SemIf demonstrates in-browser option-score readout, but decision probabilities still need separate calibration

The OpenJev browser experiment, which entered discussion on Hacker News on September 18, currently demonstrates local decision inference under the name SemIf. It compares two paths using the same model: reading scores for the allowed options directly, or generating JSON containing probabilities token by token. The project explicitly states that it is unrelated to TypeSafe; its technical focus is eliminating the cost of generating answer text. [Discussion thread](https://news.ycombinator.com/item?id=49752041), [experiment page](https://openjev.com/)

The direct-readout path maps possible answers to labels such as A through T, retrieves their logits, and applies softmax normalization only across the options supplied by the user. This process implies that the model must still process the input context, but does not need repeated decoding steps to produce a complete JSON output. For classification and workflow routing, this offers an implementation approach worth measuring. For tasks that require extended reasoning to reach a decision, whether omitting generation affects quality still requires separate testing. [Method description](https://openjev.com/)

The execution layer uses wllama and version-pinned GGUF weights. wllama is a WebAssembly binding for llama.cpp that supports WebGPU and allows the number of layers offloaded to the GPU to be adjusted with `n_gpu_layers`. The upstream project also warns that some browser compatibility modes substantially reduce performance. Test results therefore need to specify the browser, model quantization, and hardware conditions; latency cannot be inferred from parameter count alone. [wllama documentation](https://github.com/ngxson/wllama)

The upstream project also recommends splitting large model weights into files of no more than 512 MB each to improve downloads and reduce out-of-memory issues, and suggests choosing among Q4, Q5, and Q6 quantization. These deployment requirements mean that even if an individual decision is fast, the initial download and the resources needed to keep the model in memory may still dominate the user experience. Successfully loading a small model is also insufficient evidence that it is suitable for the task. [Model preparation guide](https://github.com/ngxson/wllama)

The page displays download, loading, warm-up, input processing, and generation times separately. The two methods run sequentially, with direct readout first. This avoids simultaneous GPU contention, but the effects of caching and execution order on the comparison still need to be checked. More critically, the probabilities across options are uncalibrated and exclude answers outside the supplied choices. A value close to one cannot be treated as a guarantee that the answer is reliable. [Measurements and limitations](https://openjev.com/)

Developers integrating this kind of readout into production routing should use an independent labeled dataset to assess error rates, sensitivity to option order, and abstention thresholds, while separately testing Traditional Chinese and domain-specific terminology. At the time of this review, the implementation explanation linked from the site returned a 404, leaving insufficient information to fully reproduce its evaluation process. For now, it is best viewed as a hands-on inference experiment. It does not establish that SemIf has the same capabilities as Jev or that its confidence estimates are calibrated well enough for deployment.