A typed readout for open models
You define a state and one or more bounded questions. AnyJev returns distributions for categorical choices, a probability for yes/no questions, or an expected value over an ordered score.
Independent project guide · Reviewed September 25, 2026
AnyJev turns an open LLM into a typed decision model.
AnyJev is an open-source Python toolkit from Nokia Applied Research. It reads Choice, Score and Noul-style decisions from causal language models, then adds controls for option-order bias and probability calibration.
Important: AnyJev is not a new foundation model, an open release of TypeSafe's Jev weights, or an official Jev product. It is an independent way to build Jev-shaped decisions on models you run.
- Maintainer
- Nokia Applied Research
- License
- Apache-2.0
- Package
anyjevon PyPI- Status
- Pre-alpha
AnyJev · L0
state → typed decisionsstate
I was charged twice. Please fix this today.
choice · routeWhich team should handle this?
result.level = "L0"Illustrative outputThe useful distinction
AnyJev is a toolkit, not another set of model weights.
It keeps the causal language model you choose and changes how a bounded decision is read from it. That makes it closer to a decision layer than to a standalone model such as a fine-tuned classifier.
Bias correction and calibration levels
Raw answer-token logits can favor a label or its position. AnyJev exposes which correction level produced each decision so application code can require the level it expects.
An open copy of TypeSafe Jev
The project does not publish or reproduce TypeSafe's private weights, training data or RLCD method. Similar request shapes do not establish equivalent behavior.
Raw → L0 → L1 → L2
Four levels make the evidence carried by a decision explicit.
“No training” accurately describes L0. The higher levels add labeled examples, and L2 fits a small question-specific head without updating the base model's weights.
RawNo labels
Reads a restricted softmax over answer-token logits.
No correction for option position or label preference.
L0No labels
Rotates options and corrects an estimated label prior.
More stable, but its probabilities are not task-calibrated.
L1100–500 labels
Adds temperature calibration for one model and question.
Improves confidence estimates; it does not fix wrong rankings.
L2100–300 labels
Fits a small closed-form head on an intermediate hidden state.
Specific to one model and question; needs local hidden-state access.
Local workflow example
Route a support ticket and check a proposed action.
This small example mirrors the project's public API: a Qwen model stays local, AnyJev reads two bounded decisions, and ordinary code decides what happens next.
Install the Transformers backend
pip install "anyjev[hf]"Python 3.10 or newer is required. The first real run also downloads the open model you select.
Python
Choice + Noulfrom anyjev import Decider, Question
from anyjev.backends.hf import HFBackend
decider = Decider(HFBackend("Qwen/Qwen3-8B"))
route = Question.choice(
"Which team should handle this?",
["billing", "technical", "sales", "other"],
name="route",
)
risky = Question.noul(
"Is this tool call destructive or irreversible?",
name="risky",
)
result = decider.decide(
{"message": "I was charged twice.", "tool_call": "refund"},
[route, risky],
)
result["route"].distribution
result["risky"].p_true
result.level # "L0"Define the answer space
The application names the allowed teams and asks whether the proposed tool call is risky.
Read bounded distributions
No free-form answer is generated. The result carries a distribution or probability and the level used.
Keep policy in code
Your code sets thresholds, sends uncertain cases to review and prevents side effects when the risk is too high.
Upstream benchmark
The headline result is about safer automation, not only accuracy.
AnyJev reports the following Qwen3-8B results on 300 BANKING77 test items. These figures are author-reported from committed benchmark artifacts; Jev AI Dev has not independently rerun the model benchmark.
| Metric | Raw logits | AnyJev L0 | AnyJev L1 |
|---|---|---|---|
| Answer flips after reversing options | 0.230 | 0.073 | 0.077 |
| Accuracy | 0.747 | 0.803 | 0.807 |
| Expected calibration error | 0.240 | 0.184 | 0.095 |
| Coverage at no more than 5% error | 7.7% | 46.3% | 52.0% |
Accuracy moved from 0.747 to 0.807 in this test. The larger operational change was reported coverage at no more than 5% error: 7.7% for raw logits and 52.0% after L1 calibration. That estimate is based on a small test set and should not be transferred to another workflow.
Inspect the benchmark and ablationsCommunity video · 29 seconds
Watch a quick explanation of turning an LLM into a Jev-style decider.
Data Science in your pocket published this short introduction after AnyJev appeared. It is a third-party explanation, not an official Nokia Applied Research demo or an independent benchmark.
Watch “How to convert any LLM into Jev AI?” on YouTubeCommunity field test · September 2026
A 20-ticket test surfaced the accuracy–latency tradeoff.
WquGuru compared official Jev and three open projects on one MacBook using 20 Chinese support tickets. The thread includes video, timing and per-project accuracy.
AnyJev with Qwen3-4B reportedly answered 19 of 20 tickets correctly.
Accuracy95%19 / 20 tickets
Latency~4sreported per test
Workload20Chinese support tickets
The same thread reports 20/20 for official Jev, 65% for Laya and 35% for djev. Treat this as an early community observation—not a controlled benchmark: it uses one small question set, one machine and project-specific model configurations.
View the comparison and video on X“The latency is real. We went correctness first.”
Jiamu “Morris” Zhang said the first release had not tuned its serving path. He framed AnyJev as a method for models a team already serves and said vLLM/SGLang optimization work was starting.
Read the author's full responseA concise tour of the project and its committed benchmark.
MarkTechPost summarizes the typed readout, L0/L1 distinction and BANKING77 results. Those figures originate from the project's own benchmark rather than a separate replication.
Evidence boundary: Jev AI Dev verified that the posts contain these statements, but did not reproduce the community test. View and engagement counts are intentionally omitted because they change over time.
Before production
Know what the current release does not solve.
Typed output removes one integration problem. It does not make a semantic judgment correct, safe or transferable to a different distribution.
L0 spends more compute on wide choices
A choice with K options uses cyclic rotations, so the zero-label correction needs K prefills. It trades inference work for better order stability.
L1 and L2 need workflow labels
Calibration and heads belong to a particular model and question. They should be rebuilt or revalidated when either changes.
L2 needs hidden-state access
The current L2 path uses a local Transformers backend. Served engines such as vLLM and SGLang are listed on the project roadmap.
Published numbers are not your acceptance test
Measure accuracy, calibration, order sensitivity and abstention on representative data before a probability controls an important action.
Primary sources
Check the project at the source.
AnyJev is moving quickly. Verify the latest release, backends and limitations before you build against it.
Explore the ecosystem
Compare AnyJev with other community System One projects.
AnyJev is one approach: adapt an open causal LLM at decision time. The wider ecosystem also includes purpose-trained models, local runtimes and API-compatible experiments.