Nokia open-sources AnyJev: a training-free layer that turns any open LLM into a calibrated decision model

Nokia's applied research team has open-sourced AnyJev, a Python library that turns an open LLM into a decision model without any training. L0 removes position and label bias via cyclic shifts and batch calibration, while L1 adds temperature scaling. On Qwen3-8B with BANKING77, the order-flip rate fell from 0.230 to 0.073.

• AnyJev is on PyPI under Apache-2.0 with transformers and vLLM backends
• L0 uses K cyclic shifts and batch calibration after 8 items, ~0.25 s per decision at K=20
• On Qwen3-8B with BANKING77 the order-flip rate dropped from 0.230 to 0.073
• Auto-decidable traffic at 5% error rose from 7.7% to 52.0% with L1