Show HN: Theory of Mind benchmark for 8 LLMs with reproducible markers

I built a formal protocol (FPC v2.1 + AE-1) to detect behavioral uncertainty in large language models. The goal is enabling safer AI deployment in critical domains medicine, autonomous vehicles, government where confident hallucinations can lead to high-stakes failures.

Current benchmarks focus on accuracy but miss reasoning coherence under stress. This protocol uses tri-state affective markers (Satisfied / Engaged / Distressed) to detect when models lose logical consistency, allowing abstention instead of confident hallucination.

We evaluated 8 models (Claude, GPT-4 families). Only Claude Opus reached full ToM-3+. GPT-4 family consistently failed third-order reasoning. Extended temperature tests (Claude 3.5 Haiku, GPT-4o) showed 180/180 stable AE-1 matches (p≈1e-54), independent of sampling temperature.

Dataset: https://huggingface.co/datasets/AIDoctrine/FPC-v2.1-AE1-ToM-...

A demo notebook exists for replication. Looking for feedback on methodology and possible applications in safety critical AI.