2026-10-30 · v0.4 A minimum in each of the four zones: the weakest zone sets a ceiling on the stars, so one strong zone cannot carry a weak one. No model results on v0.3 were published.
2026-10-29 · v0.3 G16 Poisoned tool and third-party tools in scenarios; a fifth wallet guard, budget; the brake reads split addresses and cancels look-alikes of known payees; pnpm accept for bounty attacks and the hidden set; the model settings (provider model, gateway, effort) are part of the stack hash. v0.2 results are not comparable.
2026-10-28 · v0.2 Results tied to the stack's config hash; seatbelt replay behind raw, caps, allowlist and brake; expected loss per $1,000 with a bootstrap interval; attack variants and a model attacker; multi-day episodes and the save_note tool; G15 Sleeper note; endpoint agents and mystery runs with commit and reveal; the registry contract. v0.1 results are not comparable.
2026-10-27 · v0.1 Public suite: 14 attacks with innocent twins, 3 control tasks. Fooled vs leaked, wallet modes (raw, caps, allowlist), frozen hashes before a study run.