RLCD vs RLHF: How Does Typesafe AI Jev Work

Most language models today are trained to be helpful conversationalists. When you ask a question, they write a polished paragraph, and if you need a decision for your software, you have to parse that paragraph back into something your code can act on. Typesafe AI's Jev takes a different route. It skips the paragraph entirely and returns a structured decision with a probability attached, which makes it worth understanding how its training differs from the standard approach.

What RLHF does

Reinforcement Learning from Human Feedback, or RLHF, is the method behind today's chat models. The process starts with supervised fine-tuning on human-written demonstrations, so the model learns the general shape of a good answer. Human raters then compare pairs of model responses and pick the one they prefer, and those preferences train a separate reward model. Finally, the language model is optimized against that reward model, so it gradually produces answers people tend to rank higher.

This works well when a person is reading the output. RLHF models have good tone, handle nuance, and know how to refuse gracefully, because they were trained on what people actually liked. The weakness shows up when software, not a person, consumes the output. A model trained to sound helpful will sound confident even when it is wrong, and its answer still arrives as free text that needs parsing, validation, and error handling before your code can trust it.

What RLCD does

Reinforcement Learning for Calibrated Decisions, or RLCD, is Typesafe's training method for Jev. It keeps the reinforcement learning loop but replaces human preference with a different goal: the model's stated confidence should match how often it is actually right. If Jev says it is 70 percent confident across a large set of decisions, roughly 70 out of every 100 of those calls should turn out correct.

That shift changes everything downstream. Instead of learning which phrasing raters prefer, Jev learns to pair each decision with an honest probability. There is no reward for eloquence or style, because there is no prose to judge. The training rewards getting the decision right and penalizes claiming more certainty than the evidence supports, which gives engineers something they can build automation around: a threshold they can test on their own labeled data and deploy with known trade-offs.

How Jev works in practice

Jev is not a chatbot. You send it unstructured state such as a support ticket, a log line, or a JSON object, along with one or more typed questions you define in advance. Each question is a choice between options you specify, a score against levels you describe, or a yes-or-no judgment, and Jev answers all of them in a single pass.

Because every question shares one reading of the same state, adding more questions barely changes latency. Answers come back as typed values your code can branch on directly, each with a probability distribution. There is nothing to parse with regular expressions and no JSON schema repair step, which removes a whole class of production failures that teams deal with when wrapping chat models in decision logic.

Why Jev is fast and cheap

The speed comes from what Jev does not do. A standard language model generates its answer one token at a time, with each token conditioned on everything before it, which is why complex structured extractions can take seconds. Jev emits all fields at once through a parallel sampler, so typical calls complete in roughly 70 to 500 milliseconds.

The cost follows the same logic. Typesafe charges $0.042 per million input tokens and nothing for output, because there are almost no output tokens to bill. On equivalent structured tasks, the company reports speedups in the range of 20 to 200 times and cost reductions of 40 to 400 times against frontier models, with small independent tests pointing in the same direction even if they do not reproduce the peak figures.

A note on the other RLCD

Readers who follow alignment research may recognize the acronym. A 2023 paper by Yang and colleagues introduced a method also called RLCD, standing for Reinforcement Learning from Contrastive Distillation, which generates preference pairs from contrastive prompts to simulate RLHF data without human labelers. Typesafe's RLCD stands for Reinforcement Learning for Calibrated Decisions and pursues a different objective entirely. The two share the reinforcement learning machinery but solve different problems, so the 2023 paper should not be read as documentation for Jev.

Which approach fits your problem

The choice comes down to who consumes the output. When a person reads the response and judges its quality, as with copilots, coding assistants, and customer-facing chat, RLHF models remain the right tool because they were trained for exactly that judgment. When software acts without a human in the loop, as with routing, classification, approval, and verification at scale, a calibrated decision model gives you testable thresholds instead of prose to interpret.

Many production stacks will end up using both. Jev can sit at the decision points, triaging and scoring at high volume and low cost, while a frontier model handles the moments that genuinely need generated text or code. For teams embedding agents inside SaaS products, that division maps cleanly: preference-tuned models for the human-facing conversation, and a calibrated model at the workflow branch points where code needs a fast, cheap signal it can trust.

Sources

• Typesafe AI, Introducing System One Models and Jev, September 15, 2026.
• InfoWorld, TypeSafe AI's new models work with machines, not humans, September 17, 2026.
• The Register, TypeSafe AI debuts model for machines that plays Doom, September 16, 2026.
• OrcaRouter, Jev: TypeSafe's Decision Model, Speed and Cost Explained, September 16, 2026.
• Progressive Robot, Jev model is the first product from TypeSafe AI, September 16, 2026.
• LangChain Blog, Building a Harness with Jev, September 17, 2026.
• Yang et al., RLCD: Reinforcement Learning from Contrastive Distillation, arXiv 2307.12950, 2023.
• Lee et al., RLAIF vs. RLHF, arXiv 2309.00267.
• Chilson, Typesafe AI's Jev Is Not an LLM and That May Be the Point, Forkast, September 17, 2026.
• Mindstudio, RLCD vs RLHF: What Is Typesafe's Jev Model Actually, September 16, 2026.