IDU
JEV shifted one letter back, the same shift that turns IBM into HAL.
A typed-decision service: give it a state and typed questions, get typed answers with
probabilities and confidence. No text generation, so there is nothing to parse and nothing
to hallucinate. Three primitives, matching Jev's shape exactly:
| type | question | answer |
|---|---|---|
| choice | which of these options? | the option, a probability per option, confidence |
| score | where on this rubric? | a position along your levels, distribution, confidence |
| noul | is this true? | one number: the probability that it is |
Backend: convaiinnovations/laya --
421M ModernBERT-large with a decision head, Apache 2.0. Every option is scored at its own
[MASK] token and softmaxed over that question's options, and all questions in a call are
answered in one forward pass.
Honest status. The base checkpoints are near chance zero-shot on their own
typed-decisions benchmark (0.362 vs a 0.461 majority-class baseline); the card's 0.766 is a
checkpoint fine-tuned on that benchmark's training split. Laya is a fast base to specialise,
not a finished decision engine. Its shipped probabilities are already temperature-fitted
(rl_agent_config.json carries per-type and per-option-count temperatures), which is the
0.466 -> 0.081 ECE improvement the card describes — so the numbers here are the calibrated
path by default, and the temperature controls are for experimenting rather than repair.
answers
Any state, any questions -- the same request shape as Jev's API.
Temperature: 0 = keep the checkpoint's fitted value (the default and the calibrated path). Set a positive number only to override deliberately — the shipped values are already fitted, and forcing 1.0 runs the model in the uncalibrated regime.
answers
Same four questions, same three cases, through IDU and through the tuned Needle 3 adapter on dkappe/needle3-gpu. Both polarities of noul are exercised, so a 'yes to everything' bug cannot hide. The Needle side is called remotely and may be cold.
comparison