Granularity Tuning Rubric¶
Decision rubric — instantiates Equivalence-Relation Refinement and Coarsening
A scored comparison of candidate class structures on actionability, error cost, and maintenance burden, converting the split-or-merge choice into an explicit weighing rather than a hunch.
Evidence that a class is failing does not tell you what to do about it — you can split it, merge it, redraw it, or leave it and absorb the cost. A Granularity Tuning Rubric is the scoring instrument that decides among those options on stated criteria: it lays candidate class structures side by side and rates each on how well it serves action, what its errors cost, how interpretable and fair it is, and how much it burdens whoever maintains it. Its defining move is that it weighs and chooses rather than diagnoses or records — it turns "these classes feel too coarse" into a criterion-by-criterion comparison whose winner is the granularity the task actually justifies. The rubric's animating ratio is the archetype's own: does the difference in action between two classes exceed the cost of keeping them distinct?
Example¶
A company is redesigning its job architecture and must decide how many engineering levels to define. Three candidate structures are on the table: a coarse four-level ladder, the current eight-level one, and a fine twelve-level proposal. The compensation team scores each on the rubric's criteria. Actionability: do adjacent levels lead to genuinely different pay, scope, and promotion decisions? The twelve-level scheme has three pairs that resolve to identical pay bands — distinctions that change no decision. Error cost: the four-level scheme lumps senior and staff engineers together, and mis-leveling there is expensive in attrition. Maintenance burden: twelve levels mean twelve rubrics to calibrate every review cycle. Interpretability and fairness: can an engineer tell why they're at one level and not the next, and do the boundaries disadvantage any group? Scored out, the eight-level ladder wins — but the rubric also says merge the three redundant top pairs of the twelve-level idea and split one overloaded mid-level, because that is where action-difference and distinction-cost diverge most. The decision is now a defensible comparison, not the loudest VP's preference.
How it works¶
The rubric's distinguishing method is multi-criteria scoring of whole candidate structures. Rather than judging one boundary at a time, it evaluates each proposed granularity as a package against a fixed criterion set — actionability, error cost, interpretability, fairness, traceability, maintenance burden — usually weighted to reflect what this domain cares about. The scores make the tradeoffs legible: a finer structure almost always wins on precision and loses on burden, and the rubric forces that exchange into the open instead of letting one dimension dominate silently. Its output is not just a winning structure but a split/merge verdict per class: the criteria pinpoint exactly which distinctions earn their keep (adjacent classes with different action, tolerable maintenance) and which do not (identical action, high upkeep), yielding the operative rule for where to cut. It consumes evidence — ideally the failure patterns a confusion matrix surfaced — but its own job is the judgment.
Tuning parameters¶
- Criterion weights — how much actionability counts against maintenance burden, fairness, interpretability. The weights are the value judgment; publishing them is what makes the verdict contestable rather than arbitrary.
- Action-difference threshold — how different two classes' downstream actions must be to justify keeping them apart. Set high and the rubric coarsens aggressively; set low and it tolerates fine distinctions.
- Cost model resolution — whether maintenance and error costs are rough tiers or quantified figures. Quantifying sharpens comparisons but risks false precision on soft costs like user confusion.
- Candidate slate — how many alternative structures are scored. A wider slate explores more of the space but dilutes attention; too narrow and the rubric merely ratifies the status quo.
- Tie-break rule — what decides when two structures score close: default to the coarser (simplicity) or the incumbent (stability).
When it helps, and when it misleads¶
Its strength is making the split-or-merge decision explicit and reviewable — the weights, the criteria, and the action-versus-cost ratio are all on the page, so a granularity choice can be argued on its merits and revisited when the merits change. It is the antidote to relations that drift by whoever-shouts-loudest, and it directly encodes the archetype's core discipline: keep a distinction only when it changes action by more than it costs.[n1]
It misleads when its tidy scores lend false objectivity to what are really value choices buried in the weights — a rubric can make an arbitrary preference look rigorous. It is also gameable: pick the criteria or weights that favor a predetermined answer and the "comparison" is theater. And it can over-reach into fragmentation if actionability is weighted so heavily that every faint difference clears the bar. The discipline is to publish the weights and criteria so they can be challenged, to sanity-check the winning structure against real cases rather than trusting the score, and to keep the cost side honest so simplicity has a real vote.
How it implements the components¶
The rubric fills the decision-and-valuation components — the ones a scoring instrument that chooses can own:
class_granularity_target— its output is the chosen granularity: how many classes at what resolution the task justifies, selected as the top-scoring candidate structure.cost_of_distinction_model— the maintenance-burden and error-cost criteria are exactly a model of what each distinction costs, weighed against the action it buys.split_merge_decision_rule— the criterion comparison yields the operative rule for which classes to split and which to merge, per the action-difference-versus-cost ratio.
It does not gather the empirical evidence of where classes actually fail — within_class_homogeneity_test, between_class_separation_test, empirical_cluster_signal — that is the Classification Confusion or Error Matrix, which the rubric consumes; and it does not preserve continuity once the choice is made — old_new_class_mapping — that is the Equivalence Class Crosswalk Table.
Related¶
- Instantiates: Equivalence-Relation Refinement and Coarsening — the deciding instrument that chooses the target granularity and the split/merge verdict.
- Consumes: Classification Confusion or Error Matrix supplies the empirical failure evidence the rubric scores against.
- Sibling mechanisms: Classification Confusion or Error Matrix · Hierarchical Taxonomy or Lattice · Periodic Relation Granularity Review · Golden Case Benchmark Set · Split/Merge Change Log
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: The rubric computes an explicit comparison of candidate class structures on actionability, error cost, and maintenance burden.
Nearest alternative: Decision, Gate & Allocation — The scores support a split-or-merge choice, but the operative mechanism is the comparative analysis rather than the final authority disposition.
Review outcome: Adjudicated after independent review; high confidence.
Origin Attribution¶
Primary origin: Library & Information Science
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Classification theory directly confronts when categories should split or merge for useful retrieval.
Related originating lineages:
- Operations Research — Multi-criteria decision analysis supplies explicit scoring and comparison of competing category structures.
Review resolution: NISO controlled-vocabulary standards explicitly govern construction, testing, maintenance, and management of taxonomies and thesauri, where split-versus-merge granularity is a core classification concern. Library_information_science is therefore primary. Operations research contributes the explicit multi-criteria scoring of actionability, error cost, and maintenance burden. The scored tuning rubric is an encyclopedia synthesis.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Researched adjudication after independent review; high confidence.
Sources consulted:
- https://www.niso.org/publications/ansiniso-z3919-2005-r2010 — ANSI/NISO standard for constructing and managing taxonomies and controlled vocabularies.
- https://www.nasa.gov/reference/6-8-decision-analysis/ — NASA decision-analysis guidance for criteria-by-alternative matrices.
Notes¶
[n1] Multi-criteria decision analysis (MCDA) is the family of methods for choosing among options scored on several weighted, often conflicting criteria. A granularity rubric is MCDA applied to candidate class structures; its honesty rests, as MCDA's always does, on making the weights explicit and contestable. ↩