# Decision Models (DMs)
A decision model (DM) is an AI model whose output is a **decision**, not text. You give it an input and a bounded set of possible answers, and it gives you back numbers: which option, how likely yes is, what score. It's the category [[Jev]] belongs to, and I think it deserves its own name next to [[Large Language Models (LLMs)]] and [[Small Language Models (SLMs)]].
Why bother with a separate term? Because a LOT of what we ask LLMs to do today is actually deciding. Is this email urgent? Which tool should the agent call? Is this paragraph off-topic? Did the agent get stuck? We ask a text generator to write the answer as JSON, parse it, validate it, and retry when it breaks. A decision model answers that kind of question directly, and that changes the economics: when a decision costs a fraction of a cent and takes 200 ms, you can afford to ask it everywhere.
## What makes a model a decision model
- **The answer space is bounded and known upfront.** Options, yes/no, or a numeric range. If you can't list the possible answers, you need an LLM
- **The output is typed by construction.** There's no text to parse, so malformed output can't happen
- **The output is a probability.** Ideally a *calibrated* one: when it says 80%, it's right about 80% of the time. That lets you set real thresholds ("send to a human below 0.7"), something you can't do with an LLM writing "I'm 90% sure"
- **No generation loop.** The input is read once and the answer comes out of a readout layer, not token by token. That's where the speed comes from
## An old idea with a new twist
Decision models aren't new. We've had them for decades under other names:
- **Classic classifiers**: spam filters, sentiment analysis, intent routing (see [[Binary Classification]] and [[Data classification]]). Fast and cheap, but trained for ONE task with your own labeled data
- **Zero-shot classifiers**: NLI-based models (e.g., BART fine-tuned on MNLI) that score arbitrary labels without training. Closer to the idea, but limited in quality
- **Rerankers and reward models**: they score an input against a query or a preference. Decisions again, just packaged for search and for training LLMs
- **LLMs used as deciders**: prompt an LLM, force [[LLM Structured Outputs]], or read the token probabilities of each option. It works, but you pay for a text generator to pick a label
What changed in 2025-2026 is **generality**. A general decision model answers *arbitrary* questions over *arbitrary* options at inference time, with no task-specific training, at a quality close to a frontier LLM. [[Laya]] went in that direction in March 2025. [[TypeSafe AI]] made it a product with [[Jev]] in September 2026, and calls the category [[System One Models]], after [[Daniel Kahneman]]'s fast, intuitive System 1. I prefer "decision model" as the generic term: it says what the thing outputs, and it doesn't depend on one company's branding.
## DMs vs LLMs
| | Decision model | LLM |
|---|---|---|
| Output | A choice, a probability, a score | Text (possibly JSON) |
| Answer space | Bounded, defined by you | Open |
| Latency | Tens to hundreds of ms | Seconds |
| Cost | Fractions of a cent per decision | Much higher, output tokens included |
| Failure mode | A wrong but well-formed answer | Malformed output, hallucinated options, wrong answer |
| Explanation | None, just a number | Can explain (whether the explanation is faithful is another story) |
They're complementary. The pattern I find the most useful (from [[Nate B. Jones]]) is to think in three building blocks: **code** for rules and calculations, **decision models** for bounded judgment, **LLMs** for open reasoning, writing and exceptions. A decision model can even decide when an LLM is worth calling, which is [[Model routing]] done cheaply.
## Current examples
- [[Jev]] by [[TypeSafe AI]]: closed, the first commercial general decision model
- [[Laya]]: open, multilingual, published in 2025
- [[Kev]]: open models on top of [[Qwen]] with a Jev-compatible API
- [[SemIf]] (formerly OpenJev): reads the probabilities of each answer from an LLM's output, runs locally
- [[Jevlike]]: a small trainable option scorer, closer to a classic classifier
- [[fast-jev-compaction]]: an example of what DMs enable, compacting a Claude Code session by scoring each tool call instead of summarizing
## Where to use them
A decision model rewards a specific way of building: many narrow questions answered in parallel, combined in code ([[Atomic Question Decomposition]]), or fed as features to a small classical model trained on your labels ([[LLM-Generated Features for Classical ML]]).
- Classification at scale: tickets, emails, leads, posts, ads
- Guardrails and checks inside [[AI Agents]]: is this relevant, is the agent stuck, which tool, how much reasoning effort
- Asking many small questions about the same input (ten yes/no checks per paragraph for a prose linter, for instance)
- Triage before an LLM: let the DM filter, and only send the hard cases to the expensive model
## Watch out for
- **No reasoning to inspect.** You get a number, not a why. Bias and blind spots are harder to see, so you need MORE evals, not fewer
- **Calibration isn't universal.** Calibrated on the vendor's data doesn't mean calibrated on yours. Measure it on your own distribution before trusting thresholds
- **Wrong tool for open questions.** If the right answer isn't in your list of options, the model will still confidently pick one. Pair the pick with an absolute yes/no check per option, or a "does an answer exist at all?" question ([[Relative vs Absolute AI Judgments]])
- **Vendor dependence.** The best general DM today (Jev) is closed, and its terms forbid publishing benchmarks. The open alternatives are behind but closing the gap
My take: this is one of the most practical ideas in AI this year. Not because the tech is magic (it isn't), but because it makes judgment cheap enough to sprinkle everywhere. I expect many of the "small" LLM calls in my own automations to move to a decision model over time.
## References
- [Introducing System One Models & Jev (TypeSafe AI)](https://typesafe.ai/blog/introducing-system-one-models-and-jev)
- [Jev notes (Simon Willison)](https://simonwillison.net/2026/Sep/21/jev/)
- [Nate B. Jones on Jev build patterns](https://x.com/natebjones/status/2101421649774236145)
- [Jev's architecture unmasked (Archer Hume)](https://archerhume.com/posts/jevs-architecture-unmasked/)
## Related
- [[Decision Models (MoC)]]
- [[System One Models]]
- [[Jev]]
- [[TypeSafe AI]]
- [[Laya]]
- [[Kev]]
- [[SemIf]]
- [[Jevlike]]
- [[Large Language Models (LLMs)]]
- [[Small Language Models (SLMs)]]
- [[LLM Structured Outputs]]
- [[Model routing]]
- [[Binary Classification]]
- [[AI Agents]]
- [[System One Primitives]]
- [[Atomic Question Decomposition]]
- [[Relative vs Absolute AI Judgments]]
- [[Self-Consistency]]
- [[LLM-Generated Features for Classical ML]]
- [[Machine Native Intelligence]]
- [[AI Model Calibration]]
- [[Confidence-Gated Routing]]
- [[AI Model Cascades]]
- [[Jevons Paradox]]
- [[Zero-Shot Classification]]