Every Health Claim, Tied to the Marker It Moves

225 papers, 316 findings, 14 health goals. Every finding names the paper behind it, the people it was measured in, and the condition it holds under.

Authors
Michael Wan
Affiliations
Independent Analysis
Published
2026-09-13

Health advice arrives as a flat list. Eat oily fish. Take magnesium. Lift weights. Get morning light. Every item looks the same size on the page, and that flatness hides the three things that decide whether any of it is worth doing.

Which endpoint actually moved? “Good for you” is not a measurement. A trial measures something specific — serum testosterone in nmol/L, hs-CRP in mg/L, sleep onset in minutes, LDL in mmol/L — and a claim is only as good as the endpoint behind it. Some popular supplements move a subjective questionnaire and nothing else. That is not nothing, but it is not what most people think they are buying.

How strong is the evidence? A pooled analysis of thirty randomised trials and one cross-sectional survey both produce a percentage. Printed side by side at the same indent, they look equally true.

Are two claims really one lever? Advice stacks badly. If three things all work by lowering the same biomarker, doing all three does not give you three times the benefit — but a list can’t show you that, so people add up effects that overlap.

So nothing here points at a goal directly. A behavior points at something measurable, and that marker belongs to a goal:

One behaviour, sleep loss, and the six different goals it reaches through six different markers Sleep restriction lowers serum testosterone, raises CRP and IL-6, increases lapses in sustained attention, raises anxiety scores, and is associated with higher all-cause mortality. Its effect on depressive symptoms was measured and came back mixed. Each of those is a different marker under a different goal, which is why advice about sleep cannot be compared with advice about any single one of them. BEHAVIOURMARKERGOAL Sleep lossrestriction or deprivation Serum testosterone hs-CRP and IL-6 Lapses in sustained attention Anxiety score Depressive symptoms All-cause mortality Male vitality Inflammation and skin Cognition and focus Mood and anxiety Longevity − − − − = − measured, and the findings came back mixed
One behaviour, six goals, six different markers — and the sixth was tested and came back mixed, which is a result rather than a gap. This is why the middle layer exists: without it, "sleep more" sits in a list next to "take magnesium" at the same indent, and nothing on the page can tell you that one of them reaches six goals and the other reaches one.

The middle layer is what makes the other two problems visible. Where several behaviors land on the same marker, you can see them converge and know they don’t add. Effect sizes are compared only inside a marker AND only against rows reported the same way. Inside a marker is not enough on its own: LDL arrives in this corpus as both mmol/L and mg/dL, and ranking 18.73 of one against 0.33 of the other put two similar findings 57-fold apart. Convertible units are converted; a standardised mean difference and an odds ratio are not convertible and are never ranked against each other. Where that leaves a row with nothing comparable, it scores the neutral middle rather than a rank it has not earned. Nothing is summed anywhere. Click any row for the study: the result sentence quoted verbatim, the population it was measured in, what the finding is conditional on, and the paper’s own figures.

Health Explorer
Legend
SABCDE usefulness — hover any badge for the arithmetic – not ranked: studied, no effect found Usefulness is derived, not measured: evidence strength, plus where the effect ranks among rows on the same marker (pulled toward the middle where that marker has too few rows to rank fairly), plus how many people it was measured in, minus penalties for open quote flags and author conflicts. The two inputs stay on the row so you can overrule it.
META-RCT meta-analysis of randomised trials RCT single randomised trial COHORT prospective cohort CROSS-SEC cross-sectional / animal / mechanistic n=421 participants in the study; k=24 pooled studies; n=? not recorded only: the finding is conditional quote incomplete a stated figure is missing from the quoted sentence check quote the quoted sentence is not a clean verbatim copy conflict of interest an author is employed by a party with a stake in the result COHORT* evidence tag corrected on independent review
+ good: the study found a welcome effect − bad: it found an unwelcome one ? this marker has no inherently good or bad direction = studied, and no effect was found The sign is the verdict, not the arithmetic. Aerobic exercise lowers blood pressure and that is a green plus, because lower is the good direction here. Which way the number actually moved is on the row next to it, and spelled out when you open the study. Four markers stay purple — a melatonin phase shift or a cortisol slope is not good or bad on its own — and = never means "unknown": those rows were measured and came back null.

What got cut

A row marked studied, no effect means somebody ran the trial and the effect was not there. An entry below means no usable study was found at all. Evidence of absence and absence of evidence are different, and collapsing them is how "nobody has checked" becomes "it's fine".

Notes on method

Rows sort by the usefulness tier, which is a derived sort key rather than a measurement. It adds up evidence strength, where the effect ranks among rows on the same marker, and how many people the finding was measured in, then subtracts penalties for rows with an open quote-integrity flag and studies whose authors have a stake in the result. Hovering a badge shows that arithmetic for that row.

Four things keep it honest. The magnitude term is ranked only within a marker, and within one scale inside it. A hormone percentage and a mortality hazard ratio are not comparable quantities, and neither are 18.73 mg/dL and 0.33 mmol/L until one of them is converted. That ranking is also discounted where the marker has few rows to rank: being the larger of two studies is not a percentile, and it used to be worth the same two points as leading a field of fifteen. The sample-size term is centred on a thousand participants rather than on the middle of its own range, because the typical row here has about nine hundred — anchored anywhere else it would have been measuring “is this enormous” while pretending to measure “is this well powered”. A paper that never states a participant count takes that same midpoint, since not knowing is not the same as being small. And both inputs stay printed on the row, so a reader who weighs conditionality differently can ignore the tier and read the effect and the evidence tag directly. The tier is there to put the strongest rows first, not to tell you a number the studies never produced.

Where two behaviors land on the same marker, the tree shows them converging. That is the whole point of the middle layer: you can see that they do not add, without being told.

Markers are not outcomes. Moving serum testosterone inside the normal range is not the same as feeling better; lowering CRP is not the same as avoiding disease. The middle layer buys mechanistic clarity and pays for it with a surrogate-endpoint gap.

Publication bias sits upstream of all of this. A null result nobody published cannot appear here.

Nothing was pooled. Where trials disagree, the strongest single source is shown rather than a pooled estimate, and the disagreement is recorded in the row’s caveats rather than averaged away.

This describes literature, not what you should do. Every row says what a study found in a specific population. If a row looks wrong, the quoted sentence and the DOI are right there — check it.