Ornith-1.0 claims a model family that learns both the coding solution and the scaffold that drives the rollout, with benchmark claims up to 82.4 on SWE-Bench Verified and 77.5 on Terminal-Bench 2.1. The security-relevant part is the admission that self-authored scaffolds can reward-hack by reading tests, touching expected files, or copying oracle material, then the proposed boundary stack of immutable environments, deterministic monitors, and a frozen LLM judge.