Traits

A trait is an independent scoring axis on a model. Each trait is scored separately from the others, so per-axis strengths and weaknesses remain distinguishable in the output rather than collapsing into a single number.

§1Definition

A trait has a name, an optional description, and two polarity labels — short strings that name the ends of the axis. A model declares a set of traits; scoring against the model produces one score on the 0–100 scale per declared trait. Traits are independent: a trait's score is derived from the content's position along that axis only, not from any aggregation over other traits.

clarity
84
specificity
72
verification
31
Figure 1. Three trait scores on a single piece of content. Per-trait scores differ because each trait measures a distinct axis; a low score on one does not reduce the score on another.

§2Mechanism

§2.1Polarity labels

Each trait declares two polarity labels — one for the low end of the axis (near 0) and one for the high end (near 100). These labels are displayed on the score card to make the axis explicit; the numeric score alone does not convey which direction the trait measures. Polarity labels are set when the trait is declared and do not change at scoring time.

§2.2Independence

Traits do not interact during scoring. The same content scored against a model produces the same per-trait scores whether the model also declares other traits or not, because each trait's parameters are fit and applied in isolation. Interaction across traits enters only at the composite, which is a post-hoc aggregation and does not alter the per-trait scores it summarizes.

§3Interpretation

Per-trait scores carry diagnostic information that a single score cannot. A piece of content with a high score on one trait and a low score on another is reporting that the content does well on one axis and poorly on another — an observation the composite would collapse. When the decision depends on which axis is weak, the per-trait scores are the appropriate signal; when the decision depends only on overall quality, the composite is.

§4Trait counts vary by model

A model may declare as few as one trait or as many as its training supports. Profiles with more traits provide finer-grained diagnostics at the cost of more supervision and more surface area to interpret. The trait count is a property of the model, fixed when the model is trained, and not adjustable at scoring time.

§5Related concepts

  • Models — traits are declared on a model.
  • Score card — how per-trait scores are rendered together.
  • Composite — the single aggregate across traits.
  • Trait discovery — how traits are proposed when they are not declared in advance.
Scores are approximate — not a substitute for human judgment.