A trial reports a result. The result becomes a headline. The headline carries a figure. And often enough, the figure came from a different study, a different endpoint, or an analysis nobody planned to run. That gap is what this publication works in.
Not a complaint about the press. A description of what happens to a number.
On 19 August 2026, Merck and Moderna announced that a Phase 3 trial of a personalised mRNA cancer therapy had met its endpoints in melanoma. The announcement gave no effect size — no hazard ratio, no interval, no absolute rates.
The figures that appeared in the coverage, a 49% reduction in the hazard of recurrence and 59% in the hazard of distant metastasis, are real, published and correctly cited. They belong to a different trial: 157 patients, open-label, five years of follow-up. The trial that was announced enrolled 1,137 patients under a double blind, and has released no numbers at all.
Nobody lied. The specialist outlets that covered it attributed the figures correctly, and several said plainly that the Phase 3 numbers had not been released. But a reader who meets “49% reduction” three paragraphs under “Phase 3 succeeds” comes away with a sense of how large this effect is, and that sense is drawn from a study seven times smaller that could not have measured it.
That is the recurring shape. Not fabrication. A number doing work it cannot do, in a place where nobody put it deliberately.
One question, worked through until the arithmetic is visible.
Before an issue publishes it goes through five independent adversarial checks. One verifies every figure against its primary source. One tests whether anything cited has been superseded by a later readout. One argues the subject’s case against us, and sweeps every other party the piece characterises for accusations we have not evidenced. One tests whether the inferences hold — whether a hazard ratio has quietly become a count of people, whether a missing number has been read as a missing result. One reads the best coverage elsewhere and asks what this piece actually adds.
Four of the five are measured against a test file with fourteen errors deliberately planted in it. The most recent measurement found thirteen. The one it missed, and every blind spot the test has surfaced, is written down.
We are specific about this because “carefully checked” is what everyone says. Ours has a number attached, it is not 100%, and we would rather tell you that than imply otherwise.
An assessment is a position on the evidence at a date. Evidence moves.
A published issue is a snapshot. Six months later a trial reports its final analysis, a readout is superseded, a meta-analysis lands, a regulator changes a label — and the assessment either follows the evidence or quietly becomes wrong.
So the assessments are living documents, and the changelog is the record of them moving: which question changed, what moved it, which specific figures are different now, and whether the conclusion survived. If you follow a contested question because a decision depends on it, that record is worth more than any single issue, because it is the only place the shape of the evidence over time is visible.
Every published assessment is free to read, permanently, with no signup. No metering, no partial pages, no email wall in front of a conclusion.
The changelog goes to subscribers first. When the evidence on a question we have assessed moves, subscribers hear about it before it is folded into the public page.
Corrections are the exception, and they are never held back. If we published something wrong, the fix goes on the public page immediately and the correction note says what it was responding to. A publication that put its own errors behind a signup would deserve none of the trust it was asking for. The changelog is where the evidence moves; it is not where our mistakes go to wait.
And who it is not for.
It is for people who have a reason to care whether a claim holds: a patient or a family member reading about a treatment, a clinician deciding between options that are described as equivalent, a journalist who wants the primary source rather than the press release, anyone who has noticed that the confident version of a story and the careful version are different stories.
It is not a news service and it is not fast. It publishes when a question is worth the work, which is not weekly. If you want to know what happened, read the coverage — much of it is good. This is for the layer underneath it.
One issue at a time, no more than one a week, usually less.
Every issue carries a one-click unsubscribe link. We send nothing else, we do not share the list, and we do not sell anything to it.
What Holds Up is published by CivicScale, which sells healthcare software and takes no outside investment. That is a real conflict, not a theoretical one, and some of the companies whose evidence we assess operate in markets where our customers have a stake in the answer. The full position is in Who Pays for This, including what money can and cannot buy here.