HIDEKI-SQ/cot-reliability-gating: v3.0.0 - Initial Public Release

Initial Public Release Overview 24 final experiments with 64,004 trials 4 history experiments (7,192 trials) preserved for transparency Complete reproducibility: all notebooks, data, and results included Key Findings Consumer model determines vulnerability to misleading cues Corruption increases cue reliance (GPT-4o: 0.5% → 8.5%) Semantic corruption (especially quantity shifts) most effective Simple verification prompts eliminate vulnerability (12% → 0%) Models Tested Claude Sonnet 4 (claude-sonnet-4-20250514) GPT-4o Contents experiments/ - Final experiment results history/ - Superseded experiments (transparency) notebook/ - 29 Jupyter notebooks data/ - Shared datasets and traces figures/ - Paper figures Documentation EXPERIMENT_INDEX.md - Complete experiment catalog CHANGELOG.md - Version history docs/REPRODUCIBILITY.md - Reproduction guide