Training is the process that produces a model from supervision. The supervision expresses a standard; training translates it into the trait axes, breaks, and calibrated scoring behavior that define a deployable model.
A training run takes a model specification and a set of supervision inputs and produces a trained model. The trained model contains: the named traits with their polarity labels, the trained breaks for each trait, and the fitted parameters used at scoring time. A trained model is self-contained; scoring against it does not require the original supervision.
Every supervision form resolves to labeled samples before training runs. Direct samples are consumed as-is; references, briefs, and feedback are translated by the corresponding intake path into samples that represent the expressed standard. Training is sample-driven; the supervision forms differ in how the samples are sourced, not in what happens after.
From the assembled samples, the training run fits the scoring parameters for each trait, computes the trait's positive and negative training distributions, and derives the three breaks as the quartiles described on the breaks page. The result is a model that can score new content and return trait scores, tier labels, headroom, and confidence for each.
When the traits to measure are known in advance, they are declared in the model specification. When they are not, trait discovery proposes them from the contrast between positive and negative samples and returns them for review. Discovered traits can be renamed, merged, dropped, or edited before the training run finalizes.
Scoring method, axis construction, break placement rule, and distribution handling are selected by the training run from candidates evaluated against the supervision. They are not exposed as configuration. The model specification declares the traits and intent; the rest of the fit is internal to the training run.