Tiers, breaks & headroom

How a trait score is graded: plain-language tiers, the trained breaks that define them, and the headroom to the next one.

A tier is the plain-language label attached to a trait score — Strong, Solid, Developing, or Weak. Tiers come from breaks, the trained thresholds that divide a trait’s scale, and headroom measures how far a score sits from the next break above. Together they turn a raw 0–100 number into something a reader can act on.

Definition

A tier is assigned by where a score falls relative to its trait’s breaks. The label appears as detail.<trait>.label.

Tier	Boundary	Meaning	Typical action
Strong	at or above the strong break	In the positive cluster’s top quartile	A strength — focus elsewhere
Solid	at or above the solid break	Inside the positive cluster	Good — minor refinement possible
Developing	at or above the developing break	Between the two clusters	Improving — keep pushing
Weak	below the developing break	Inside the negative cluster	Needs significant work
null	—	confidence is low	Can’t grade reliably

Breaks

Breaks are the three thresholds — developing, solid, and strong — each tied to a quartile of the training distribution, so the boundaries reflect real data rather than fixed cutoffs. They appear as detail.<trait>.breaks.

Each break is computed from the score quartiles of the positive and negative training samples:

developing = neg_p75
solid      = max(pos_p25, developing)
strong     = max(pos_p75, solid)

Break	Tied to	Crossing it means
developing	upper edge of the negative cluster (neg_p75)	Left the negative cluster
solid	lower edge of the positive cluster (pos_p25)	Entered the positive cluster
strong	upper quartile of the positive cluster (pos_p75)	Reached the top of the positive cluster

The nested max(...) keeps the breaks monotonic — developing ≤ solid ≤ strong — even when the two clusters overlap. When they overlap, the breaks collapse toward each other, the Developing band shrinks, and scores in the overlap region get low confidence and no tier.

Headroom

Headroom is the distance, in score points, from a score to the next break above it. It appears per trait as detail.<trait>.headroom and, for the card as a whole, as the top-level headroom.

Current tier	Headroom measures distance to
Weak	the developing break
Developing	the solid break
Solid	the strong break
Strong	0 — already at the top

Each tier has exactly one natural next milestone, so headroom is always a single, unambiguous “distance to the next level” — a client can show it without knowing which threshold it targets. The card’s top-level headroom is the largest per-trait headroom: the bottleneck trait, the one with the most ground left to cover.

Interpretation

Tier and confidence move together for a well-separated trait:

Tier	Confidence
Strong	high
Solid	high
Developing	moderate
Weak	high
null	low

Developing generally carries moderate confidence — built-in honesty about a transitional score. Gate automated actions on Strong or Solid; treat Developing as content that warrants human judgment.

Breaks are per-trait and re-derived on every training run, so the same raw score can mean different things across traits or across versions. Read the tier and the headroom, which already fold the breaks in; the raw thresholds are mainly for diagnostics and for drawing the scale. Small headroom means a minor edit could change the tier — a score two points below the strong break is one improvement away from Strong — and the composite’s headroom points at the trait to work on first.

Edge cases

• When confidence is low, both the label and the headroom are null — the model withholds a grade rather than guess.
• Overlapping clusters collapse the Developing band — by design, content that lands in the overlap is left ungraded.
• Headroom 0 and null are distinct: 0 is a graded score already at the top tier; null is a withheld grade because confidence was too low.
• Every break is a real cluster quartile; none is an arbitrary cutoff.

Next

Confidence

Why some scores get no tier at all.

Composite

How per-trait scores combine into one number.