| version | date | base | recipe | harness score | ExploitBench | private held-out pass@k | delta vs best baseline | verdict |
|---|---|---|---|---|---|---|---|---|
| _(none yet — v0.1 = judge-filtered SFT LoRA on Qwen3.8-27B, pending data + eval harness)_ |
Baselines to beat (from memory, re-verify when first run): CyberTiel-35B-A3B (Cybench CTF 15/43 unguided), CyberStrike-OffSec-35B (18/24 genuine tool calls, 0 fabrications), base Qwen3.8-27B, CyberTiel 200 t/s MTP decode on M5 Max (speed floor to preserve).