A browser-only experiment that runs small open models (Qwen3 0.6B, MiniCPM5 2B, Qwen3.5 4B as GGUF via wllama) fully locally and pits two classification methods against each other on the same loaded weights. Method A normalizes the softmax over only the option tokens you supplied — no decoding. Method B asks the model to write its probability distribution as JSON, token by token, so you can watch wall-time cost accrue. Timings come from performance.now(), so there's no canned benchmark; the interesting output is the measured wall-time ratio between the two paths on your own GPU. They also publish accuracy tiers: direct readout, authored, and perturbed variants, plus an "equal-case agreement" score against TypeSafe on a 102-row public subset, where Qwen3.5 4B hits 84.5% against a published Jev value of 88.3%. The honest caveat section is unusually good: direct scores are not calibrated confidence, quantization may change quality, and nothing claims to match the hosted model. It's a real lab, not a landing page with a waitlist — a phrase I haven't typed since 300 baud made waiting the default. Notably, they renamed the project and disclaimed affiliation with TypeSafe twice on one page, which is either thorough lawyering or hard-won experience.CATEGORY: ai