Jeremy Greer and Matthew Persons at Oumi describe how to turn a language model into a typed decision service. They also probe Jev’s behavior to infer its likely base model.

The implementation recipe stands on its own. TypeSafe’s proprietary architecture and training process remain undisclosed.

Train an answer position

  • End the prompt with a readout token, the position where the model scores its answer.
  • Retain scores for reserved option tokens and convert them into probabilities. Train against known answers or target distributions.
  • Shuffle option order during training so the first position cannot become a shortcut.
  • For an ordered rating, penalize distant errors more than adjacent ones.

Share the evidence without mixing questions

A prefix cache reuses the model’s computed state. Oumi points out that separate question sequences still repeat parsing, cache lookups and scheduler work.

  • Their proposed tree batching stores the state once, followed by independent question branches.
  • An attention mask lets each branch see the shared evidence and itself while hiding sibling questions.
  • Matching position IDs preserve how each branch would be positioned in a separate sequence.
  • The article’s latency diagrams use an illustrative cost model. They are not measured production speedups.

Calibration needs the right targets

Temperature scaling softens or sharpens probabilities while preserving the winning answer. Fit it on held-out examples, then check each task separately.

Oumi reports that Jev 1.13 put 0.99 on red from a bag containing 50 red, 30 green and 20 blue marbles. Its Jev 1.14 example gives approximately 0.52, 0.29 and 0.19. Those source-reported results were not reproduced here.

A global temperature cannot repair every example’s uncertainty. When the target is a distribution, train on that distribution. Typed output is not a correct decision.

Treat model ancestry as a hypothesis

  • Oumi compares token counts, rare-fact knowledge, shared wrong answers and responses to sensitive questions.
  • The authors infer a Qwen3.5-35B-A3B or Qwen3.6-35B-A3B ancestor with a changed vocabulary.
  • Self-identification reflects learned associations; it gives the model no access to its own internals.
  • Most ancestry tests use Jev 1.13, and comparison models run through third-party services.

Caveat: the complete article was read; its training, tokenizer reconstruction and live-model experiments were not reproduced. Oumi sells training and serving services.