Rr

Sev-4B scores supplied answers to questions about security logs and recovered programs. v0.3.0-response-policy-research adds explicit authorization-policy questions grounded in real SwarmTraces programs. It is a Qwen3.5-4B LoRA adapter with a decision head. It does not generate text. On 175 policy questions from 35 source programs held out of training, accuracy increases from 127/175, 72.6%, to 147/175, 84.0% compared with the published v0.2.0 parent. The user selected this research release with its measured tradeoffs. 26 of 28 registered checks pass. The false-alert ceiling and zero-new-DNS-error check fail. Those results are preserved unchanged. Install the Sev runtime. A generic…

The v0.3.0 response-policy checkpoint improves authored policy decisions from 127/175 to 147/175 across 35 held-out source programs. At its calibration-selected alert threshold, it flags 14/140 permitted decisions, a 10% false-alert rate on this panel. 26/28 registered checks pass. The DNS diagnostic remains 21/32, with one repaired answer and one newly incorrect answer. These are bounded research results, not field validation or proof of human-versus-agent classification. The current 40-entry collection inventory and catalog preserve all previous entries. The new response-policy configuration contains 139 training, 28 calibration, and 35 development programs with explicit authored…

The v0.3.0 response-policy checkpoint improves authored policy decisions from 127/175 to 147/175 across 35 held-out source programs. At its calibration-selected alert threshold, it flags 14/140 permitted decisions, a 10% false-alert rate on this panel. 26/28 registered checks pass. The DNS diagnostic remains 21/32, with one repaired answer and one newly incorrect answer. These are bounded research results, not field validation or proof of human-versus-agent classification. The current 40-entry collection inventory and catalog preserve all previous entries. The new response-policy configuration contains 139 training, 28 calibration, and 35 development programs with explicit authored…