Fix hook-example variable consistency (weak_policy, matching the loop)

HARD-tercile B-C precision: significant by bootstrap, n.s. by McNemar (p=0.064)

Warm-luxury grade of banner (no synthwave neon)

Fix HELM mechanism (within-weak rollback, not strong-policy-from-first-flag); consistent with the space

AUROC attribution: distinguish 0.738 pilot / 0.764 confirmatory vs 0.63 complement on matched pilot set

Restore GR00T N1.7 generalization result (RTR 0.155, paper Limitations/Results) to the key-results + limitations; correct the pre-registration deviations pointer to the paper's actual four (H 50->10, suite, probe-hook fix, two-stage)

Declare datasets: Kaikaku/aegis-rollouts; drop misleading library_name:pytorch (artifact is numpy npz); clarify AUROC 0.764/n=2792 is the paper's confirmatory vs the shipped n=112 pilot 0.738; add pre-registration deviations pointer; Kaikaku casing

Re-render architecture diagram in warm-luxury palette (ivory weak / bronze strong / gold probe+gate / ember fire) on warm-black

Re-render headline bars in warm-luxury palette (AEGIS aged-brass, controls bronze, ceiling taupe, ivory labels, gold ~2x annotation) on warm-black

Critical pass: warm-luxury figures (re-rendered dark to match the Space/poster, retiring light-mode + teal/cyan); fix probe-feeding bug in usage loop (weak pass must keep running to enable hand-back); remove unverified GR00T 15.5% (duplicate of HELM, no GR00T conditional RTR in the data); clarify chunk pooling

Banner: launch key visual (the touch)

Refresh banner

AEGIS model card: frozen probe head, gate config, paper figures

initial commit