Jev: When AI Judgment Becomes a Software Primitive, How Does the Boundary of One-Person Companies Change?

Typical autoregressive LLMs are pretrained to learn the next token, with sequence probability written as P(w₁…wₜ | x) = ∏ P(wᵢ | x, w₍<ᵢ₎). Post-training improves behavior through human preferences, verifiable rewards, etc. RLHF rewards answers humans prefer, but "satisfactory answers" do not automatically equal "accurate probability estimates of correctness." RLVR can reward correct answers, but also cannot derive calibrated probabilities solely from accuracy. TypeSafe calls its approach RLCD, Reinforcement Learning for Calibrated Decisions: training directly for decisions and probabilities. Its official tutorials also place RLCD as a post-pretraining training branch. Therefore, the category shift should be understood as a change in optimization objective and computational interface, not as proof that the foundation has no relation to language models.

A statistical example explains this objective difference: the Brier score for binary prediction is (p-y)², where p is predicted probability and y is the actual outcome (0 or 1). It penalizes confident errors; in expectation, honestly reporting conditional probabilities beats arbitrarily overconfident statements. Such strictly proper scoring rules show the mathematical basis for "training probability honesty." This is only an explanatory example, not Jev's disclosed loss function. Official materials reviewed do not provide sufficient information to reproduce the complete RLCD objective function, reward construction, and ablation experiments.

Jev's basic interface is state + questions → answers: input can be text or structured state, and output consists of three primitives defined before the call. Choice selects among up to 255 predefined options and returns probabilities for each, suitable for classification, routing, and selecting existing candidates. Score provides distributions and weighted scores for 2-10 ordered levels defined by text, suitable for severity, relevance, and priority. Noul returns the probability of "yes" in range 0-1, suitable for relevance and condition satisfaction.

Given a legal option set S, type safety means P(output not in S) = 0. For example, if the program provides only three verified paper IDs, the model has no channel to output a fourth fabricated ID. This eliminates one degree of generation freedom. However, the model may still choose the wrong paper or judge "unsupported" as "supported." Therefore, "mathematically impossible to be type-wrong" cannot be extended to "mathematically impossible to be fact-wrong." The zero-error guarantee in official release clearly falls on schema matching. Engineering design must also leave exits such as "other," "not mentioned," and "insufficient evidence" in the candidate set.

Typical autoregressive models decode step by step; even if the answer is JSON, subsequent tokens depend on prior context. Jev's officially described mechanism shares input state, evaluating multiple questions in parallel and in isolation within a single call, returning results directly rather than token-by-token. For example, processing a reader feedback can simultaneously judge topic, urgency, whether it points out factual errors, and whether it proposes collaboration. The program receives results and then decides which judgments are useful. Officially this pattern is called speculative fan-out: precomputing branches that may be needed, reducing serial round-trips.

Shared state also changes how software is organized: instead of repeatedly passing in the same article and asking ten labels sequentially, ten clearly defined questions can be placed in a single request. This reduces repeated input and network round-trips. However, the questions themselves still count against input budget, and batching more questions is constrained by context and service limits; "free output" should not be understood as unlimited judgment with no computational cost. There are two boundaries. First, parallel evaluation does not mean events are statistically independent; probabilities cannot be arbitrarily multiplied. Second, next steps requiring new evidence still need retrieval or additional calls.

If a model reports approximately 0.8 for a batch of events, calibration means about 80% of those events occur. It is a population frequency property, not a guarantee for any single judgment; "more confident means more accurate" itself is also weaker than strict calibration. There is a detail that can easily lead to erroneous deployment: Jev's confidence field is a concentration measure calculated from the probability distribution, and cannot be directly interpreted as the probability that the selected answer is correct. Choice and Score return full distributions; Noul has only binary probability with no independent confidence field. When computing expected loss, use business-validated event probabilities rather than mixing the two fields.

The Jevons paradox behind Jev's name provides an economic explanation for this path. In the 19th century, Jevons noted that improving fuel efficiency could expand its uses and ultimately increase total consumption. Transferred to AI, the inference is: when judgment becomes cheap enough, software will begin executing judgments that previously were not worth executing. Where human classification was once done only for the most important reports, in the future multiple judgments can be attached to every message, every retrieval fragment, and every tool call. Assuming per-call cost drops 100x and demand grows 1000x, total expenditure still grows 10x. But this growth depends on demand elasticity, business benefit, and error cost; "cost decrease" cannot be written as an inevitable law of demand explosion.

For OPCs, the scarcer resource is likely attention. Ten thousand judgments per day, even with only 1% escalated to humans, means a hundred reviews; two minutes each is two hundred minutes. Therefore, in addition to controlling errors, escalations must be controlled, duplicate events merged, and highest-loss items prioritized. Research-oriented OPCs can start with a low-risk process, use hundreds to thousands of human-annotated materials to establish a baseline, run without affecting actual processing, then gradually open automated classification. Focus on recording miss rates, false-positive rates, manual processing volume, tail latency, and total cost per valid result.

A viable combination is: code handles deterministic checks, Jev handles high-frequency, well-scoped judgments, reasoning LLMs handle complex or low-confidence cases, and humans handle high-cost disputes and final responsibility. LLMs can also help design questions and evaluation sets during development. System One and System Two describe task division here; using Jev to check LLM outputs adds a verification layer, but the checker itself must also undergo independent evaluation. Similarly, it is not advisable to ask Jev for a broad score like "is this research worth publishing." A more operable design is to separately judge whether material is new, whether arguments have sources, and whether evidence contradicts itself, then combine via code. This way business priorities are controlled by the operator, and the model doesn't have to secretly trade off all objectives in a single score.