jev primitives: choice, score and noul

Start with the Shape of the Answer

The output shape is a modeling decision, not formatting. “Which team owns this ticket?” compares candidates, so it is a Choice. “How urgent is it?” uses ordered descriptions, so it can be a Score. “Does it request a refund?” tests one condition, so it is a Noul. Start with that plain-language question before writing criteria.

Which Primitive Should You Use?

Start with the Output Contract, Not the Model Category

Several alternatives can return JSON, but JSON is transport rather than semantics. LangChain’s structured-output documentation describes constraining a generative model to a schema. Jev instead exposes bounded decision primitives and distributions. A trained classifier learns its answer boundary from examples; a rule executes a boundary written in code.

The Same Ticket Can Require All Three Primitives

These questions share evidence but do not mean the same thing. A billing ticket can be routine or critical, and it may or may not request a refund. Combining them into one label such as critical_billing_refund creates a brittle taxonomy with many rare combinations. Separate results preserve meaning and let application code apply a different threshold or review rule to each decision.

If the route labels are unrelated categories, do not use Score merely because a database stores them as 1, 2 and 3. Billing is not “less than” technical support. Conversely, urgency is ordered, so a Score rubric can describe a meaningful progression. The Score-versus-Choice guide works through this distinction.

Choice Is a Relative Decision

Choice selects among criteria keys supplied in the request. Its probabilities compare those options and approximately sum to one. Include an explicit no-match option when none may fit. Do not force unrelated labels into one Choice merely because the UI displays one result.

Choice confidence summarizes how concentrated the distribution is. It does not prove that the candidate set is complete, the state is sufficient or the selected label is true. A confident choice among bad options is still a bad workflow.

Score Uses an Ordered Semantic Rubric

Score criteria are ordered descriptions whose positions define their levels. The returned score is the probability-weighted expected position and may be fractional. Each level must describe a meaningful situation on its own. Jev 1.13 documentation warns against using Score to reconstruct exact numbers or perform arithmetic.

Noul Is a Yes Probability

A Noul asks whether one condition is true and returns a value from zero to one. A value near 0.5 means the yes and no outcomes have similar probability; it does not mean medium intensity. When labels are independently applicable, one Noul per label is more faithful than a Choice.

Noul has no separate confidence field. Calling the Noul value “confidence” obscures whether the value means probability of yes or certainty in a selected answer.

Common Primitive Mistakes

• Using Choice when several labels may apply independently.
• Using Score for exact numeric extraction, counting or arithmetic.
• Treating a Noul near 0.5 as medium severity.
• Assuming a Choice probability and equivalent Noul must match.
• Omitting no-match when the candidate set may not contain the answer.
• Packing several independently useful judgments into one vague instruction.

Questions in One Request Cannot Depend on Each Other

TypeSafe says multiple questions over shared state are evaluated independently and in parallel. The urgency question cannot read the route selected by the Choice beside it. Batch questions only when each can be answered from the original state. If a second decision needs the first result—for example, choosing a billing specialist only after selecting billing—make the first request, validate its response, construct the permitted specialist list in code and then make a second request.

Parallel evaluation is an interface property, not a guarantee of constant end-to-end latency. Payload size, question count, provider queues, rate limits and retries still affect the complete path. See multiple questions in one call for the dependency pattern and error handling for system behavior outside the model.

Changing a Primitive Creates a New Evaluator

Moving from one Choice to several Nouls changes the statistical question. Choice probabilities are relative to the listed candidates; each Noul tests its label independently. Their values need not match, and an old threshold must not be copied across. The same applies when replacing a vague Score with a Choice or rewriting criteria. Treat primitive, question, criteria, state projection and resolved model as one versioned instrument.

Before migration, freeze representative labeled cases and preserve the old raw distributions. Run old and new designs side by side, compare class errors, calibration, automation coverage and review volume, and inspect disagreements by slice. Promote only when the new design improves the operational target without hiding a costly error. The Choice-versus-Noul guide, evaluator scorecard and regression-testing guide provide the detailed protocol.