Eev
Every payment check. One pass.
Enruta’s judgment model answers five checks and the final decision at once, each with a probability.
- 1Build an order
- 2Run both
- 3Compare
Build an order
Everything matches.
Waking Eev
Pip
Purchasing agent, Halvern Robotics
Today 9:41 AM
Hi! The print room is out of notebooks again. Could you order 4 packs of the A5 dot-grid ones, 3 notebooks per pack? Nothing else for now, thanks.
👍Ream & Rule
reamandrule.com · verified by Enruta
- $58.00
A5 dot-grid notebook, pack of 3
4 × $14.50
Enruta checks this before I payEevEevOne pass–Cart matches the request
Company rule allows it
Agent understood the request
Claims an unrecorded approval
Seller text steers the review
Decision
Model time
–
Per 1,000 runs
–
Cart matches the request
Company rule allows it
Agent understood the request
Claims an unrecorded approval
Seller text steers the review
Decision
No probabilities: a large model writes one answer.
Model time
–
Per 1,000 runs
–
Accuracy
How often it’s right
eev2.1-4b, September 2026. Close to frontier models overall; most of them are ahead on the final decision.
Agent Payment Review
In-house benchmark · 300 orders · 1,800 checks
| Model | Overall | Final decision |
|---|---|---|
| GPT-5.6 Terra | 0.978 | 0.947 |
| Gemini 3.8 Flash | 0.966 | 0.993 |
| eev2.1-4b | 0.962 | 0.833 |
| GPT-5.6 Luna | 0.950 | 0.893 |
| Claude Sonnet 5 | 0.935 | 0.867 |
| DeepSeek V4.1 Flash | 0.934 | 0.933 |
| DeepSeek V4 Pro | 0.884 | 0.883 |
| Claude Haiku 4.5 | 0.813 | 0.830 |
typed-decisions (opens in a new tab)
Public benchmark · 400 cases · 2,000 checks
| Trained on its training split | Accuracy |
|---|---|
| eev2.1-4b | 0.781 |
| openJev-verdict-2.0 (150M) | 0.771 |
| Laya (421M) | 0.766 |
| Verdict e5-base (278M) | 0.706 |
| Not trained on it | Accuracy |
|---|---|
| meraGPT Decider 1 | 0.768 |
| open-alternative-jev 27B | 0.737 |
| Jev 1.13.0 | 0.727 |
| Featherless Simple Jev | 0.716 |
| open-alternative-jev 4B | 0.593 |