AnyJev L0 gets strong accuracy out of an unmodified Gemma 4 31B (87.9% vs Mouse-1's 84.9% on public tool benchmarks), but its probabilities are badly calibrated, so a 0.9 doesn't mean 90%. Mouse-1 is the pick when you act on the probability itself: it leads When2Call (85.8% vs 80.9%) and lets you fit a confidence cutoff on your own data.
AnyJev L0 (by MorrisZJ, on the AnyJev engine) is a recipe rather than a model. It runs stock gemma-4-31B-it in FP8 under vLLM, restricts the output to the label tokens, averages two option orderings and applies a running prior. Nothing is trained, so nothing is published but code.
At a glance
| Mouse-1 | AnyJev L0 | |
|---|---|---|
| Public tool benchmarks (avg of 4) | 84.9% | 87.9% |
| When2Call | 85.8% | 80.9% |
| Calibration error on the board | Not on the board | 0.24 |
| Fitted confidence cutoffs | /v1/calibrations | Not offered |
| What you run | Hosted API | Gemma 4 31B on your own vLLM |
| Training | Trained decision model | None needed |
| Prompt-injection results | Published: 0% flipped | None published |
Public tool benchmarks
Average of BFCL, API-Bank, When2Call and PhishNChips (%) · Higher is better
| AnyJev L0 | 87.9% |
|---|---|
| Mouse-1 | 84.9% |
Function calling, exact match (%) · Higher is better
| AnyJev L0 | 98.0% |
|---|---|
| Mouse-1 | 97.6% |
Next API out of 53 (%) · Higher is better
| AnyJev L0 | 85.0% |
|---|---|
| Mouse-1 | 83.7% |
Answer, call, ask or decline (%) · Higher is better
| Mouse-1 | 85.8% |
|---|---|
| AnyJev L0 | 80.9% |
Phishing or legitimate email (%) · Higher is better
| AnyJev L0 | 87.8% |
|---|---|
| Mouse-1 | 72.3% |
AnyJev leads on PhishNChips (87.8 vs 72.3), API-Bank and, narrowly, BFCL. Mouse-1 leads on When2Call by almost five points.
Why calibration decides this one
A decision model is useful in a harness because you can act on its probability: auto-approve above 0.95, ask a human between 0.5 and 0.95, block below that. That only works if the numbers mean what they say. AnyJev's expected calibration error on the board is 0.24, meaning its stated confidence is off by about 24 points on average. Most trained decision models on the same board are at 0.02–0.05.
Mouse-1 goes a step further than good default calibration. POST /v1/calibrations fits a cutoff on 200–400 of your own labelled decisions, so each answer carries an abstain flag set for the error rate you choose.
How we measured
- AnyJev's scores and calibration error are from the Jev Decision Index board as of October 20, 2026. Mouse-1 is not on the board; its benchmark scores come from our own run of the board's kit on the same rows, which does not produce a board calibration figure.
- We have not run AnyJev through our agent decision suites.
Sources
Checked October 20, 2026.
See also: Mouse-1 vs Decider 31B · Mouse-1 vs Torchcast Decision 27B · All comparisons
FAQ
What is AnyJev L0?
AnyJev L0 is a training-free inference technique, not a trained model. It serves the stock Gemma 4 31B instruction model in FP8 through vLLM and reads decisions from restricted label probabilities, averaging over two option orders with a running prior. No weights are published because none are trained.
Is AnyJev L0 more accurate than Mouse-1?
On the four public tool benchmarks of the Jev Decision Index, AnyJev L0 averages 87.9% against Mouse-1's 84.9%. Mouse-1 is ahead on When2Call: 85.8% against 80.9%.
Are AnyJev's probabilities calibrated?
Not well. Its expected calibration error on the Decision Index board is 0.24, against roughly 0.02 to 0.05 for most trained decision models there. That makes its probabilities hard to threshold on.