Lesson summary

Typed judgments provide evidence; deterministic code controls the action.

1. Noul returns a probability from 0 to 1. Choice and Score also return a confidence field.
2. No. Questions independently read the same state. Code must compose a later call if a question needs a previous answer.
3. No. All mass on level 1 and an even split across levels 0 and 2 both have mean 1.
4. Across many comparable cases assigned 0.80, the outcome should occur about 80% of the time.
5. Code compares its confidence against the application threshold. Raising the threshold can reject the same response.

Stage 1 of 7

A ticket needs a decision

A ticket arrives: PDF export hangs, but CSV still works for some teammates. Our application needs to decide where it goes. A paragraph about the ticket is useful to a person; a routing function needs bounded values it can test.

Jev, TypeSafe’s System One model, takes state and named questions and returns typed answers. State means the evidence supplied to the call. Criteria define the possible answers. A probability expresses how likely an answer is.

We will follow one invented response through three question types, parallel evaluation, uncertainty and the code that acts on it. The values in this lesson are teaching fixtures, not measured Jev responses.

Stage 2 of 7

Three question types, three answer shapes

Start by giving each judgment its own question. All three read ticket T-104. Category asks what kind of ticket it is; repro asks whether reproduction steps are present; severity asks how badly the feature is broken.

A Choice distributes probability across the alternatives we supplied: bug, billing and feature. The returned choice is one of those keys; it cannot invent another key. A valid key can still be the wrong judgment.

Repro is a Noul: it returns the probability of yes, here 0.90. That is not a Boolean until our code applies a threshold, and there is no separate confidence field.

Severity has ordered levels. A Score returns probabilities over our rubric, then their weighted mean: 0×0.05 + 1×0.25 + 2×0.70 gives 1.65. That is a position on our rubric, not a measured fraction of affected users. Choice and Score also return confidence, which we will inspect shortly.

All three answers refer to the same ticket. Their output types differ, but each question can be evaluated from that shared state.

Stage 3 of 7

Questions do not wait for one another

Those questions share the same ticket, so none needs to wait for another answer. The top trace represents conventional autoregressive output: each generated token depends on the prefix already produced.

Jev’s documented interface evaluates questions in parallel and in isolation against the shared state. Follow the three question lanes: their answers become available together. This is a dependency schematic, not measured milliseconds or a disclosed network diagram.

Category does not feed into severity inside this request. We can ask for severity speculatively and ignore it later if category is billing. A question that truly needs a previous answer requires code to compose another call.

TypeSafe describes a parallel sampler and a new architecture, but the public sources do not disclose enough to reconstruct their algorithm. The useful contract is narrower: one state, independent questions, typed results.

Stage 4 of 7

A mean can hide uncertainty

A returned value can hide how uncertain the judgment was. Look at two possible severity distributions on the same three-level rubric. One places all its mass on level 1; the other splits it between levels 0 and 2.

Both distributions have a weighted mean of 1.00, despite making very different claims. The concentrated answer selects a middle level decisively; the split answer is uncertain between the extremes.

That is why Choice and Score also return the full distribution and a confidence statistic derived from its shape. Confidence is not an independent estimate that the answer is correct. This diagram compares distribution shapes without inventing the service’s exact confidence formula.

Change the distribution and inspect what the unchanged mean fails to tell you. Keep the probabilities when a single score would hide a distinction your policy needs.

Stage 5 of 7

Check probabilities against outcomes

Even a sharply concentrated distribution can be wrong. To evaluate a probability, we need outcomes from many cases. This illustrative cohort contains one hundred tickets that all received Noul 0.80 for the same yes/no question.

In the matched cohort, eighty yes cases agree with that probability. Each filled cell is an observed yes, and each outlined cell is an observed no. Compare the predicted rate with the observed rate for the selected cohort.

Switch to the shifted cohort. Fifty-five yes outcomes against the same prediction expose overprediction in this cohort. Calibration compares probabilities with observed frequencies across comparable cases; it cannot certify an individual answer.

RLCD, Reinforcement Learning for Calibrated Decisions, is TypeSafe’s name for training toward calibrated probabilities. The public description gives the target, not a reproducible reward function. These synthetic outcomes explain how to check that target; they do not establish Jev’s accuracy on your workload.

Stage 6 of 7

Code owns the final branch

Return to T-104 with the original three answers. The application receives category, severity and repro separately. Its policy can inspect each value without asking the model to write a plan.

First compare category confidence, 0.85, with the policy threshold, 0.70. That gate controls whether code proceeds automatically. It does not change the returned probabilities.

Because category is bug, code uses severity 1.65 and repro probability 0.90. It tests severity above 1.50 and repro above 0.60. Both predicates must pass.

The ticket goes to engineering. Raise the confidence threshold to 0.90: the model has said exactly the same thing, but code sends the ticket to human review before evaluating the bug branch.

Application code combines independent judgments and selects the action.

Try both confidence thresholds, then the illustrative billing case. When billing passes the confidence gate, code ignores the irrelevant severity and repro answers. The branch belongs to the application.

Independent questions provide judgments; they do not jointly enforce business rules. Type-safe output can still contain a wrong judgment. TypeSafe documents errors with arithmetic, indirection and adversarial state: keep exact calculations, permissions and structural invariants in code, and test the questions on labeled cases from your workload.

Stage 7 of 7

Connect the whole decision path

Choice selects a supplied alternative, Noul returns a yes probability, and Score averages ordered rubric levels. Each has a specific role in the same typed response.

Parallel questions share evidence without consuming each other’s answers. Code can ask speculative questions in one call and use only the results relevant to its branch.

The same mean can hide very different distributions. Retain the probabilities and treat confidence as a summary of their shape, not a promise of correctness.

Calibration needs observed outcomes across a population. An individual answer, even one with high probability, cannot establish that relationship.

Code decides which judgments matter and what they permit. Changing a threshold can change the route without changing the model’s answer.

The complete path is state and criteria, independent typed judgments, inspected uncertainty, then a policy-controlled action. Jev supplies a decision interface inside a program; the program still owns the consequences.

Research behind the lesson

Primary TypeSafe documentation checked September 22, 2026. Example tickets, distributions and calibration cohorts are illustrative. The diagrams model the published interface, not undisclosed network layers, sampler code or an RLCD reward function.

• TypeSafe API: state, questions and typed answers
• Choice: bounded alternatives
• Score: a probability-weighted rubric
• Noul: probability of a yes answer
• Confidence and its distinction from probability
• AI primer: RLCD and calibration
• Speculative fan-out
• Jev 1.13: documented failure modes
• Launch: parallel sampling and disclosure boundaries