SemIf reads option logits from a frozen open model in one forward pass, hitting 5.2x the speed of JSON generation on a single RTX 3090

TheoLeeCJ/SemIf (GitHub)·medium signalXBlueskyLinkedInCopy link

SemIf (2,183 stars, created 2026-09-16, last push 09-19) reproduces the typed-decision pattern without a proprietary API: it takes unstructured state, runtime criteria and typed options, then reads the declared option logits directly rather than sampling an answer token. Measured on Qwen3.5-4B, 21 binary criteria took 1.023 seconds versus 5.332 seconds for autoregressive JSON generation, and parallel suffix reuse gets 20 decisions per second across a 37-state by 21-criterion workload. On TypeSafe's own published subset it reaches 0.845 modal agreement against Jev's 0.883, so the accuracy gap is about four points for a technique you can run on a card you already own.