— this page is the wiki’s own epistemology. It carries no source attributions: every claim
is the fabric’s reasoning over method rules (the telos’s dose-response section) and already-held concept
pages cited by wikilink. Like The Estimate-to-Action Gap, sources: [] is legitimate here.
In this domain, the evidence structurally yields floors, ranges, directions, and harm-ceilings — but not a point OPTIMUM. A study can show below here is deficiency, this arm beat that arm, more in this range still helps, past here it harms — but it essentially cannot demonstrate this exact intake is best. The claim is epistemic (an optimum is not derivable), not biological (not there is no best dose). Three distinct structural routes each terminate in no observed peak, and they converge on one honest output: name a data-supported range or direction bounded by a floor and a harm-ceiling, and disclaim the point-optimum. No single method rule or concept page states this; it is induced across them.
Three routes to no peak — each holds for a different reason
Routes 2 and 3 are the two genuinely independent legs — each fails to yield an optimum even if the other did not hold, and neither needs the curve to be flat, so the principle does not rest on any one shape claim. Route 2 (the peak is beyond the sampled range) holds even under perfect measurement; Route 3 (noise smears a point into a band) holds even for a curve with a real interior peak. Route 1 is a demoted third — it fires only where a flat region is actually observed, and an observed flat top is often itself downstream of Route 3 (below), so it leans on the other two rather than standing beside them. The honest count is two independent routes plus one weak, largely-Route-3-downstream add-on — not three co-equal legs.
Route 3 (the master driver) — measurement noise yields both broad ranges AND nulls
Dietary exposure is measured by asking people, and the error is the domain’s binding constraint -> Measurement Error in Dietary Assessment. It is one cause with two effects: regression dilution attenuates the slope toward the null (a real gradient reads as no gradient), and imprecision widens the interval (a point reads as a broad range). So the same instrument that hides an effect also smears whatever survives into a band — you cannot read an optimum off a band. Compound many noisy exposures (a whole diet is many nutrient effects at once) and the uncertainties propagate: the aggregate is broader and more null-straddling still.
- Precisify it correctly. This is not random error that averages to a clean normal as n grows — that would buy precision. It is systematic error that does not average out (differential under-reporting, the flat-slope syndrome compressing the exposure range, correlated validation error), an irreducible noise floor no sample size overcomes — which is exactly why mega-cohorts still return broad, null-tending nutrition results.
- Guard — pull to null is the central tendency, not a theorem. Non-differential error biases toward the null only in the univariate, unconfounded case; with imperfectly-measured covariates the bias runs in any direction, and differential error can inflate or reverse an effect -> Measurement Error in Dietary Assessment. So: tends-to-broad-and-null, with real exceptions — not a law that everything collapses to zero.
Route 2 (the cleanest case) — a monotone effect has no observed peak at all
For an effect still rising at the edge of the observed data (fibre, sodium and free sugars are each monotone over their studied range — the falsified knees-and-plateaus audit, below), the apparent optimum is simply the top of the sampled range — set by the study’s often-arbitrary cut-off, not by the curve. A different study with a slightly different cut-off would report a different optimum, and you cannot extrapolate above the data. This generalizes a held rule — where a threshold appears in guidance, the first hypothesis is that it marks the edge of the evidence, not a feature of the curve (SACN’s 30 g fibre is where the confidence intervals widen, not where returns flatten) — lifting it from the deficiency threshold to the upper bound: same error, the study boundary mistaken for a target .
- Do NOT import a harm-ceiling or a diminishing return here. Both are by definition absent from a monotone effect (if you had observed one, the effect would not be monotone). The honest output for the pure monotone case is therefore the barest of all: a direction, open-topped within the data, with no evidenced stopping point — not an optimum, and not even a bounded range. This is the cleanest no-optimum case: the peak is not indistinct, it is simply beyond the data horizon, unobserved and possibly non-existent.
Route 1 (the weakest, held conditionally) — on a genuinely flat region the optimum is vacuous
Where a curve reaches a genuine plateau — muscle-protein synthesis flattens above a breakpoint estimated at 1.62 g/kg/day — no point on the plateau is a peak: past the knee, more buys nothing, so every adequate intake is equivalent and optimum is near-vacuous rather than merely unmeasured. The flat region is therefore itself the optimum, which generalizes the telos’s a minimum effective dose is a region, not a number and flat regions tolerate imprecision cheaply from the minimum-effective-dose to the optimum: a flat-topped optimum simply IS a region -> The Estimate-to-Action Gap (the region, not a point step). Two cautions keep this honest: the breakpoint’s own location is uncertain (95% CI 1.03-2.20) — that interval is on where the knee sits, not the width of the flat band, so it is not itself the plateau; and the plateau is open-topped, bounded above only by the harm-ceiling or a logistical cap, not by an optimum.
-
The bare target aim for 1.62 is borderline-false, and fails in three distinct ways. Stated as a point it launders a barely-identified knee into a precise target — it (i) asserts an identification precision the CI denies: the true knee could sit anywhere in 1.03-2.20, either side of 1.62, so a hard 1.6 reads as a certainty the meta-regression does not carry; and (ii) silently drops the rising left arm — below the (uncertain) breakpoint more protein still helps, so if the true knee sits nearer 2.2 a person told 1.6 is under-dosed and forfeits the objective -> Protein and Resistance Training for Muscle and Strength (which already says hold the number loosely and biases UP to 1.8-2.2 for hypertrophy for exactly this reason). And (iii) it trusts a centre and an interval the estimator itself cannot deliver — a level deeper than (i)-(ii), which still take 1.03-2.20 at face value. A break-point is a non-regular parameter (it indexes which observations fall on which side of the knee — a discrete, non-smooth feature), and precisely there the normal approximation is least safe: this is not the CI is worthless (the CLT earns normality for regular estimators at adequate n), it is that a changepoint’s sampling distribution is routinely skewed, sometimes multimodal, and its standard-error interval mis-covers (its coverage is not the nominal 95%). Two consequences follow. The reported point 1.62 is the mode of the likelihood, which coincides with the mean/median only under symmetry — and 1.62 is essentially the CI midpoint (mean of 1.03 and 2.20 = 1.615), so aim for 1.62 just collapses the interval to a summary warranted only under that symmetry (Decision relevance, below). And the tell sits on the page: 1.62 (1.03, 2.20) is near-symmetric (0.59 below, 0.58 above) — the signature, most plausibly, of an estimate +/- 1.96 x SE normal-theory interval, strongly suggesting symmetry was assumed in rather than found (near-symmetry is only consistent with Wald — a profile-likelihood or bootstrap CI could be near-symmetric too, and 0.59 vs 0.58 is symmetric only to rounding — not proof of it). So the neatness is not reassurance: this apparently-Wald CI is probably mis-shaped and its centre not a trustworthy target. The honest object is the region with its CI — whose tidy symmetry is itself no evidence of tight identification — not the point.
-
The flagship plateau is mechanism-grounded, NOT statistically established — how thin an observed plateau is in this domain. Morton’s 1.62 knee — this page’s worked genuine plateau — rests on the acute-MPS-saturation mechanism plus a biphasic (knee+plateau) fit that was not significant (p=0.079, R2=0.19); and on the baseline-intake axis a linear (monotone) regression explained significantly more variance than the biphasic one -> Protein and Resistance Training for Muscle and Strength. So even the corpus’s best-motivated plateau is a mechanism-supported knee, not a demonstrated one — and on one axis the data significantly prefer a straight line (Route 2), which is why Route 1 leans on the others rather than standing beside them. (A per-total-body-mass target is also off-support for a body composition Morton never sampled — no BMI/body-fat reported, energy-restricted and obese cohorts excluded — the studied-range boundary binding a covariate, not just the dose axis; full treatment on the protein page.)
- A second, independent lab reaches the same monotone-not-knee on the protein curve — Route 2, not
Route 1. Refalo 2025’s meta-regression of the energy-deficit protein->FFM relationship found a
linear model beat quadratic and cubic (97% probability monotone-positive; no knee, no plateau
over the analyzed 0.8-3.2 g/kgBM range) -> Protein Intake During Energy Restriction. Different lab
(Deakin/AUT vs Phillips’ McMaster), different population (deficit vs energy balance), no shared
author, so this is a genuine
[E-independent]corroboration that the protein curve is monotone over the studied range rather than plateaued — strengthening the demotion of Route 1: even the domain’s flagship candidate plateau reads as an open-topped monotone line once a second lab looks. Its ES-zero crossings (1.9 g/kgBM, 2.5 g/kgFFM) are exactly Route-2 sampling-edge artifacts to read as a floor + direction, never a peak. (inferred from Refalo et al., 2025) - A THIRD analysis puts the inflection somewhere ELSE — the knee is population- and estimator-dependent, so there is no single quantity to be off about. Tagawa 2020 (Miyachi lab, no Morton/Refalo author) pooled 138 RT+non-RT trials and put the general-population diminishing-returns inflection at 1.3 g/kg BW/d via a multivariate-adjusted spline — yet that spline stayed positively correlated with LBM across the whole 0.5-3.5 g/kg range (monotone, no plateau), and its RT subgroup kept rising past 1.3 (verbatim quotes on the protein page). (Not clean independence of Morton — Tagawa cites Morton and shares RT trials; the independence is real only for its without-RT leg — so read this as population-dependence, not a third vote.) Across Morton (RT ~1.6/absent), Tagawa (mixed ~1.3), and Refalo (deficit, monotone-no-knee) the inflection moves with population and estimator and no analysis locates a true plateau. That is the underivability: not a numeric disagreement about one quantity, but the absence of a single quantity to derive -> Protein and Resistance Training for Muscle and Strength. (inferred from Tagawa et al., 2020)
- A second, independent lab reaches the same monotone-not-knee on the protein curve — Route 2, not
Route 1. Refalo 2025’s meta-regression of the energy-deficit protein->FFM relationship found a
linear model beat quadratic and cubic (97% probability monotone-positive; no knee, no plateau
over the analyzed 0.8-3.2 g/kgBM range) -> Protein Intake During Energy Restriction. Different lab
(Deakin/AUT vs Phillips’ McMaster), different population (deficit vs energy balance), no shared
author, so this is a genuine
-
Guard — this route is DEMOTED, and must not rest on plateaus everywhere. The knees-and-plateaus prior was falsified on the corpus (fibre/sodium/sugar are monotone over the studied range; objective activity shows no plateau), so flatness is not the general shape — route 1 fires only where a flat region is actually observed. Worse, an observed flat top is often downstream of route 3, not an independent biological fact: measurement error’s flat-slope compression erases a knee but never manufactures one -> Measurement Error in Dietary Assessment, so a measured plateau is weak evidence of a true one. Route 1 is real where a curve is genuinely flat, but it carries the least weight and leans on the other two.
- A worked FOOD instance of both halves of the guard — nuts -> mortality
[2026-08-13]. Aune 2016’s nut dose-response is genuinely observed to flatten at ~15-20 g/day for CHD/CVD/all-cause (a real Route-1 plateau, unlike fibre/sodium/sugar) -> Nut Consumption and Mortality — a second observed plateau to set beside ESC fruit/veg, so the falsification’s demotion (not its reversal) is what nuts support. But it is FFQ-measured, so the flat top is exactly the Route-3-downstream case: the measured plateau is weak evidence of a true one. And it carries the boundary-as-target launder in the wild — Aune names 20 g/day «the optimal intake» because returns flatten there, i.e. a knee-region relabelled a point-optimum. Read it as a floor-for-most-of-the-benefit with its studied range, never a target. - A SECOND worked food instance — yogurt -> T2D, the same Route-3-downstream plateau
[2026-09-05]. Gijsbers 2016’s dairy-T2D dose-response MA observes a nonlinear inverse for yogurt that flattens at ~80 g/day (RR 0.86 vs 0 g/day, then no further decrease above 80 g/d) -> Dairy and Cardiometabolic Health, Fermented Foods and Health — a third observed food plateau beside nuts and ESC fruit/veg, again FFQ-measured, so again the measured plateau is weak evidence of a true one (the flat top is exactly what a reverse-causation ceiling or flat-slope compression would also produce). One instructive contrast with the nuts case: Gijsbers reports the 80 g/d knee without relabelling it an optimal intake — the boundary-as-target launder is absent here, so the honest object (a knee-region with its studied range, read as a floor for most of the benefit) survives on the page as written. (inferred from Gijsbers et al., 2016)
- A worked FOOD instance of both halves of the guard — nuts -> mortality
What survives — BOUNDED, not nihilism
No point-optimum is not no knowledge. Up to three things stay demonstrable, and all are actionable — but WHICH of them a given exposure yields depends on its curve: a bounded nutrient-status curve can show all three, whereas a monotone exposure (Route 2) yields a floor and a direction but no observed ceiling at all. Do not assert the full triad by default:
- The FLOOR — a deficiency/requirement threshold, below which harm is real (the demonstrated lower arm on Deficiency Repletion vs Enhancement: repleting the deficient moves hard outcomes).
- The within-range DIRECTION — more (or less) in the studied range still helps, an evidence-indexed directional claim (every-reduction-pays for free sugars; reduce sodium across the typical range).
- The harm-CEILING — an upper bound past which the exposure harms (the RCT-demonstrated upper arm on Deficiency Repletion vs Enhancement: antioxidant/beta-carotene mortality and cancer).
Only the point-optimum and the boundary-as-target are disclaimed. For the exposures that DO have both bounds, the recommendation apparatus mirrors this exactly — RDA + UL is a floor and a ceiling, never an optimal intake. Where both bounds exist they are the STEEP, well-identified ends of a bounded net curve, and the flat interior between them is where noise and net-effect dominate and identification fails; a monotone exposure has only the floor end identified.
The domain bound — clean measurement ESCAPES the noise floor
This principle is scoped to the self-reported / free-living / whole-diet subdomain — the streetlight’s dark half. Where the exposure is measured cleanly, the range narrows and a real effect (even a near-optimal target) becomes identifiable:
- a recovery biomarker (urinary sodium -> absolute 24-h intake -> the tight sodium/BP slope),
- Mendelian randomization (a genetic instrument free of the reporting error),
- controlled feeding (the exposure is delivered, not recalled).
This is precisely why the fabric up-weights MR, trials and biomarkers — they buy back the identification that self-report loses. The escape is itself gated (a biomarker tracks intake only where the body does not regulate or synthesise the analyte -> Measurement Error in Dietary Assessment), so it is available for some exposures and not others. Naming the bound keeps the principle from hardening into a blanket nutrition can never be precise — LDL->CVD and sodium->BP are the standing counter-examples.
The is/ought line — satisfice is a posture, not a finding
The optimum is not derivable is epistemic. Therefore satisfice — clear the floor, avoid the ceiling, move in the evidenced direction, and stop is a decision posture adopted under that limit: reasonable, but a step from is to ought that the evidence does not itself prescribe. It must not be smuggled back as though the data mandated it (the same error the fabric flags as the descriptive->normative category error -> The Descriptive-Normative Category Error). The wiki states the epistemic limit; the choice to satisfice under it, and how much of the range to occupy, is weighed at Layer 3 with the person’s costs and preferences.
One distinction that must not blur — allocation vs dose
The fabric does not stop optimizing. It optimizes ALLOCATION — Layer 1 ranks levers by effect size x certainty and spends attention on the largest remediable gap -> Layer 1 - Ranking Interventions for a Stratum. It satisfices DOSE — per lever, clear the floor and stay in the range rather than chase a point-optimum this domain cannot locate. Satisfice is a claim about the dose, never about where to spend attention.
Relation to the neighbours (distinct decisions, not restatement)
- The Estimate-to-Action Gap (this cluster’s nucleus) governs Layer-2->3: given a parameter, how do you transform it into a decision (its first step is region, not a point). This page is one level upstream — it explains why the parameter is a range/direction and not a point in the first place, systematically, in this domain, and where that fails (the clean-measurement escape). The gap page consumes what this page explains.
- The U-Shaped Association Artifact is the specific case where the harm-ceiling arm of an observational curve may itself be an artifact — a reason even a bounding arm can be untrustworthy.
- Deficiency Repletion vs Enhancement is the seed instance: the floor (repletion) is demonstrable and the ceiling (enhancement-to-harm) is demonstrable, but the flat replete middle yields nothing to optimize toward — this principle read on one nutrient’s status-dependent curve.
Decision relevance
- Read an apparent optimum or threshold as a study-boundary artifact until shown otherwise. The first hypothesis for any best intake or cutpoint is that it marks the edge of the evidence (or the middle of a noise band), not a curve feature. Ask which route produced it before treating it as a target.
- State the honest object: a floor, a direction, a ceiling — not a number. Where the exposure is cleanly measured, a narrow identified target is legitimate; say which regime you are in.
- Never cite a threshold/target bare — carry its two load-bearing facts. A single number silently drops exactly the two things the routes above say it cannot own: its confidence interval (identification uncertainty — how far the true value could sit from the estimate, driven by sampling and heterogeneity and, in this domain dominantly, the Route-3 measurement noise) and its studied range (extrapolation boundary — Route 2, above which there is no data). A bare figure launders a point-estimate or a study-edge into a target; 1.62 g/kg, CI 1.03-2.20, studied to ~2.2 g/kg carries the decision, a bare 1.6 g/kg does not. And a carried CI is necessary but not self-certifying: a symmetric standard-error interval on a non-regular parameter (a break-point / changepoint) mis-covers (its coverage is not the nominal 95%), so its tidy symmetry is no proof of tight identification — carry the interval, but do not read its neatness as reassurance (Route 1’s 1.62 (1.03, 2.20) is the worked case). Where the exposure is cleanly measured the CI may legitimately be tight — that is a reason to report it narrow, never a licence to omit it.
- Do not collapse a CI to its midpoint. The midpoint of an interval is a privileged summary only under symmetry — exactly the property a non-regular break-point estimator may lack (Route 1 (iii)). So aim for 1.62 is not a neutral default: 1.62 ~ mean(1.03, 2.20), i.e. it silently picks the CI midpoint and thereby inherits (iii)‘s defect. The tell that the midpoint is not automatically right: for a bounded safety quantity — say an arsenic exposure limit with CI (a, b) — nobody would set the limit at the midpoint. Which summary to pick instead is not settled here — that is a loss-function question weighed at Layer 3 -> the is/ought section above; the identification point is only that the midpoint is a choice, not the estimate.
- This is an open loop. Nothing here grades a chosen dose against a realized outcome; the wiki verifies only the would-form (would a well-informed advisor decline to name a point-optimum on this evidence?), never that a person was better off for satisficing.