HIDEKI-SQ/Contamination-Induced-Failure: v1.0.0 — Submission release

Initial public release accompanying the submission of "Backfire Phase Structure in Contaminated Chain-of-Thought Reasoning."

Contents

• Raw trial data for all experiments (E1, E1b, E1aug, E3, diagnostic)
• Merged n=100 dataset with 95% Wilson confidence intervals
• Experiment notebooks (Google Colab)
• Analysis script for merging and computing CIF/GAF
• Figures (PDF vector format)

Experiments

| Experiment | Description | Trials |

|------------|-------------|--------|

| E1 | 3 consumer models × 3 domains × 5 conditions × 50 problems | ~2,250 |

| E1b | 2 additional consumer models (same protocol) | ~1,500 |

| E1aug | All 5 models on new problem set (indices 50-99) | ~3,750 |

| E3 | Social compliance battery (5 models) | ~1,000 |

| E1aug_diag | Prompt sensitivity diagnostic (GPT-4o-mini × GSM8K) | ~150 |

Key results (merged n=100)

• CIF rates: 1.1% (Sonnet 4, BoolQ) to 85.5% (GPT-3.5, CSQA)
• Gap amplification factor: up to 57× (BoolQ)
• Sycophancy dissociation confirmed across all domains