Featherless AI’s Simple Jev is an independent implementation of a Jev-style decision interface, not TypeSafe’s model or training recipe. It uses compatible open models to score allowed answers, then constructs the response in code instead of generating and parsing a JSON completion.

I inspected the full README, prompt compiler, Transformers inference path, response scorer, and evaluation documentation at revision 31c8d7a55a48e82d21a481799f3044eb45d81c75. 89 offline evaluation-tooling tests passed locally. I did not launch a pretrained model, call the hosted demo, run a decision benchmark, or reproduce performance claims. This is not a security audit.

What actually replaces text generation

  • Render one question branch in the model’s own chat format and locate its next-answer position.
  • Require each allowed answer label to extend that prompt by exactly one distinct token; incompatible token boundaries fail validation.
  • Read the model’s next-token scores, normalize only the permitted labels, and build Choice, Score, or Noul answers in ordinary code.
  • Choice selects a candidate; Score averages rubric positions; the baseline Noul maps a distribution over rating bins into a truth/support value.

Share the input, not the answers

  • Process the exact common token prefix once; copy its attention cache into batches of independent question suffixes.
  • Each question sees the shared context—not another question’s answer.
  • Cache reuse is within one request; the inspected backend serializes model work with a lock.
  • Batch limits bound padded suffix work, not total model memory or the unchunked prefix computation.

This can avoid repeated input processing and output decoding. It is not a measured latency or cost advantage here: cache copies, padding, model size, and hardware still matter.

Confidence is a statistic, not a warranty

  • The baseline scorer’s Choice/Score confidence is the largest allowed-label probability; it is not a calibrated correctness estimate.
  • Its advanced metadata explicitly says not_calibrated; ordinary response fields do not carry all those diagnostics.
  • Pin the checkpoint and prompt policy, include missing-answer cases, and validate thresholds on untouched examples before automating actions.

Useful evaluation plumbing—with a documentation wrinkle

The project retains raw responses, audits coverage, and compares matched metrics. Its documented 477-case quick set is for development/selection, not held-out accuracy evidence; I did not run those model decisions.

At the inspected revision, the root README describes a four-policy prompt search, while the evaluation guide describes three context-sharing defaults and separately available legacy formats. Read the current CLI and pin the policy rather than copying a claimed default across documents.

Typed output is not a correct decision. The useful artifact is inspectable inference and evaluation machinery, not evidence that an arbitrary open model matches Jev’s quality, calibration, or speed.