Skip to content

Similarity Dimension Rubric

Classification artifact — instantiates Prototype-Centered Category Modeling

Names the dimensions along which resemblance to the prototype is judged, sets their weights (which can shift by context), and fences off the dimensions that must never count.

A Similarity Dimension Rubric makes explicit what "similar to the prototype" actually means. It enumerates the specific dimensions along which resemblance is assessed, assigns each a weight (which may change with context), and — the part that distinguishes it — names the dimensions that are off-limits. Its distinctive move is decomposition: it breaks the fuzzy global question "how typical is this?" into named, scorable axes, converting an intuition into an auditable instrument. It is the only mechanism in the set that governs which features may legitimately count toward membership, which is what lets a graded category be inspected, taught, and contested rather than merely felt.

Example

A data team is hiring an analyst. "Hire people like our best analyst" would quietly import whatever their best analyst happened to be — including irrelevant things. Instead they write a rubric. The dimensions of resemblance are SQL fluency, statistical reasoning, clarity in communicating findings, and comfort defending numbers to skeptical stakeholders. The weights are set by the role: this team's analysts spend more time defending results than building models, so "stakeholder skepticism" is weighted up and pure modeling weighted down. And there is an explicit exclusion list — university prestige, candidate name, years since graduation — features that correlate with the incumbent prototype but must not enter the judgment. A candidate's fit is now scored on legitimate, weighted axes, and the fence keeps "resembles our prototype" from decaying into "resembles our prototype's demographics."

How it works

  • Decompose the resemblance. Turn one global similarity into named dimensions, each with scoring anchors, so judgments can be compared and defended.
  • Weight by context. The same category weights its dimensions differently in different settings; the rubric records the weights and the conditions under which they change.
  • Fence the illegitimate. Explicitly list the dimensions that must not enter the judgment — proxies for protected attributes, features irrelevant to the category's purpose. The fence is the guardrail, stated up front rather than hoped for.
  • Choose the combination rule. State whether a high score on one axis can offset a low score on another, or whether some dimensions must pass on their own.

Tuning parameters

  • Dimension granularity — a few broad axes versus many fine ones. Fine axes are precise but expensive and tend to double-count correlated features.
  • Weighting scheme and context-sensitivity — fixed weights versus per-context weights. More context-sensitivity fits reality better but is harder to keep consistent and auditable.
  • Exclusion strictness — how aggressively proxies are fenced off. Strict fences protect fairness but can discard genuinely predictive signal — a real, unavoidable tension.
  • Anchoring — whether each dimension carries scored example anchors (borrowed from an example deck) or only bare labels; anchors improve rater agreement.
  • Compensatory versus conjunctive — whether dimensions trade off or some are must-pass gates.

When it helps, and when it misleads

Its strength is that it turns "I know it when I see it" into an inspectable, teachable, contestable instrument; the weights make context-dependence explicit instead of hidden, and the exclusion list is a concrete, structural fairness lever rather than a good intention. Its failure modes are subtle. A rubric can look rigorous while over-weighting a salient but irrelevant dimension — false precision laundering a resemblance judgment that is really the representativeness heuristic in a spreadsheet.[1] Fencing off one proxy can merely hide it, if a correlated feature is left in for the judgment to lean on. The classic misuse is writing the rubric after the winner is chosen and reverse-engineering the weights to justify the decision. The discipline that keeps it honest is to fix the dimensions and weights before seeing the cases, to validate them against real outcomes, and to audit for the proxies the fence missed.

How it implements the components

Similarity Dimension Rubric fills the components that define and govern the resemblance judgment:

  • similarity_dimension_map — the rubric is the enumerated map of the dimensions along which similarity to the prototype is measured.
  • context_weighting_rule — the per-dimension weights, and the rule for how they shift by context, are stated in the rubric.
  • bias_and_stereotype_guardrail — the explicit exclusion list fences off the dimensions that would let the prototype harden into a stereotype.

It does not supply the boundary/minimal-pair cases — those are the Near-Miss Comparison Set — nor does it run the classification: the Nearest-Neighbor or Exemplar Classifier applies the metric this rubric defines, and the Typicality Rating Exercise supplies the anchors and graded scale.

Notes

The rubric defines the metric every other mechanism silently relies on — the classifier's distance function, the embedding map's axes, the "one deciding feature" of a near-miss pair. Because those tools inherit whatever notion of similarity is in force, making it explicit here is precisely what lets them be audited rather than trusted.

References

[1] The representativeness heuristic (Tversky & Kahneman): judging category membership by resemblance to a mental prototype, which can override base rates and legitimate evidence. A dimension rubric disciplines it by fixing in advance which resemblances count and how much they weigh. ↩