Traits

An axis of judgment a model measures — one question, one score.

A trait is a single axis of judgment: one question a model asks of content, such as “how clear is this?” or “is this on topic?” A model has one or more traits, and scoring returns one score per trait alongside a composite.

Definition

A trait is learned from labelled examples — content paired with a polarity or quality label — not from rules, prompts, or hand-tuned weights. Once trained, a trait is a fixed geometry in embedding space; scoring projects content onto that geometry to produce a universal score $s \in [0, 1]$, reported on a 0–100 scale.

Mechanism

Each trait carries the metadata needed to understand what it measures. These fields appear on a model card and through the list_traits operation:

Field	Meaning
key	Stable identifier — the key used in the score card’s scores and detail.
name	Human-readable trait name.
description	What the trait measures.
positive_label / negative_label	The trait’s two poles, e.g. Clear / Obscure.
kind	Definition mode — extrinsic or intrinsic (explained below).

A trait is defined in one of two modes:

• Extrinsic — the labelled samples are the trait; the axis is re-derived from them on every training run. This is the default for authored traits.
• Intrinsic — the axis is the trait; a fixed direction recovered from the data that re-training calibrates around but never rotates. Discovery produces intrinsic traits.

Separately, each trait has a type — its geometry — which determines the shape of the native score it reports.

Interpretation

A trait answers exactly one question, so a model with several traits reports several independent judgments.

• Scores are comparable within a trait: 80 is better than 60 on the same trait.
• Scores are not comparable across traits of different types — an 80 on a topic trait and an 80 on a spectrum trait describe different geometric facts.
• The composite folds the per-trait scores into one number.

Edge cases

• A model may have a single trait; its composite then equals that trait’s score.
• The poles are descriptive, not a fixed ranking — a trait can be designed so that either pole is the desirable one.

Next

Score types

The four geometries a trait can take.

Composite

How per-trait scores combine into one number.