Tiers

A tier is a categorical label assigned to a score based on where the score falls among the trait's trained boundaries. Four tiers are defined: Strong, Solid, Developing, Weak.

§1Definition

A tier is the label returned alongside a trait score. The label is determined by comparing the score against three boundaries derived from the trait's training distributions. The four tiers partition the score axis as follows.

TierConditionSource boundary
Strongscore ≥ strongUpper quartile of the positive training distribution
Solidsolid ≤ score < strongLower quartile of the positive training distribution
Developingdeveloping ≤ score < solidUpper quartile of the negative training distribution
Weakscore < developingWithin or below the negative training distribution

Table 1. Tier labels and the score conditions that assign them. Boundary names (strong, solid, developing) refer to the trained thresholds described on breaks.

§2Mechanism

§2.1Boundaries are per-trait

The three boundaries are computed during training from the quartiles of the trait's positive and negative sample distributions. Because each trait has its own training distributions, each trait has its own boundaries. A score of 70 on one trait may fall in the Solid band; a score of 70 on another trait may fall in the Developing band.

Weak 28
Developing 55
Solid 72
Strong 89
Figure 1. Four representative scores and the tier each resolves to under one trait's trained boundaries. Another trait in the same model may assign a different tier to the same score.

§2.2Label assignment

At scoring time, the score is compared against the trait's boundaries from highest to lowest. The first condition the score satisfies determines the tier. No tier is computed for the composite — the composite is reported as a score only.

§3Interpretation

Each tier expresses where the score sits relative to the training distributions that defined the trait.

TierWhat it says about the score
StrongIn the upper quartile of the positive training distribution. At or above the strongest examples the trait was trained on.
SolidWithin the positive training distribution's inter-quartile range. Comparable to the bulk of training examples that exemplified the trait.
DevelopingBetween the positive and negative distributions. Above the bulk of examples that did not exemplify the trait, below the bulk that did.
WeakAt or below the upper quartile of the negative training distribution. Within the range of examples that did not exemplify the trait.

§4Edge cases

§4.1Withheld label

When the positive and negative training distributions overlap enough that the boundaries collapse or invert, no tier label is assigned. The score is returned with a null tier. The same condition produces low confidence, so the two signals appear together.

§4.2Developing and moderate confidence

Scores in the Developing band fall between the two training distributions, where the system has less direct training signal than in either distribution's bulk. Developing scores are therefore always reported with moderate confidence. All other tiers pair with high confidence for well-separated training distributions.

§5Related concepts

  • Breaks — the trained boundaries that define the tier cutoffs.
  • Headroom — the score distance from the current tier to the next tier's boundary.
  • Confidence — the reliability signal paired with every tier assignment, and the reason a tier may be withheld.