METHOD

How a second measurement is made

Every paper on this site follows the same ten steps, in this order. They exist so that a reader can judge the process and not only the result, and so that we can’t skip one when the finding is exciting.

The shape

The unit is a second measurement: one sentence an organisation published about itself, recomputed by a path it doesn’t control, with the data, the scripts and a stranger’s reproduction beside it.

The instrument is what turns a paper into something that keeps happening: a recheck that runs on a schedule, a clock that counts silence, a table that watches the claims that passed by a hair, and a stated ending for each paper, so that “fixed”, “still broken on day 45” or “crossed the line” is an event the page writes by itself.

The record is what makes the instrument trustworthy: every version of every page in a witnessed log, the mistakes listed in the open, a review gate before anything is published, this page, and a backlog of claims not yet taken. The product is not any one paper. It is that the record can be checked.

The relationships are the part that compounds. A round trip earns a seal and a row. A reviewer sees the draft first. A dependent is told when a clock runs out, and not before.

The pipeline runs backlog, kill condition, paper, recheck, ending. The backlog is public; anyone may measure a claim before we do.

Where this goes, if it keeps that shape, is a standing registry of public promises about logs, datasets and supply chains, each with its measured truth, its margin and its ending, maintained by rechecks rather than by hand, and read by the people who have to act on those promises.

1. Quote the claim

A measurement starts from the organisation’s own words: a policy, a whitepaper, a documentation page, a memo. The sentence is copied verbatim, the page is saved, and the saved copy is hashed and published as an exhibit. If the words aren’t theirs, there is nothing to measure.

2. Name the kill condition

Before any data is pulled we write down the single result that would mean there is no story: the claim holds exactly, the data isn’t public, the rule was withdrawn. If that result arrives, the thread stops the same day, and the check goes into the null rows if it was worth running at all. Fifteen of the twenty-eight checks so far ended there.

The same step now asks who has counted this before: a government monitor, an inspector general, a journal, a newsroom. In regulated fields the count usually exists, and the page then opens with their number and says what it adds, which is almost always the join to what the enforcer did about it. On 3 October 2026 a finished draft on medical-device adverse-event reporting was withdrawn before publication when the same measure turned up in the BMJ; the five papers that had prior measurements now carry them as the first row of their tables.

3. Take a second path

The number is recomputed with our own code, written from the public specification rather than the operator’s tooling, against data anyone can fetch. Where two independent paths exist, both are run and compared: a validator console and our own validator; a vendor’s API and the raw repository behind it.

4. Rerun before publishing

A finding has to survive a second, separate run, on a different day or from a different vantage point. The two runs are both published.

5. The review gate

Before a page goes up, an independent pass re-derives every number on it from the exhibits, checks every quotation against the saved source, and verifies the checksums. The first full pass across all fifteen papers, on 3 October 2026, found errors on twelve of them; they are listed under our own mistakes. The gate is written down at REVIEW.md.

6. Stamp it

The SHA-256 of the page as readers receive it is appended to our own transparency log, which seven outside witnesses cosign and which is anchored to Bitcoin. Every revision gets a new leaf and the old ones stay. Each paper also gets a DOI on Zenodo, with the page bytes, a PDF and the scripts.

7. The subject sees it first

The organisation measured gets the finding before or at publication, with numbered questions. Their reply is printed on the page as written and scored question by question. We take no positions and no payments, and show the draft to no one else, except a named reviewer where one exists.

8. The clock, both ways

Thirty days without a reply and the status line turns red and counts. A reply that changes what we measured, confirmed by the recheck, earns a seal the operator can keep, and the paper carries the time they took in green. The round trips that worked are listed beside the ones that didn’t.

9. Keep measuring

Every paper gets a scheduled recheck, daily or weekly, so the page says the same day when a fix lands or a number moves. The claims that held by a hair are watched first; a margin that crosses the line is the next finding.

10. Say what a stranger can reproduce

Every paper ends with the commands to run. On 3 October 2026 we had each one executed from the public repository on a clean machine by someone who had not seen the work; the result is one line on each paper and a table on the track record. Where a stranger cannot reproduce something (a certificate chain only we hold, a twelve-gigabyte corpus), the page says so instead of implying otherwise.

What this is not

Not an audit, not a certification, not a service to the organisations measured and not a service against them. It is a second look at a public number by someone with no stake in it, published with enough that a third look is cheap.