“What should the agent do next?”
Choose between checking the order, searching the help center, or escalating to a human. Ateve Jev returns a probability for each option and the most likely choice, guiding your agent’s next action.
// ATEVE JEV · DECISION MODEL
v1Better decisions. Smarter actions.
Ateve Jev evaluates context, makes a choice, and returns probability estimates.
Help your AI agents make better choices.
“Where is my refund?” “Where is my refund? I sent the jacket back two weeks ago.” Pick the next step for this request.
“Is the returned item unused”
Check whether the returned item is unused based on details such as “Never worn. Tags are still on.” Ateve Jev returns the probability of yes, helping your agent decide whether further review is needed.
“How helpful is this draft reply?”
Rate a draft such as “Your refund is being processed” for clarity and useful next steps. Ateve Jev returns probabilities for each rating level and an expected score, helping your agent decide whether to revise.
// BENCHMARKS
Performance across decision tasks.
Ateve Jev goes beyond workflow automation to verify content, assess code, and rank search results. Launch evaluations cover multiple domains and languages.
REPORTED LAUNCH RESULTS■ ateve-jev-v1
| # | TASK | BENCHMARK | RESULT |
|---|---|---|---|
| 01 | General decisions | Public Test Set | 87.92 |
| 02 | Choice / Boolean / Rating | Choice / Noul / Score¹ | 89.99% / 88.98% / 71.51% |
| 03 | Intent classification | MASSIVE | 89.64% |
| 04 | Scenario classification | MASSIVE | 92.97% |
| 05 | Banking intent classification | Banking77 | 90.83% |
| 06 | Factual consistency | VitaminC | 90.54% |
| 07 | Chinese NLI | OCNLI | 84.00% |
| 08 | Adversarial inference | ANLI · R1 / R2 / R3 | 78.10% / 64.40% / 60.08% |
| 09 | Structured decisions | Nimble | 87.04% |
| 10 | Code judgment | CodeJudgeBench | 58.58% |
| 11 | Code judgment | Themis / CodeCriticBench | 82.86% / 62.12% |
| 12 | Retrieval reranking | SciFact · nDCG@10 / MRR@10² | 0.7964 / 0.7701 |
¹ Score: most likely level. ² SciFact: same candidates; higher is better.
// USE CASES
Where Ateve Jev gets to work
Explore how Ateve Jev routes tasks, keeps context relevant, and chooses the next action in browsers.
// MODEL ROUTER
The right model. Chosen automatically.
From a typo fix to an API migration, different tasks need different models. Ateve Jev returns a choice and probabilities for your model options. Your application uses that judgement to route the request and let the selected model handle the task.
// CONTEXT COMPACTION
Fast compaction. Keep your agent moving.
Ateve Jev scores past tool calls and results in batches, helping the application trim stale context without generating a new summary. Retained content stays verbatim, and the agent continues with a lighter history.
// BROWSER USE
Fast decisions. Fluid browser actions.
From entering cities to displaying matching flights, every step needs a timely decision. Ateve Jev selects the next action from the current page state, keeping the workflow moving as the page changes.
// WHY RESPONSE TIME MATTERS - GAME DEMOS
Timely decisions.
Actions that keep up.
A falling block or an approaching fireball leaves little time to choose. These demos make the value of a timely response visible.
Tetri Tower
Choose before the piece lands.
SINGLE GAME316lines in one game
Jev selects how to rotate and place each piece on the current board; Ateve harness carries out the move. Timely decisions leave room to adjust before the piece locks into place.
Note: The Tetris benchmark was conducted using Ateve Jev within the Ateve Harness framework.
Super Mario Bros
Catch the moment to jump or dodge
WORLD 1-1 → 1-44stages cleared consecutively
Gaps and fireballs change the next move. Jev selects movement and jumps from the current game state, and Ateve Harness executes them. A quick response lets the action happen while the opportunity is still there.
Note: The Super Mario benchmark was conducted using Ateve Jev within the Ateve Harness framework.
// PRODUCTS
All of our products
Search is live today. The rest of the stack is on its way — same key, same base URL, same structured contract.
Search API
Context-aware search that understands query intent and retrieves the most relevant evidence.
Try Search API for freeFetch API
Turn unstructured content from web pages, documents, and PDFs into clean, structured data your applications can use.
Coming soon
Decision model
Ateve Jev evaluates context and estimates probabilities to help your AI agents make better choices and take smarter actions.
Coming soon// FROM INFORMATION TO ACTION
From knowing to doing
Ateve connects AI to the real world through search and source verification. Jev extends that mission into decisions your software can act on.
Try Ateve Jev for free