Introducing Simplex FastCD: fast decisions for interactive systems

Many decisions in an interactive system cannot wait for a long answer.

Is the user still speaking? Is this a good moment to initiate an interaction? When several events arrive together, which should be handled first? The output may be a single choice, but that choice determines what the rest of the system does next.

Meet Simplex FastCD (Fast Constrained Decision-making), our fast decision model adapted from open Qwen-family models. We have used this approach in our interactive models and robot systems. We are now opening it up for others to explore.

From internal decisions to an accessible model

We have long used System-one models for conversational turn detection, proactive triggers, social perception and event prioritization. Some compact decision stages have a 50–200 ms response budget: decide whether to act and which action to take, then let the rest of the system carry it out.

That range is an application design target, not a latency promise for every FastCD request. Context length, question count, images, reasoning settings and service load all affect response time.

Jev brought wider attention to the potential of these models in robotics, autonomous driving and Browser Use. It also encouraged us to turn our own work into something people could try. Watching a constrained choice drive the next step on a computer or a robot was a striking moment for us.

Decisions that fit into code

FastCD accepts text, structured context and images, and returns structured decisions. Supply the current state, a question and the allowed options; the model makes a choice within those boundaries.

• Choice: select one option from up to 1,024 candidates and inspect the candidate distribution.
• Multi-choice: select multiple relevant candidates for tasks such as multi-label decisions and identifying simultaneous events.
• Yes/no: obtain the probability of “yes” and use an application-specific trigger threshold.
• Ordered scoring: evaluate a state against ordered criteria for ranking or downstream rules.
• Multimodal input: combine images with a question to make visually grounded decisions.

An interactive system might choose among continuing to listen, responding or deferring an action. A visual application might select several objects that meet a condition. The application then combines the decision with its own state and execution constraints.

Compared with Jev’s public interface, our main extensions are image input, a larger candidate set and native multi-choice. At the time of writing, Jev documents a 255-option Choice limit and provides Choice, Noul and Score primitives. Multi-label tasks can also be composed from individual judgments; FastCD gives them a direct multi-choice interface.

What our confidence evaluation shows

We care about both getting answers right and identifying answers that are more likely to be wrong. In an internal comparison of 2,000 paired questions, we used the maximum candidate probability as a common confidence measure for FastCD and Jev.

Metric	FastCD	Jev
Error-detection AUROC ↑	0.745	0.751
Agreement with reference labels ↑	65.85%	73.35%
Calibration error, 10-bin ECE ↓	7.59%	2.92%

The evaluation uses the public typed-decisions dataset, with reference labels derived from teacher distributions rather than human ground truth. AUROC measures how well confidence separates correct and incorrect answers. The values are numerically close, while Jev has better label agreement and calibration in this comparison. Similar AUROC does not establish equal overall performance or demonstrate that we reproduced Jev’s RLCD training results.

Candidate probabilities and confidence can help decide when to proceed, retry or route a question elsewhere. They should not automatically be read as calibrated probabilities that an answer is correct. Thresholds need validation on the application’s own tasks.

We have also seen useful results in internal interaction tests. On 25 fixed, handwritten Chinese interaction states, a decision about whether partial speech recognition justified prefetching matched 23 expected answers, with a mean request time of about 144 ms. This small synthetic test demonstrates feasibility under one configuration, not an overall success rate with real users.

Demos to follow

We will add short demonstrations showing the input, candidate options, decision and resulting action together. The following spaces are reserved for those videos.

Demo 01

Demo 01 · Computer interaction — coming later

Planned content: choose the next action from a screen state and a set of candidates, then observe the page update.

Demo 02

Demo 02 · Robot interaction — coming later

Planned content: decide when to respond and which event to prioritize when several things happen together.

Demo 03

Demo 03 · Multimodal multi-choice — coming later

Planned content: select multiple candidates using an image and a written condition.

A free preview, with more work ahead

We are deploying FastCD on our own cluster and making the initial preview free. Deployment and the experience are still evolving, and concurrency and rate limits will apply during the free period.

Start with the FastCD Playground, or use the API documentation to connect a concrete decision to your application.

We will continue organizing evaluations and demos. As time allows, we also hope to share our perspective on RLCD and further research, potentially including open-source work. For now, we want to put fast decisions into more hands and see what people build.

References

• TypeSafe: introducing System One and Jev
• Jev API documentation and model documentation, checked September 21, 2026
• typed-decisions dataset