Score types

Each trait has one of four types, fixing its geometry, score range, and native metric.

Every trait has a type that fixes three things at once: the geometry it uses to judge content, the range its native score lives in, and the question that score answers. The universal 0–100 score is the same across all types; the type is what gives each trait its own native signal.

Definition

A trait type is a triple — (geometry, score range, native metric). Four types are defined; a trait declares exactly one.

Type	Question	Geometry
topic	Is this about X?	Cosine similarity to a learned model of the topic.
spectrum	Where on this scale?	Projection onto a learned pole-to-pole axis.
claim	Does this confirm a statement?	A relevance subspace combined with an agreement axis.
outlier	Is this unusual for a cohort?	Mahalanobis distance to a reference cohort.

Mechanism

Each trait reports two score channels, both returned in the score card:

• Universal score — a value in $[0, 1]$, reported 0–100. This is the cross-type display number: it drives the composite, the tiers, and anything that compares or combines traits.
• Native score — one or more typed fields, named for the geometry’s own question. Native recovers the signal the universal score saturates or collapses. It is reported in the field’s own range, never rescaled to 0–100.

The native fields appear as detail.<trait>.native:

Type	Native field(s)	Range	Reports
topic	topicality	$[-1, 1]$	Drift toward the topic (positive) vs ambient content (negative).
spectrum	polarity	$[-1, 1]$	Which pole (sign) and how strongly (magnitude).
claim	relevance, agreement	$[0, 1]$, $[0, 1]$	Whether content is about the claim, and — given it is — whether it confirms ($>0.5$) or contradicts ($<0.5$).
outlier	cohort_similarity, anomaly_z	$[-1, 1]$, $(-\infty, \infty)$	Raw similarity to the cohort, and a standardized standoff that ranks content deep in the tail.

Interpretation

• Use the universal score for display, gating, and any cross-trait reasoning — it is the only channel that composes.
• Use the native score for ranking and diagnosis, where the universal score saturates. Two clearly anomalous texts both read near 100 universal; their anomaly_z values still separate them.
• A claim universal score near the middle is ambiguous on its own — off-topic and on-topic-but-neutral both land there. The (relevance, agreement) pair disambiguates.
• Never compare scores across types.

Edge cases

• Moderation is packaged classification — a configuration of these primitives, not a fifth type.
• A single-pole spectrum trait derives its opposite pole automatically, so it still projects onto a full axis.

Next

Traits

What a trait is and what it measures.

Score card

Where these scores are returned.