— this page is the wiki’s own epistemology, the positive half of method-risk R21. It carries no source attributions; every claim is the fabric’s reasoning, and the numeric example (protein 1.62 g/kg, CI 1.03-2.20) is extracted on Protein and Resistance Training for Muscle and Strength, cited there.
An RCT or systematic review estimates a PARAMETER; a decision for a person is a TRANSFORMATION of it. A trial reports the average causal effect of an idealised exposure, in a reference class, on a measured outcome, with an interval. None of those four qualifiers is the thing a person needs to decide. The gap between the two is not a defect in the study — the study is not wrong; it is simply not a decision — and closing that gap is a determinate set of steps the study does not and cannot perform. This page names the steps. It is where both the value and the error of applying evidence live, because every step is also a lever a motivated reasoner can pull to reach a conclusion they already held.
The two symmetric meta-errors
The gap has two failure modes, opposite in direction and equally wrong:
- Scientism — reify the point estimate as the decision. Read a wide interval as a precise number, a population average as the personal effect, an idealised dose as the executed one. The error of taking the parameter for the answer.
- Science-scepticism — dismiss the study because it doesn’t apply to me. The error of refusing the parameter as an input at all.
The honest path uses the parameter THROUGH the transformation, dismissing neither. The transformation steps below are exactly the joints at which either error enters — and each step is dual-use: the same move that makes a good decision is the cover for a rationalised one.
The transformation steps
Each step converts one property of the parameter into one property of the decision. Each links to where the fabric works it in detail.
-
Region, not a point. A wide confidence interval or a flat curve is a region; reading a single number off it is false precision. The worked case is protein: the muscle-protein plateau is 1.62 g/kg/day with a 95% CI of 1.03-2.20 — a wide interval routinely collapsed to a bare 1.6 g/kg target. A minimum effective dose is a region, not a number (the telos), and where the curve is flat the number does not carry the decision -> Surrogate Outcomes, Protein and Resistance Training for Muscle and Strength.
-
The loss function — bias away from the costly tail. A decision needs the asymmetry of costs, the tail risk, and the reversibility that the point estimate omits. Where one side of the estimate is cheap to over-shoot and the other is expensive, aim away from the expensive side. This rule is direction-agnostic: for a hypertrophy objective, protein overshoot is low-harm for healthy kidneys, so bias UP; for training intensity, overshoot loads the injury tail (often irreversible), so bias toward the margin, DOWN. That the same rule gives opposite directions in two cases is the proof it is bias away from the costly tail, not a disguised always err aggressive or always err conservative -> The U-Shaped Association Artifact (tail asymmetry), Better than What (net effect on the outcome menu).
-
Assigned is not actual — offset the instruction. The trial measured an assigned, idealised exposure; a person executes a drifted one. When the drift is directional and measured, the optimal instruction is the physiological target plus a bias-correction offset, aimed so the drifted execution lands on target. Worked case (illustrative — the drift premise is reported, not a fabric-held finding): trainees are reported to over-estimate reps-in-reserve (leave more reps than they believe, especially the less experienced), so leave 3 RIR as instructed would land at ~5-6 actual and under-stimulate. Guards: the offset is load-limited (over-instructing toward failure is safe on high-rep isolation, but re-loads the injury tail on heavy compounds — keep the explicit margin there); the drift is heterogeneous (some under-reach, some over-reach — the offset is person-specific and opposite for the two); and it must be calibrated to measured drift, never asserted — an un-measured I need to go harder is the rationalisation for what one wanted -> Better than What (assigned-advice vs actual-exposure).
-
Transport — by the mechanism’s support factors, not demographics. The parameter is bound to a reference class; transporting it here means checking that the support factors the mechanism needs are present, not that the demographics match. This dimension has its own two symmetric failures, and they live at different telos layers: over-widening (stretch the class to applies to everyone — a Layer-2 / Estimate danger, the researcher’s error, forced by the streetlight when only the easy population was studied) versus over-narrowing (shrink the class to doesn’t apply to me — a Layer-3 / Recommend danger, the individual’s error, the No-True-Scotsman / R21-science-scepticism move). They are the biases of different actors, not competitors for one slot -> R21, R20 (motivated decomposition is the over-narrowing instance).
-
Practice fills the absence — conditionally. Where the parameter is absent, a well-selected practice can lead the evidence: selective pressure is search-without-theory, and a field’s accumulated lore may encode real signal that formal study lags by years. But this is valid only under three conditions — objective alignment (the field selects on your patient-important outcome, not a surrogate traded against it), dense/fast feedback, and ruthless-versus-confounded selection. It is double-edged: the same argument predicts a field’s health-harming practices are equally selected-for, so applied honestly it warns as much as it licenses. And it can never license a particular practice — practice-ahead-of-proof and practice-stuck-in-superstition look identical from inside; the argument implies the population of practices holds some signal, not which one. The worked contrast: bodybuilding’s ~2 g/kg protein heuristic led the evidence (objective-aligned: it selects on muscle, which is the objective; overshoot cheap), while its train to failure lore does not (loads the injury tail) — trust practice where objective-aligned AND overshoot-is-cheap -> Limits of Evidence, Surrogate Outcomes.
The dual-use guard
— this page IS method-risk R21’s positive half, and R21 is the shadow half of the same machinery; they must be read together.
Every step above is simultaneously the correct decision move and a rationalisation vector. Decompose to the right sub-question, read the region, weight the tail, offset for drift, check objective-alignment — each is how a good analyst works, and each is the cover for the study doesn’t apply to MY situation. The per-step discriminator, applied to every transformation:
-
Is this step EVIDENCED/calibrated, or merely ASSERTED? The finer category has its own outcome evidence (not a hopeful label); the drift is measured (not imagined); the objective genuinely aligns (not a mismatched surrogate); the cost structure is real (not conjured to justify the direction wanted).
-
The uniformity tell — the backstop. A determined rationaliser always claims each step is evidenced, so no single-step check is decisive. The meta-signal is the pattern: if the transformations always land you where you already wanted to be, that uniformity is the signature — regardless of how defensible each step looks in isolation. The roughly-right answer is usually the less-exciting one; arriving repeatedly at the exciting one is the flag.
Decision relevance
- Do not report a parameter as a decision, and do not refuse a parameter because it is not one. State the estimate, then perform the transformation out loud — which region, which loss function, which drift, which support factors — so the reader can audit each step for evidenced-versus-asserted.
- The transformation is the analyst’s job and the analyst’s exposure. It is where the wiki adds value over a bare study summary (the RAG null), and precisely where its own motivated reasoning would enter. Run the uniformity tell on the wiki’s own outputs, not only on a user’s.
- This is an open loop. None of these steps grades a decision against a realized outcome; the wiki can verify the would-form (would a well-informed advisor transform the parameter this way?), never that anyone was better off.