Training

How supervision becomes a fixed scoring geometry — embedding math, no language model.

Training turns a model’s supervision into a fixed scoring geometry. It fits each trait’s axis in embedding space and calibrates the breaks that separate score tiers, producing a version that serves deterministic scores.

Definition

Training is the lifecycle transition that produces a trained model. It consumes the model’s samples, fits the geometry of each trait, and records the calibrated thresholds as a version. Training runs no language model — it is embedding math, and its output is fully determined by its inputs.

Mechanism

A model can be specified in full — every example supplied — or minimally, with only the essentials of each trait: a name, a question, or a pair of poles. Before the math runs, a build step hydrates a minimal specification into a complete one, synthesizing the examples and pole exemplars the author did not supply. The training engine always sees a fully-specified model.

Synthesis happens only at build time. It never runs at score time, and it is separate from the training math — so a trained model is reproducible and its scores carry no language-model dependency. A fully-specified model makes the build a no-op: the engine sees exactly the examples supplied.

Training also calibrates each trait against its own training distribution. It places the breaks at cluster quartiles, sets the confidence regions, and fixes the tier boundaries. Calibration is per-trait and re-derived on every run, so the same raw score can map to different tiers across traits or across versions.

Effort

How much cost and time a build spends is set by effort, an authoring level from low to max that defaults to high. Higher effort spends more on synthesis and searches more thoroughly for the best scoring configuration, raising quality at proportional cost. The full set of levels is documented with the authoring schemas.

Effort applies to a cold start or an explicit re-tune. Routine background retrains reuse the existing configuration and run no search.

Interpretation

• Cold start is training the first time: a model becomes scorable from whatever supervision it has, with no prior usage data required.
• After the first run, feedback accumulates new samples and a background retrain folds them in, producing a fresh version without re-searching the configuration. This keeps a serving model current cheaply.
• An explicit re-tune searches for a new configuration — worth the cost when the data has changed enough to warrant it.

Edge cases

• Training is asynchronous: the model reports busy while a run executes and returns to ready (or failed) when it finishes.
• A re-tune on a ready model runs a full search; without it, a ready model trains incrementally through the background retrain rather than re-searching.
• Background retrains run no configuration search, so they are fast; only a cold start or an explicit re-tune searches.

Next

Tiers & breaks

The thresholds training calibrates.

Versions

The snapshot each run records.