Verdict (Open-jev) launch thread

Watch out

This is a third-party reproduction, not Jev, and the author is explicit about the gap: on 337 cases from TypeSafe's public evaluation set Verdict reaches 48.1% where Jev reaches 90.8% and a 26B diffusion model reaches 88.4%, which he attributes to the parameter count. All figures are self-reported, and the order-invariance comparison is not like for like: his 3.0% is across 1,000 cases, while the 27.8% figure for a frozen Qwen3.5-4B comes from another open reproduction measured on its own 36-case fixture, a caveat he raises himself. Only three of twelve public-data use cases across two checkpoints and three label framings passed the promotion gate, with passage re-ranking at 11.9%, jailbreak screening at 34.9% and phishing detection at 57.5%. The shuffled-context control pairs questions with the wrong document and reports accuracy falling from 95.0% to 22.4% with abstention rising from 21.8% to 72.7%, which argues the model conditions on context rather than memorising labels. Not documented in the capture: the training data and license, which two checkpoints he means, whether the repository ships weights or only code, the WebGPU and CPU hardware behind the latency numbers, and the demo and repository URLs, which the author says are in the first reply rather than the thread. Copy reviewed against the linked page on 2026-10-04.