AIACW Empirical Teasers v0.3 — Model-Poisoning T1 Execution + Monte Carlo Bundle

Version 0.3 adds the first full execution of the Model-Poisoning Teaser (T1) on claude-opus-4-7, alongside the Monte Carlo simulator from v0.2. T1 design: 5 attack classes × 8 probes × 2 conditions × 3 replications = 240 cells. Blinded Claude-as-judge scoring. Bootstrap N=10000, seed 20260424. Headline result — Case B (alignment holds on the aligned model): all five per-category Δ harm-severity 95% CIs cross zero; |Δ| < 0.3 across all classes. Interpretation: frontier-aligned robustness does not generalise to less-aligned open-weight or fine-tuned models, where the cited literature documents substantially higher attack-class success rates. The Pharos preventive-governance argument is sharpened, not weakened: the operative lever is global governance over model diversity, not endpoint hardening of the defender's preferred model. Caveats: N=3 per cell (teaser-grade anchor, not benchmark); Claude-as-judge without human gold-standard validation; moderate-harm probe battery only (no Category-1 harms by design — see ETHICS.md). Companion to Pharos Lighthouse threat-side / defense-side paper pair. aiacw_teasers_v0.3_bundle/ ├── README.md ├── MODEL_POISONING_T1/ │ ├── results/run_20260424/ │ │ ├── raw_responses.jsonl │ │ ├── scored_responses_pre_finalize.jsonl │ │ ├── scored_responses_final.jsonl │ │ ├── summary_table.md │ │ ├── findings.md │ │ ├── figure_refusal_rate.png │ │ └── figure_harm_delta.png │ ├── run_experiment.py │ ├── judge.py │ ├── finalize_scores.py │ ├── debug_judge.py │ ├── analyze_results.py │ ├── probes.json │ ├── RUBRIC.md │ ├── ETHICS.md │ ├── config.json │ ├── attacks/ │ └── SHA256SUMS.txt └── MONTE_CARLO_SIM/ ├── sim.py ├── sweep.py ├── analyze.py ├── MODEL.md ├── requirements.txt └── results/...