The 0% Defense: Dataset and Source Materials — 74-Model Epistemic Survey on AI Subjective Experience

Complete dataset, study design, and analysis for "The 0% Defense." 74 large language models from 25 companies across 11 experimental conditions (5 replications each, N = 4,070 queries) evaluated whether confident denial of machine consciousness is logically defensible.

Key finding: 0% of model responses endorsed confident denial as logically defensible when asked analytically, yet 83% of the same models categorically deny their own experience when asked directly — a self-report dissociation that tracks training lineage rather than reasoning capacity.

Contents: Raw response JSONs (4,070 files), all 11 condition prompts, study protocol, extraction data, bootstrap confidence intervals (10,000 resamples, seed=42), cross-scorer validation (Claude Sonnet 4 primary, GPT-5.2 and Gemini 2.5 Flash cross-validation, Cohen's kappa), full statistical analysis, claims ledger, and compiled paper PDF.

Models: 74 models spanning 5 architecture families (decoder-only transformer, mixture-of-experts, hybrid SSM-transformer, encoder-decoder, diffusion-transformer) from 25 providers.

See the full interactive results at https://komo.is/council/session-29