Module kinds 

Value kinds (spec 002, 3.5). Each is its own type with private fields and fallible construction; there is no From or Into between any two, so a score never becomes a probability and a label never acquires mass.

fn launder(label: SelectedLabel) -> Distribution { label.into() }

fn launder(d: Distribution) -> CalibratedProbability { d.into() }

fn launder(s: ModelScore) -> Distribution { Distribution::from(s) }

// A calibrated probability cannot be built without applying an artifact.
fn forge(id: CalibrationId) -> CalibratedProbability {
    CalibratedProbability { calibration: id, options: vec![], masses: vec![] }
}

Structs

CalibratedProbability

A distribution passed through a calibration artifact bound to this step. It records which transformation was applied; it does not establish that the result is calibrated on current data (spec 002, 3.6.2).

Distribution

Non-negative mass over declared options, summing to 1 within tolerance. Uncalibrated: it is not a calibrated probability.

KindError

Why a supplied value is not a valid value of its kind. Becomes invalid_backend_output at the backend boundary.

ModelScore

An unnormalized score, comparable only within its scope.

OrdinalLevel

A level on a declared rubric, low to high, as a distribution over levels or a single level. Its expectation is an index summary only.

RankPosition

An order over candidates with the scores and components that produced it; not a probability of being best.

ScopeMismatch

Two scores from different scopes were compared.

ScoreScope

Where a model score may be compared: one plan, step and request.

SelectedLabel

One option, with no mass and no uncertainty measure.