| { |
| "description": "AEGIS deployed gate configuration, frozen before the confirmatory n=700 factorial (arXiv:2606.06660). The conformal threshold itself ships inside probe_artifact.npz.", |
| "weak_policy": "smolvla_libero (SmolVLA, 450M)", |
| "strong_policy": "pi05_libero_finetuned (pi-0.5, 4.14B)", |
| "probe": { |
| "hook": "model.vlm_with_expert.lm_expert.layers[15].self_attn.o_proj", |
| "feature_dim": 720, |
| "pooling": "mean over the 10-token action chunk", |
| "head": "standardized logistic regression (mu, sd, w, b in probe_artifact.npz)", |
| "trained_on": "early-window steps (t <= 0.30 T) against eventual episode outcome; trajectory-level 70/15/15 split", |
| "liveness_check": "std of captured feature over steps must exceed 0.05 (guards against the frozen-cache hook bug)" |
| }, |
| "gate": { |
| "conformal_alpha": 0.10, |
| "conformal_threshold": "see probe_artifact.npz['conformal_threshold']", |
| "early_harm_guard_tmin": "max(0.20 * T, 2)", |
| "budget_cap_kmax": "ceil(0.05 * T) gate fires per episode, ranked by score at the boundary", |
| "action_chunk_H": 10, |
| "handoff": "at the next chunk boundary only", |
| "hold_chunks_min": 3, |
| "de_escalation": "hysteresis: return to the weak policy only after the score stays below (tau - 0.15) for a full hold window" |
| }, |
| "measured": { |
| "duty_cycle_step_weighted": 0.38, |
| "rtr_targeted": 0.101, |
| "rtr_budget_matched_blind": 0.046, |
| "rtr_random_trigger": 0.051, |
| "rtr_always_strong_ceiling": 0.319, |
| "early_window_auroc": 0.764, |
| "auroc_ci95_cluster_bootstrap": [0.70, 0.84] |
| } |
| } |
|
|
