Models

A learned geometry in embedding space that applies a standard of judgment to content.

A model applies a standard of judgment to content. It reads a piece of text and places it relative to a standard learned from examples, reporting one or more trait scores and a composite as a score card.

Definition

A model is a learned geometry in an embedding space. Each trait it measures is an axis fit to labelled examples — content paired with a polarity or quality label. The standard of judgment lives in that geometry: examples that demonstrate the standard, distilled into a fixed shape any new content can be measured against.

A model is learned from demonstrated examples — written rules, natural-language prompts, and hand-tuned weights play no part. Curating examples requires only recognizing the standard, not articulating it; the model captures what its author knows how to recognize and makes that judgment repeatable.

Mechanism

Building a model and using it are distinct steps:

Step	What happens
Define	Declare the model’s traits and supply supervision — the examples that demonstrate the standard.
Train	Training fits each trait’s geometry and calibrates the thresholds that separate tiers.
Score	The trained model projects new content onto each trait and returns a score card.

Scoring runs no language model. A trained model is fixed embedding geometry, so the same content and the same model version always produce the same score card. A model is small and fast, and behaves identically across every surface that calls it — the REST API, the MCP server, or a model-card read.

Each model has a stable handle within a namespace — for example u22a8.commit-message. The public catalog under u22a8.* ships ready-to-use models; an account authors its own in namespaces it owns.

Interpretation

A model is a unit of judgment, scoped to one standard. Several traits on one model report several judgments about the same content at once; the composite folds them into a single number. To measure a different standard, use — or build — a different model.

Because scoring is deterministic and carries no per-call language-model inference, a model behaves like infrastructure: a gate in a pipeline, a check in review, a ranking signal. Its judgment is consistent with the examples it was taught, repeatable across runs.

Next

Traits

The axes of judgment a model measures.

Model lifecycle

How a model goes from draft to serving scores.