Every effect estimate is implicitly relative to a comparator — the counterfactual the exposure displaces. A bare “X is harmful” or “X is beneficial” is incomplete until it answers instead of what? Change the comparator and the sign and size of the effect change with it. It lives here as a claim-graph concept so it can be woven and challenged, not only applied at the point of recommendation.
The claim is not that comparisons are hard — it is that the comparator is usually left unstated, and the unstated default is doing the work. When a study “removes” a nutrient, the calories are replaced by something; when a person “stops” a behaviour, the time/energy goes somewhere. The estimate is a contrast, never an absolute, so the omitted arm is a hidden premise.
The worked instances (already source-grounded on their pages)
Each restates a finding held on the linked page; the selection and framing as comparator-instances is this page’s synthesis.
- Saturated fat — the harm is the replacement. The effect of cutting SFA is not one number: it depends on what replaces it (polyunsaturated fat lowers CVD risk; refined carbohydrate does not). The recommendation itself splits by replacement nutrient -> Saturated Fat Intake and Replacement. This is the cleanest instance: the exposure is identical, the comparator sets the answer.
- Free sugars — isocaloric swap is null. Swapping free sugars for other carbohydrate at equal energy does not move weight; the harm is the energy, surfaced only when the comparator is held at equal calories -> Free Sugars Intake. A comparator-blind “sugar is fattening” mistakes the molecule for the energy it carries.
- Dietary patterns — each wins only against its comparator. No pattern is best in the absolute; the evidenced benefits are “better than” a specified alternative (Mediterranean vs a typical Western diet), which is why “which diet is best?” has no comparator-free answer.
- “Is the food category doing any work?” — the sibling failure: a category (an isolate vs the whole food; a label spanning heterogeneous items) can carry an effect that is really the comparator’s or a component’s -> Is the Food Category Doing Any Work.
- Massage therapy — one intervention, three comparators, three answers. The cleanest within-review demonstration: the same intervention scores a large effect vs no treatment, roughly half that vs sham, and a small sub-clinical effect vs an active comparator (and loses at 6 months), because the loose controls leave attention/touch/placebo in the estimate -> Massage Therapy for Pain and Function. The comparator gradient here is the finding.
- White meat — the benefit may be what it displaces, and the swap itself is unmeasured. Higher poultry intake associates with a small all-cause mortality reduction (neutral on CV mortality and, per Ramel 2023, null and WCRF-graded as substantial effects unlikely on CVD mortality and T2D), but a high-poultry diet is also a low-red-meat diet and often a more-plant-protein one, so the signal could be crediting poultry for the removal of red meat or the addition of plant protein rather than poultry itself. The sharper comparator point: all three held SR/MAs measure intake of poultry, not the substitution of red meat with white — a limitation the newest source (Ramel 2023) names explicitly, since foods replace rather than add to a diet — so the decision that matters (swap red for white?) is left to imported substitution studies, not the intake analyses -> Poultry and White Meat Consumption.
The failure mode it names
A comparator-blind claim — “seed oils are toxic,” “red meat is harmful,” “sugar is bad” — asserts a contrast while hiding its second arm. It is not necessarily wrong; it is unfinished, and its truth value can flip with the omitted comparator. The diagnostic move: before accepting or acting on an effect, name the counterfactual and ask whether the estimate survives a realistic one (not an idealized one — judged against what the person would actually do instead).
Where it sits
The placement and cross-links below are this page’s reasoning.
- It is the decision-side twin of Surrogate Outcomes (what the effect is measured on) and The Estimate-to-Action Gap (turning an estimate into an act): together they are the three standing checks between a study result and a recommendation.
- It grounds the Layer-3 frame-as-substitutions rule (CLAUDE.md): a recommendation is judged against the realistic alternative, not an ideal — so the comparator is elicited per person, not assumed.
- Method grounding is top-down from the meta-method corpus (counterfactual/contrast reasoning; quality as fit-to-question) — cited up, not as a domain source.
Synthesis
The reason this earns a page rather than a line: the same move — find the unstated comparator — resolves otherwise-unrelated disputes (SFA, sugars, patterns, seed oils, isolates) into one question, and the resolution differs by instance. That is a configurative synthesis: no common magnitude, a shared structure. The value is diagnostic, not a pooled number.