Jevfree repurposes EmbeddingGemma 2 for Jev-shaped questions in the browser. Instead of generating a response, it compares the situation with possible answers in a shared numerical representation. The author calls it an experiment, not a Jev replacement.

I read the full article and inspected the repository, including its decision scorer and model worker. I did not run browser inference, benchmark it or perform a security audit.

Cache answers; compare the new situation

  • Each text option includes the question instructions and its description, then gets embedded and cached.
  • The scorer subtracts the options’ shared average vector and normalizes them, emphasizing differences within that question. This is centroid subtraction, not a full statistical whitening transform.
  • It compares the situation with those vectors and uses softmax—a formula that turns scores into a distribution—to produce answer weights.
  • Boolean questions compare true and false descriptions; score questions compute a weighted average over rubric levels.

“Probability” is not measured reliability

  • The inspected code sets a default softmax temperature of 0.05; changing this setting changes how decisive the distribution looks, without demonstrating better decisions.
  • Its confidence combines the gap between the two leading options with how concentrated the distribution is. It is not an empirical estimate that the answer is correct.
  • The article’s dinner example selects takeout and says the meal cannot be prepared quickly. That illustrates an additional limitation: questions are scored against the original situation, not conditioned on earlier selected answers.
  • Typed JSON is useful plumbing, not proof of sound reasoning. The author explicitly warns about negation, numbers and tasks requiring reasoning.

Local inference has a first-load cost

  • The worker selects WebGPU when available and otherwise WASM; without half-precision support it falls back to quantized weights.
  • The author reports an approximately 1.5 GB initial download for the default half-precision model. That size and real-device performance were not reproduced here.
  • The implementation uses an embedding model, not TypeSafe’s hosted Jev. Its parser accepts Jev-style requests, but noul still requires a condition, and responses use results, not the hosted API’s answers envelope.

Use the demo as a local similarity-routing baseline or teaching tool. Before using it for consequential routing, test real labels, negation, absent answers and confidence thresholds; do not inherit Jev’s behavioral guarantees merely because the request looks familiar.