Category Language Audit¶
Method — instantiates Abstraction–Substrate Traceability Guardrail
Reviews labels, reports, forms, dashboards, and interfaces for wording that turns classifications into essences or facts beyond their warrant.
A Category Language Audit is a systematic sweep of the words an organization uses to name its abstractions, hunting for the grammar that quietly promotes a classification into an essence. Its single idea is that reification often happens in a verb: the difference between "this account scored high-risk on the fraud model" and "this account is fraudulent" is not analytical, it is grammatical, and the second sentence has smuggled in a claim the model never made. The audit does not touch the model, the data, or the interface plumbing — it works purely on the language layer, finding every place where a copula ("is"), a nominalization ("a low performer"), or a dropped qualifier has converted a measured, bounded, revisable estimate into a flat statement of fact, and rewriting it to carry its warrant. It is the mechanism that keeps the abstraction and the substrate distinguishable in speech.
Example¶
A large employer runs annual performance calibration. The talent system emits a distribution and, for the bottom band, prints "Low Performer" on the review form, the succession dashboard, and the automated email to managers. A Category Language Audit walks each surface. On the review form it finds "Employee is a low performer" and flags the copula: the rating is a relative ranking against this cycle's rubric and this manager pool, not a property of the person. It rewrites to "Rated in the lowest band this cycle on the four calibrated competencies." On the succession dashboard it finds the standing label "Low Performers (14)" — a noun that has turned a cycle-specific score into a caste — and changes it to "Rated lowest band, 2024 cycle." In the manager email it finds "flagged as a flight risk," an inference presented as a status, and rewrites it to state what was actually measured and for what use.
The setup takes a style guide and a reviewer with veto power; the outcome is that no downstream reader can mistake a ranking for a diagnosis, because the language itself now names the measurement, the period, and the scope every time the label appears.
How it works¶
The audit is language-first and inventory-driven:
- Enumerate the surfaces. List every place the contested label is rendered as words — forms, letters, dashboards, dropdown values, tooltips, exported reports, chatbot replies.
- Apply a de-reifying rewrite rule. For each occurrence, test the wording against a small grammar of failure: essentializing copulas ("is a…"), standing nouns that outlive the measurement, dropped hedges, and inferences dressed as observations. Rewrite to the form "[measured thing] estimates/rated [X] for [use], as of [period]."
- Restate the claim once, canonically. Alongside the rewrites, produce the single authoritative sentence of what the label claims — its warrant — so every surface can point back to it rather than reinventing a caption.
- Set the guard, not just the fix. Add the rule to the style guide and to templates so newly generated text inherits it, rather than re-drifting after the one-time cleanup.
The distinguishing discipline is that it operates on text, not on data or display timing: it makes the words honest wherever they are written.
Tuning parameters¶
- Rewrite aggressiveness — from light hedging ("estimated") to full restatement with period and scope. Heavier rewrites are unambiguous but verbose; caption real estate and reader fatigue push back.
- Surface coverage — whether the audit reaches only high-stakes documents or every tooltip and dropdown. Wider coverage closes leakage but multiplies the maintenance surface.
- Canonical-claim granularity — one warrant sentence per label versus per use-context. Finer granularity is more accurate but harder to keep synchronized.
- Enforcement point — advisory style guide versus a template/lint gate that blocks non-compliant strings. Hard gates hold the line but can misfire on legitimate prose.
When it helps, and when it misleads¶
Its strength is that it attacks reification where it actually spreads — in ordinary language and institutional memory — and it is cheap relative to model work: changing "is" to "scored" costs nothing and removes a whole class of misreading. It is the direct counter to essentialist drift, the slide by which a descriptive category hardens into a supposed inner nature.[n1]
It misleads when tidy wording is mistaken for a fixed substrate. An audit can make a badly-warranted score read beautifully — "estimated, as of this quarter, for this use" — while the estimate underneath is garbage; careful captions can launder a bad abstraction as easily as they can honor a good one. It also decays: rewrite once and the next template, the next vendor screen, the next generated summary reintroduces the copula. The classic misuse is treating the audit as a one-off copy edit rather than a standing rule wired into templates. The guarding discipline is to fix the rule and the template, not just the string, and to keep the audit strictly about language — never letting a clean caption stand in for actually checking whether the label is valid.
How it implements the components¶
Category Language Audit fills the language-and-claim components — the part of the guardrail that lives in words:
de_reification_language_rule— it is the rule: a grammar of essentializing patterns plus rewrites that keep classification-as-estimate distinct from fact, enforced across every text surface.representational_claim_statement— it produces the canonical one-sentence statement of what each label claims (measured thing, use, period, exclusions) that the rewrites reference.
It does not display those caveats at the decision moment (boundary_disclosure_surface) — that is Point-of-Use Reification Warning, whose interface card the audit's wording feeds; nor does it monitor or adjudicate when a label diverges from reality (substrate_divergence_channel, proxy_optimization_monitor) — those belong to Counterexample Case Review and Proxy Drift Dashboard.
Related¶
- Instantiates: Abstraction–Substrate Traceability Guardrail — the audit keeps the abstraction and its substrate distinguishable in language.
- Sibling mechanisms: Point-of-Use Reification Warning · Model Card or Datasheet Linkage · Counterexample Case Review · Map–Territory Review Checklist · Proxy Drift Dashboard
Editorial Notes¶
Form Classification¶
Form family: Assessment, Review & Assurance
Rationale: Reviews labels, reports, forms, dashboards, and interfaces for wording that turns classifications into essences or facts beyond their warrant, making its operative form a bounded evaluation of existing evidence or work that produces a finding or disposition.
Independent corroboration: The frozen evidence defines Category Language Audit as 'Reviews labels, reports, forms, dashboards, and interfaces for wording that turns classifications into essences or facts beyond their warrant', so its operative form is Assessment, Review & Assurance.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Linguistics & Semiotics
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Universal
Rationale: Linguistic analysis supplies close review of labels and grammatical constructions that turn contingent classifications into essential, factual-sounding predicates.
Related originating lineages:
- Psychology — Psychological essentialism explains why category labels are read as hidden, stable natures.
- Sociology & Anthropology — Critical classification research supplies attention to status and power effects of institutional wording.
Review resolution: Linguistics and discourse analysis are primary because the audit examines copulas, nominalizations, modality, and labels that reify a classification. Psychology contributes essentialism research and sociology contributes institutional classification effects, making the generalized audit a cross-disciplinary synthesis with universal applicability.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] Psychological essentialism is the tendency to treat a category label as pointing to a hidden, unchanging inner nature shared by its members. It is why "is a low performer" reads as a claim about the person rather than about a measurement, and it is exactly the slide the audit's rewrite rule is built to interrupt. ↩