— the wiki’s own epistemology, the consolidating genus behind method-risk R21. The reasoning is the fabric’s; the one source quotation (EFSA, in the UL instance and the corrective) is an illustration of a body naming the error, not a derivation of it.
A quantity or label defined DESCRIPTIVELY or operationally is silently promoted to a NORMATIVE or causal claim. A percentile becomes a target; a deficiency floor becomes an adequacy goal; a population ratio becomes an individual verdict; a manufacturing class becomes a health judgment; a statistical threshold becomes important. This is the is-to-ought / measure-to-meaning jump — Ryle’s category mistake, Hume’s is/ought — applied to clinical and nutritional numbers. The number is not wrong; it is being read as a kind of thing it is not.
The instances — each a type-pair
Each row is one measured/defined object (the is) promoted to a norm (the ought):
-
RDA — a deficiency floor (covers 97.5% of the population’s requirement; a population, deficiency outcome) promoted to an adequacy target. It is deployed as one: food-label % Daily Value and 100% of your needs frame it as a goal to hit. A floor presented as a target systematically misleads anyone with an objective above deficiency (muscle, healthspan, performance) — for whom 100% RDA reads as done when it is the start -> Deficiency Repletion vs Enhancement.
-
Biomarker reference interval — the descriptive central 95% (2.5th-97.5th percentile) of some population, promoted to normal = healthy. This species has its own sub-mechanism, below.
-
5 a day — a population dose-response floor over a heterogeneous basket, promoted to a per-person target with interchangeable units. The fabric already critiques this: a cabbage portion is not grapes is not juice -> Plant Foods.
-
Avoid processed — a manufacturing class (NOVA) promoted to a health verdict. The health variance within a NOVA class plausibly swamps the variance between classes, so the label does not carry the outcome — though the fabric holds this as a qualitative judgment, rarely measured, a claim about what would be found rather than a reported statistic -> Is the Food Category Doing Any Work.
-
Dietary cholesterol — a food’s cholesterol content (a compositional descriptor, ~180 mg per egg) promoted to a serum-cholesterol / CVD effect (a causal claim): eggs are high in cholesterol therefore raise blood cholesterol and heart risk. The dietary->serum transmission is weak for all but a minority of “hyper-responders” (absorption and endogenous synthesis compensate), so the second step is the descriptive->causal jump. The right-typed instrument is the evidenced outcome dose-response — which is near-null (Godos 2020: moderate intake CVD SRR 0.95, CIs kiss 1.00) — not the cholesterol content, and the causal lever is apoB particle number reached by other routes -> Eggs Dietary Cholesterol and Cardiovascular Risk, LDL ApoB and Cumulative Exposure.
-
BMI — a population screening ratio promoted to an individual adiposity/health verdict.
-
Significant — a statistical threshold promoted to important / real / large.
-
Evidence-based — a method descriptor promoted to true.
-
Tolerable Upper Intake Level (UL) — a risk-assessment threshold (the intake below which no adverse effect is expected — descriptive) read as a target (an intake to aim at — normative). The two differ in kind, and EFSA 2022 on dietary sugars is the rare case of a body naming the error to pre-empt it: «The Panel wishes to clarify that a UL is not a recommended level of intake» (European Food Safety Authority, 2022). It runs in both directions on sugars: EFSA could set no UL («as low as possible»), so reading that as softer than WHO’s
<10%target — or reading WHO’s<10%as a safe threshold — both promote one type into the other -> Free Sugars Intake, Which Objective Moved This Recommendation. A body that produces the descriptive object and explicitly withholds the normative one is the constructive corrective enacted (below).
Calibration provenance — the reference-interval sub-mechanism
The reference-interval instance carries a mechanism the others share in weaker form: a number means only what its calibration population and purpose support, and both are routinely mismatched to how it is used. The population that gets a marker measured is not representative:
- Sick-enriched. Indirect / data-mined intervals (built from the lab’s own tested database) are enriched for symptomatic, worked-up, diseased people — you get tested because something prompted it.
- Self-selected. Even direct healthy-volunteer intervals carry volunteer self-selection and unscreened subclinical disease.
- Missing whole strata (the streetlight, applied to calibration). Entire strata are structurally absent because there is no clinical indication to test them — a young-adult LDL reference range is either extrapolated or built on a doubly-selected (young AND indicated, e.g. familial hypercholesterolaemia) sample. That is the insufficient-evidence state, not a known range. And the dark region is where the decision has most leverage: LDL is a cumulative lifetime exposure, so the un-measured young window is where modifiable area-under-the-curve is largest -> the streetlight is darkest where it matters most -> LDL ApoB and Cumulative Exposure.
- Frequency-distorted. The most-monitored patients contribute the most measurements (measurement-weighted is not person-weighted). This is a real artifact, but its direction is treatment-specific, not simply inflate toward sick: it inflates for markers monitored at their disease value (HbA1c in diabetics, creatinine in CKD), and can deflate for treated-to-target ones (LDL, glucose — frequent testers are largely on statins / being titrated, pulling naive data-mined ranges down; proper methodology excludes lipid-lowering-therapy patients). Reason the sign per marker; do not assume it.
Directional consequence. For markers with an evidenced marker-to-hard-outcome transmission (LDL/apoB -> ASCVD via genetics/MR/RCT; fasting glucose/HbA1c; blood pressure), the risk-optimal value sits at or below the low edge of the range — so within normal range is falsely reassuring there -> Metrics for Targeted Health Guidance.
Why it happens — communicability, not usually a lie
The promotion is rarely deception; it is communicability. A single target or label deploys more easily than a floor, a distribution, or a dose-response curve — so the descriptive object is flattened into a normative one at the delivery layer, where it reaches people. This is the which objective moved this lens applied inward: communicability is the non-health objective driving the type-promotion, and the harm often lives in the gap between the expert definition and the deployed perception (the RDA is defined correctly as a floor and deployed as a target).
The corrective is CONSTRUCTIVE, not debunking
The catch is not ignore the number. It is a type-discipline in three moves:
- Ask what KIND of thing this is — descriptive or normative? A percentile, a floor, a manufacturing class, a threshold?
- Use the descriptive for description only — a reference range tells you where a population sat, not where you should be.
- Get the right-typed instrument for the decision — an evidenced causal dose-response to a hard outcome (LDL/apoB -> ASCVD, not the reference range; a target legitimate only by its evidenced transmission -> Surrogate Outcomes) or the specific mechanism/composition (the fat/salt/sugar and food matrix, not the NOVA label). A real category-error catch upgrades the instrument. The LDL case is the worked example: the right-typed instrument is CTT’s per-mmol event dose-response (monotone, no threshold, benefit below the “normal” range) plus Marston’s apoB-particle metric — not the LDL-C or apoB manufacturer reference range -> LDL Lowering and Cardiovascular Events.
The dual-use guard
— shares R21/R20’s dual-use structure with The Estimate-to-Action Gap.
That’s just a descriptive range, not a target is both the correct diagnosis and the science-sceptic’s universal dismissal of any number they dislike. The discriminator is whether the catch points to a better-typed replacement (constructive — upgrades to an evidenced dose-response or a named mechanism) or merely discards the number (evasive — the R21 motivated skip). A motivated version throws the instrument away; an honest one replaces it with the right-typed one.
- Where normal is not optimal cuts. Toward don’t be reassured for the evidenced markers (LDL/apoB/glucose/BP), NOT toward chase optimal labs for the unevidenced ones (the optimal TSH / testosterone / vitamin-D / ferritin ranges pushed without hard-outcome evidence — the surrogate-vs-target trap; hsCRP is the fabric’s worked skip-case — statins move the marker ~42-45%, but it stays a surrogate, not a proven causal lever -> Metrics for Targeted Health Guidance).
Decision relevance
- Appraise the DEPLOYED number, not just the defined one. A number’s institutional function can diverge from its definition, and the divergence is often where the harm is (the RDA-as-% -DV case).
- Where an evidenced causal dose-response exists, USE IT and let the reference range go — it sidesteps the calibration-population problem entirely. The reference range is the fallback only where no causal dose-response is held, and there its provenance problems make it a weak instrument.
- Open loop. This grades the type of a claim, never its validity against a realized outcome.