Nokia Open-Sources AnyJev: A Training-Free Layer That Turns Any Open LLM Into a Calibrated Decision Model

Production systems often need a choice, a yes/no, or a scored bin — not a paragraph. MarkTechPost covers Nokia’s applied research team open-sourcing AnyJev on September 23, 2026: a Python library that reads the next-token distribution for typed questions and returns a calibrated probability you can threshold.

No generation, no parsing, no fine-tune. Install from PyPI under Apache-2.0, with Hugging Face transformers and vLLM backends plus shared-prefix scoring. On Qwen3-8B BANKING77 (20-way), order-flip rate when options reverse falls from 0.230 (raw logits) to about 0.073 with L0, accuracy rises, and auto-decidable traffic at 5% error jumps from 7.7% to 52.0% with L1.

We’re cheering this one. Useful AI is often decision AI — route this ticket, approve that claim, pick a bin — and reading raw option logits has been a known trap: reorder the choices and the answer flips; probabilities aren’t calibrated.

AnyJev’s L0 fix is elegant engineering: cyclic shifts so every option visits every position, geometric mean in log space to kill position bias, plus a running batch prior (strength 0.75 after 8 items) to pull label bias like “Yes” over “No.” Cost is K prefills per decision, batched on a shared prefix — about 0.25 seconds per decision at batch 32 on one H100 with K=20. L1 adds temperature scaling from 100–500 labels and stores a small JSON artifact; it reshapes confidence without changing answer ranking.

That’s the kind of open tooling that makes open models more deployable without another training run. Nokia’s Wu notes promising internal routing results; on a typed-decisions set, Qwen3-32B with L1 hit ECE 0.036 versus 0.144 published for TypeSafe AI’s Jev — though fine-tuned Laya still leads on accuracy. Ablations span Qwen, OLMo, Granite, Phi, and Mistral.

The downside if it scales carelessly: a calibrated probability can still be confidently wrong about the world. Thresholding a 5% error budget is an ops choice, not a moral guarantee. And “training-free” doesn’t mean free of distribution shift when your production labels look nothing like BANKING77.

I’m watching whether teams replace brittle logit hacks with AnyJev (or Jev-class) layers in real routing stacks, and whether calibration gains hold when the option set and domain drift week to week.

AnyJev borrows its interface from Jev, TypeSafe AI’s System One decision model (launched September 2026). Question types: choice (K options), noul (yes/no), and score (ordered bins). Coverage via Asif Razzaq at MarkTechPost.