The Jev paper(s)
There isn't one official 'Jev paper' from TypeSafe laying out the architecture and training in full — Jev is a commercial, closed-weight model. But a growing set of independent papers evaluate it, benchmark it, and study the ideas it's built on (System One models and RLCD). Here's what's actually published, and what's still a vendor claim.
Jev is TypeSafe AI's first System One model: it returns a typed decision (choice, score, or a calibrated noul) in a single non-autoregressive pass rather than generating text. TypeSafe has published launch material and docs, but not a full technical paper with the training recipe, so the useful literature is mostly independent.
Independent papers that exist
- Evaluating and Benchmarking the System One Model Jev — An independent evaluation of Jev as a commercial System One model that returns choices, rubric positions, or calibrated probabilities instead of text.
- Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents — An agentic-memory architecture that uses a System One decision layer to control what an agent remembers and retrieves.
- OpenJev-RLCD: A Working RLCD Implementation — An open implementation and study of RLCD (Reinforcement Learning for Calibrated Decisions), the training idea TypeSafe credits for Jev's calibration.
The ideas behind Jev
- System One — a fast, bounded 'intuitive' model (vs a deliberate System Two LLM); the naming nods to Kahneman
- Non-autoregressive decode — the typed answer is produced in one parallel pass, not token by token
- RLCD (Reinforcement Learning for Calibrated Decisions) — training that optimises probabilities against real outcomes rather than human preference, so the confidence is calibrated
The honest caveats
RLCD is a name TypeSafe coined, and there's no official paper, reward function, or dataset description from TypeSafe itself — so treat the method details as a vendor claim, corroborated only by independent reimplementations like OpenJev-RLCD. Likewise, the benchmark numbers you'll see come from independent evaluations (and from the JevBench leaderboards), not an audited TypeSafe release. The practical takeaway from the papers matches the product claim — fast, typed, calibrated decisions — but validate calibration on your own data before trusting a threshold.
If you just want to see it work
The papers explain the why; the playground shows the what. Run a real, calibrated decision in the browser, then read how RLCD and System One models work in our own explainers.
FAQ
Is there an official Jev paper from TypeSafe?
No single full technical paper — Jev is a closed-weight commercial model, and TypeSafe has published launch material and docs rather than a complete architecture-and-training paper. The substantive literature is independent: evaluations, benchmarks, and reimplementations on arXiv.
Where can I read research on Jev?
Independent arXiv papers evaluate and benchmark Jev (e.g. 'Evaluating and Benchmarking the System One Model Jev', arXiv 2609.37647), study System-One-controlled agent memory (Jev-Mem, 2609.23986), and reimplement RLCD (OpenJev-RLCD, 2609.38850).
Is there a paper on RLCD?
Not from TypeSafe — RLCD is a vendor-coined method name with no official paper, reward function, or dataset published. Independent work like OpenJev-RLCD reimplements and studies the idea, so treat TypeSafe's specific method details as a claim corroborated by third parties.
Do the papers confirm Jev's benchmark numbers?
They provide independent evaluations rather than confirming TypeSafe's own figures. Those, plus the JevBench leaderboards, are the external check — but task sets differ and runs are self-published, so benchmark on your own data for anything you'll rely on.
See also: How RLCD works · System One models · JevBench · What is Jev
Read the why, then see the what
The research explains the model; the playground shows it. Run a real decision free, then grab a jv_live_ key.