THE MENTAL MODEL
A probabilistic function for software.
A traditional function expects exact inputs and follows rules written in code. A language model accepts ambiguous input but returns an open-ended string. Jev occupies a different point: it accepts ambiguous state, evaluates a finite decision space, and returns values that ordinary code can consume without parsing prose.
state + typed questions→probabilities + answersYour application still owns the workflow. It decides which state to send, which outcomes are legal, when confidence is high enough to act, and which cases require review.
THREE PRIMITIVES
Noul, Choice, and Score.
Noul: how likely is this statement?
Noul handles a yes-or-no proposition and returns a probability between zero and one. Use it for checks such as “Does this request require a refund?” or “Is this tool call risky?”
Choice: which option best fits?
Choice selects one named option from criteria you define and preserves the full probability distribution. It is useful for queue routing, tool selection, intent classification, and next-action decisions.
Score: where does this state sit on a scale?
Score evaluates an ordered rubric with two to ten levels. The probability-weighted result can sit between levels, making it useful for urgency, quality, severity, and priority.
GOOD FIT
Use Jev where the outcomes are known first.
Choose the next tool, subagent, retry path, or escalation.
Detect intent, score urgency, and route the request.
Estimate whether an action matches a policy before execution.
Judge whether an output satisfies explicit criteria.
LIMITS
What Jev is not.
- It is not a chatbot or writing model.
- It does not replace deterministic validation, permissions, or business rules.
- A high confidence value does not prove an answer is correct.
- It is a poor fit when the valid outcome cannot be described in advance.
- It should not autonomously execute high-impact actions without policy and review thresholds.
PRODUCTION PATTERN
Model the judgment. Keep the system deterministic.
- Define the state. Include only the context needed for the decision.
- Define legal outcomes. Make every option distinct and operational.
- Calibrate thresholds. Test against labeled examples from your workflow.
- Handle uncertainty. Route low-confidence cases to review or a safer fallback.
- Log the exact model. Keep the request, response, version, and downstream outcome together.