Jev makes decisions instead of writing. Where should it sit in an AI system?
A practical look at TypeSafe AI’s System One model, its typed decision contract, and how teams can evaluate bounded automation without confusing schema safety with correctness.
A surprising amount of production AI work is not asking for a paragraph. It is asking for a bounded judgment: which queue, which risk band, whether to continue, or when to ask for help. Teams often use a generative model for those decisions, then add prompts, JSON schemas, validators, and retries around it.
TypeSafe AI proposes a different primitive. Its first public “System One Model,” Jev, accepts application state and typed questions, then returns decisions rather than free-form text. The company launched Jev in early access on 15 September 2026 and positions it as intelligence designed to sit inside software.
System pattern
A decision model belongs inside a governed path
The model supplies a probabilistic judgment. Application policy still determines what that judgment is allowed to trigger.
State + typed questions
A ticket, record, trace, or document plus the choices, rubric, or yes/no judgment the system needs.
Typed probabilistic decision
A Choice, Score, or Noul answer shaped for software, with probabilities and confidence where defined.
Thresholds + permissions
Code checks confidence, business risk, authorization, and fallback rules before choosing the next step.
Run bounded code
Use a validated, high-confidence decision inside an explicitly permitted path.
Hand off to an LLM
Let a generative model write, plan, or explain after the route is selected.
Request human review
Stop when confidence, risk, or policy does not support autonomous action.
The output contract is the product
Jev’s public API describes one request containing shared state and one or more named questions. Each question has a predefined type, and each returned answer matches that type. The model does not need to write JSON-shaped prose and hope the application can parse it.
Choice
Select one option from a set you define.
Example: Route this ticket to billing, support, fraud, or retention.
Score
Rate the state against an ordered rubric.
Example: Place this incident between low and critical urgency.
Noul
Return the probability that a yes/no statement is true.
Example: Does this response require a policy review?
This changes an important failure mode. The application no longer has to recover structure from an unconstrained string. It does not remove the harder question: whether the selected answer is correct.
Important distinction
Schema safety is not decision accuracy
TypeSafe says Jev cannot make type errors because the possible output shape is fixed in advance. That is a useful engineering guarantee. It should not be translated into “the model cannot be wrong.” A model can always choose the wrong valid option, assign the wrong score, or produce a poorly calibrated probability.
The practical control is not a slogan. It is a workload-specific evaluation: compare predictions with reviewed labels, inspect performance by segment, test adversarial and ambiguous inputs, and verify that a confidence threshold produces an acceptable error rate. Vendor speed and cost results are useful hypotheses, not substitutes for that evidence.
Where it could fit—and where it should not
The strongest candidates are repeated, bounded decisions whose valid outcomes are known before the call: routing work, scoring relevance, flagging policy conditions, ranking candidates, or verifying an agent trace. Fast decisions may also make new real-time paths economical.
Jev is not a replacement for a model that must draft an answer, explain a trade-off, form a novel plan, or synthesize evidence. It is also not a permission system. High-impact actions still need deterministic authorization, auditable policy, and a safe path for uncertainty.
Recohut perspective
Recohut AI Model Landscape
A visual map of models by their typical output and the breadth of tasks they address.
Recohut perspective
Three rules for evaluating a decision model
Evaluate the decision, not the demo
Build a labelled set from the real workload. Measure accuracy, calibration, latency, and the cost of each wrong branch before comparing it with rules or an LLM.
Keep confidence separate from permission
A high-confidence answer is still a prediction. Business policy, user permissions, and action risk should independently determine whether software may proceed.
Design the fallback before the threshold
Every uncertain or disallowed path needs somewhere safe to go: deterministic code, a generative model, a human reviewer, or no action at all.
From model to workflow
Designing an agent decision path?
Recohut helps teams place models, policies, tools, and human review inside measurable production systems.