TypeSafe: Jev 1.13

Jev 1.13 is TypeSafe’s versioned System One model for structured decisions. Its producer identifier is jev-1.13.0. It evaluates text or textual JSON state against typed questions and returns values and probability distributions directly, without composing a written answer. The edition suits classification, routing, condition checks and rubric judgments whose possible outcomes are known before the request.

State and focused question preparation

Supply a string, a JSON object or an array of text values as the state. Use named fields when the evidence contains related records, a conversation and a policy. Retain the parts each question needs and refer to a specific field or item when that identifies the evidence clearly. Jev 1.13 is text only; images, audio, video and binaries need preparation into text or structured textual fields outside this model.

Each question has a type, instructions and any criteria its type requires. Instructions and criterion descriptions can use text or JSON structure. The producer returns answers under the question identifiers, but those identifiers are not part of inference. Write the complete question rather than relying on an identifier such as urgent to imply the intended criterion. A useful support state can contain the message and impact details while separate questions define its category and severity.

Choice: a supplied category and its distribution

Choice selects the highest-probability option from the unordered set defined by the caller. Its criteria map contains the option keys and descriptions, with null allowed when an option needs no extra explanation. The documented limit is 255 options per Choice. Add an other or no-match option when a taxonomy does not cover all possible evidence, and explain boundaries that a short category name does not settle.

The answer contains the selected option, a probability for every option and confidence. Keep the distribution when two candidates are close instead of treating the selected label as certainty. The selected category can identify a proposed route; the surrounding application decides whether and how to perform that route.

Score: an expected level, not an exact measurement

Score uses an ordered array of described levels, with at least two recommended and up to ten supported. Indices start at 0. The numerical result is the probability-weighted average of the level indices, so it may fall between levels. The answer’s legend relates the indices to their descriptions, and the level probabilities show how the judgment is distributed.

For severity, describe the difference between cosmetic impact, a usable workaround and a blocked task. A fractional score can help prioritize or aggregate judgments, while selecting the most probable category is a different operation. TypeSafe’s edition-specific guide warns that Score is weak for reconstructing an exact physical or numerical quantity by interpolating between levels. Keep precise arithmetic and measurements outside the model.

Noul and confidence have distinct meanings

Noul returns the probability, from 0 to 1, that a defined condition is true. Optional true and false criteria clarify the boundary. It has no separate confidence field. A probability near 0.5 indicates uncertainty about yes versus no, not a middle value on a graded scale.

Choice and Score confidence are derived from their own distributions. For Choice with at least two options (n > 1) and selected probability p_max, the documented formula is (p_max − 1/n) / (1 − 1/n): it compares the top outcome with an even split. Score confidence also accounts for the probability-weighted distance from the most likely level, normalized against the spread of a uniform distribution and floored at 0. Thus adjacent ambiguous levels and opposite-end ambiguity have different meanings. Keep the full distribution when the task needs more detail than a single statistic.

Independent judgments over shared evidence

Questions in one request see the same state and are evaluated independently. A result from one question is not hidden context for another. Split a broad judgment into focused concerns, then interpret or combine the returned values according to the task. An issue can have a category, an impact rating and a relevance check without turning those independent criteria into one vague priority label.

If a later question needs evidence fetched or options chosen from an earlier result, make that dependency explicit in a later request. For candidate extraction, first obtain the available values from the source, then use a bounded Choice to select among them. Jev 1.13 is not trained for free-form text generation and should not be described as producing an arbitrary new string or a narrative explanation.

Edition-specific context and language limits

The maker lists a 64k-token budget for the state plus all questions combined, and a 32k-token budget for the state plus the longest single question. Both constraints matter. A state that fits the total budget can still leave too little room for an individual question’s instructions and options. Remove unrelated material while keeping the evidence that determines the answer.

English is the edition’s primary training language and the producer reports its best accuracy there. Other languages, including CJK scripts, are accepted but do not have equal accuracy. Use representative source material from the intended language when assessing how probabilities or category boundaries should be interpreted. A text-only input contract does not establish universal multilingual performance.

Published substantive limitations

TypeSafe documents literal reading of scoping words and negations, unreliable counting and arithmetic, difficulty comparing dates as ordered quantities, and reduced reliability with multiple layers of indirection. It recommends doing exact computation in code and giving the model the semantic judgment. Numerical encodings can be less effective than meaningful names, and a large state filled with unrelated detail can distract from the relevant evidence.

The same guide says adversarial content in state can steer the answer; contradictory instructions and criteria can confuse it; and Choice option order can sometimes bias a result toward an early option. Define the actual condition, align instructions with criteria, reduce unnecessary hops, and compare borderline judgments against the source. Keep the Jev 1.13 limits in mind when interpreting its judgments.

A useful Jev 1.13 artifact preserves the state identifier, complete question, options or levels, typed result, probabilities and relevant confidence statistic. Keep the version explicit when results or thresholds depend on this edition. Review uncertain cases against the evidence and let the application perform exact calculations and workflow actions; the model supplies a focused judgment with numerical uncertainty.