TL;DR
A new write-up on the typed-decision AI wave put two names side by side: Jev, the model from TypeSafe AI that just raised 870 million dollars at a 7.5 billion dollar valuation (led by a16z), and VIDRAFT's Darwin-ZTC, described as the open-weights alternative. The thesis both share: agents should not ask a chat model to decide. We have been building exactly that, in the open, with a zero-token method.
What the article says
The piece frames the shift cleanly. Jev, built by a former OpenAI researcher, does not generate text for routing and classification. It returns structured decisions: a choice from fixed options, a score on an ordered scale, or a probability for a yes or no statement. A safety node can then do something like if (safety >= 0.9) execute; with no text to parse. OpenAI reportedly followed with a Decisions API within two weeks.
Then it names the open side: VIDRAFT's Darwin-27B-ZTC-v2, an open-weights model that scores decisions with ZTC (Zero-Token Confidence), generating no tokens at all.
Why typed, zero-token decisions matter
Most agent stacks still ask a chat model to pick a route or a label, then parse the prose back into a decision. That is slow, non-deterministic, and every generated token is a chance to drift or hallucinate. A typed decision removes the prose: the model reads the problem in one forward pass and returns the decision directly. With ZTC, it generates zero tokens, so the result is deterministic and cheap enough to run on every request.
# one forward pass, zero generated tokens
h = model(text).last_hidden_state[last_non_pad_token]
score = ((h - mu) / sd) @ w # calibrated probe -> typed decision
The open side is usable today
The difference between the closed and open sides is simple: you do not have to wait for a funded product to try the idea. Our decision stack is public under Apache-2.0.
- ZTC method and inference code: https://github.com/final-bench/ztc
- Darwin-27B-ZTC-v2 and the Decision benchmark write-up: https://github.com/final-bench/s1mb
- ONGRID, an ontology graph plus retrieval plus typed decision engine: https://github.com/final-bench/ongrid
- Models and demos: https://huggingface.co/FINAL-Bench
FAQ
What is a typed decision model?
A model that returns a structured, typed result such as a choice, a score, or a probability that software consumes directly, instead of free-form text a program must parse.
What is ZTC (Zero-Token Confidence)?
A method that produces a decision in a single forward pass and generates zero tokens, by applying a calibrated probe to the model's final hidden state. It is deterministic and cheap.
Is there an open alternative to closed typed-decision products?
Yes. VIDRAFT's Darwin-ZTC and the ZTC method are open under Apache-2.0, which is why the article names them as the open-weights alternative.
Why not just use a chat model to decide?
Generation is overhead for a decision, it is non-deterministic, and parsing text back into a decision adds drift. A typed, zero-token decision avoids all three.
Top comments (0)