AnyJev Makes Open LLMs Typed Deciders

AnyJev targets a practical weakness in raw LLM classification: token probabilities often reflect formatting biases as much as the underlying decision.

• –L0 reduces position and label bias without changing model weights or requiring supervised data.
• –The approach can raise automation confidence while preserving the flexibility of open-weight models.
• –Rotation-based inference costs more compute, scaling with the number of options, though adaptive stopping can reduce the overhead.
• –Its probabilities are more useful than raw logits, but they should not be treated as fully calibrated certainty without additional validation or labeled calibration.
• –The library is especially relevant for agent routing, moderation, triage, and tool-use decisions where structured outputs matter.