Terminology governance · a worked audit excerpt
What audit output looks like.
The vocabulary-layer essay argues that terms are retrieval keys and idiom is a tax. An argument deserves evidence, so here is an excerpt from a real terminology audit — run with the method's own rubric over the author's own five-product operations suite (a regulated manufacturing niche): twenty client-facing assets selected by reach × authority × longevity, the full forty-entry term-inventory scan, and consistency measured across every recurring concept.
Findings are anonymized the way the rubric requires — author-blind, severity attached to assets, assets cited by code. The corpus is private, so the anonymization is real.
The three rates.
A terminology audit reads out in three numbers. First, S1 count — charged or exclusionary terms on client-visible surfaces. Here: zero, across all forty inventory entries and every surface. A clean result is a real and creditable result, and it is stated plainly rather than dressed up.
Second and third are where this corpus pays: seven of seven recurring concepts — every concept appearing in three or more assets — carry two or more names, with canonical-name shares ranging 50–88%. No freight problem; a pure retrieval problem. Every split concept is a split retrieval key: the same question phrased with the minority term ranks the wrong assets, for humans and machines alike.
Three findings, severity-rated.
Finding 1 · S2 — the entity split: customer / client / account. One relationship entity, three names, seven of twenty assets. "Account" leads with a 71% share, but "customer" and "client" both survive on client-visible surfaces across products that share a user base — and the winning name is itself already claimed: the same word names an unrelated operational concept elsewhere in the corpus, a collision the terminology standard must settle in the same decision. Proposed fix: canonical account for the relationship entity, with the colliding sense renamed or always qualified; customer only for commerce contexts (a person buying at the counter); client deprecated.
Finding 2 · S2 — the production-unit split: batch / lot / run. Ten of twenty assets, three names, 70% canonical share. The team's own engineering log records a naming collision on "batch" being deliberately dodged once before — this drift has already billed the team a decision. Proposed fix: canonical run for a production unit; lot reserved for contexts where the regulator's own form says lot; batch deprecated.
Finding 3 · S2 — the destructive-verb split: void / delete / remove. Three verbs for irreversible-looking actions, 50% share. The product's actual behavior is auditable reversal — and the one verb that says so, "void," appears exactly once. Proposed fix: void for auditable reversals; remove for detaching an association; delete banned from client surfaces unless data is truly destroyed.
Severity here is the rubric's exposure × freight scale: none of these carry freight, all of them sit on client-visible surfaces — S2, the fix-this-quarter tier.
The method, honestly scoped.
The rubric is the three-week pilot: twenty assets, two reviewers, findings that cite a forty-entry term inventory by number so reviewers flag identically. This excerpt ran the scan and scoring stages in a single pass — the full pilot adds a second reviewer and a reconciliation step, and disagreement between reviewers is itself a finding: it marks the territory a terminology standard must decide.
One asset cluster was withheld from this excerpt because its vocabulary would identify the corpus's niche precisely — the anonymization is allowed to cost a finding. What remains is representative: no charged language, universal concept drift, and three fixes any curator could ship in a week.
Run this on your own corpus
The rubric, the forty-entry inventory, the scoring sheet, and the full pilot blueprint ship in the toolkit's Terminology Governance Pack. The essay that makes the argument is free: the vocabulary layer.