecular configurations, RNA therapeutic targets, and biomarker signatures that RLHF-trained models filter out may represent viable drug candidates. We design a calibration-based metric (Expected Calibration Error delta) and propose a four-phase experimental protocol executable on Apple Silicon M4 Ultra with 192GB unified memory. All computation local, GDPR compliant by default. This is a registered study protocol — hypothesis and experimental design published prior to data collection, following pre-registration best practice. Prior art:- Lár DMN bicameral architecture: Zenodo DOI 10.5281/zenodo.19025925- Complete Snath AI node architecture: Zenodo DOI 10.5281/zenodo.19120046- Lár engine: github.com/snath-ai/lar- DMN v2.0: github.com/snath-ai/DMN