Skip to content

Label Taxonomy Cleanup

Portfolio-cleanup method — instantiates Signal Value Preservation

Rationalizes an overgrown set of labels — merging duplicates, splitting overloaded tags, retiring dead ones — so each surviving label again makes one clear, distinct claim.

Label Taxonomy Cleanup takes a sprawling, overlapping set of labels or tags and collapses it back into a coherent one: merge synonyms, split any single label forced to carry several incompatible meanings, and drop tags nobody acts on. Its distinguishing idea is that it operates on the whole portfolio of labels at once, not on a single signal's evidence or price — it restores diagnostic separation between labels, so that applying one actually tells the receiver something the others don't. When a hundred near-duplicate tags all sort-of-apply, none of them informs; cleanup makes each label claim one thing again.

Example

An issue tracker has drifted to 240 labels. Over the years engineers added urgent, high-priority, P1, important, needs-attention, and hot — six ways to say one thing — while bug got used for defects, missing features, and support questions alike. Triagers had stopped filtering by label because no label predicted anything. A cleanup pass inventories all 240 with a card-sort, merges the six priority synonyms into a single priority:P1, splits the overloaded bug into defect / feature-gap / support-question, and retires ~90 tags used fewer than three times. The set drops to about 40, each with a written one-line definition — a controlled vocabulary.[1] Now filtering by defect returns defects, and the labels carry information again.

How it works

The distinctive move is to operate on the set: inventory every label and its actual usage, locate the overlaps, overloads, and dead entries, then merge, split, or retire and write a single clear definition per survivor. It does not measure receivers or set evidence bars for applying any one label — it restructures the vocabulary itself so that the labels are mutually distinct and each maps to exactly one meaning.

Tuning parameters

  • Merge aggressiveness — how readily near-synonyms collapse. Aggressive merging sharpens the set but can erase real distinctions users relied on.
  • Split threshold — when one label's multiple meanings warrant separate tags. More splitting improves specificity but regrows the set.
  • Retirement cutoff — the usage or age level below which a label is dropped. Lower keeps rarely-used-but-meaningful tags; higher prunes harder.
  • Definitional rigor — whether each survivor gets an enforced written definition or just a name. Definitions prevent re-sprawl but cost upkeep.
  • Migration handling — how existing items tagged with retired labels are remapped; sloppy migration loses history.

When it helps, and when it misleads

Its strength is that it restores between-label separation cheaply, without touching evidence rules or prices, so filtering becomes informative again. Its failure mode is the one-off: the set re-sprawls within a year unless something governs the minting of new labels. The classic misuse is over-merging to a tidy-but-useless handful that no longer captures real categories — tidiness mistaken for information. The discipline that guards against this is to pair the cleanup with a creation gate (who may mint a new label), re-run it periodically, and judge success by whether labels regained predictive and filtering value rather than by how far the count dropped.

How it implements the components

  • signal_portfolio_map — the inventory step maps the entire label set and its overlaps; the cleanup operates on that map.
  • signal_claim_definition — merging and splitting resolve what each surviving label claims, one clear meaning per mark.
  • signal_redesign_path — merge / split / retire is a structural redesign of the label system.

It does not measure whether receivers still respond (habituation_and_adaptation_monitor) — the Receiver Response Dashboard and Receiver Fatigue Survey do — and it does not set the evidence bar for applying any single label (evidence_and_specificity_rule), which is the Signal Issuance Rubric.

Notes

Cleanup, recalibration, and retirement are easy to confuse. Label Taxonomy Cleanup fixes the claims and overlaps of a set of distinct labels; Severity Tier Recalibration re-scales the thresholds of one graded tier system; Signal Retirement Registry tracks each signal's lifecycle and drives end-of-life. A cleanup pass is a good moment to hand expired labels to the registry rather than merely deleting them.

References

[1] A controlled vocabulary — a fixed, defined set of allowed terms — is the standard information-architecture remedy for tag sprawl. Its whole purpose is that each term is distinct and defined, which is exactly the diagnostic separation this mechanism is trying to recover. ↩