How we check a page before it goes up

What we compare, what we publish, and what we do not claim.

What we compare

Every check on this site puts two things next to each other and records whether they match. Three pairings carry almost all of the work.

Provider documentation against a sentence in a guide. A version number, a plan limit, a file-size ceiling or a pixel size is read from a page we can name and open, on a date we write down. The guide then quotes the words the figure was taken from, so a reader can see whether it was copied or rounded up. Documentation also ages: a provider can edit a help page the day after we read it, and the guide is wrong from that moment until somebody opens the page again.

A tool's output against the example printed on that tool's page. The table in the guide and the table the page draws in the browser are the same table: one set of column names, one set of state words, one export. A screenshot is described by the page it came from and the date the build was captured, and it is always the output of a tool on this site.

A checklist item against the exported file. The columns a tool prints are the columns in the published file: claim-ledger.csv carries Claim, Source location, Status, Correction and Notes, and action-register.csv carries Deliverable, Owner, Due date, Source timestamp, Status and Notes. Nothing is renamed on the way out to the file, because a rename is how a column quietly stops being checkable.

When a comparison fails, the failure stays visible. The correction field holds what the source actually says, the row keeps both wordings, and the export carries the row out of the page. Rewriting the sentence and leaving the row as it was would remove the only useful part of the record.

What we do not do

This is not a hands-on test report, and this page is not a benchmark of anything. We do not install software, run the same task through two tools and print what happened, and no number on this site is a measurement we took ourselves. No page carries a score, a star count, a percentage or a ranking, and no comparison table picks a winner.

We also do not act on a reader's behalf. Nothing here sends, publishes, uploads, deploys or submits work for anyone: there is no mail form, no account, no publishing button and no integration that pushes a file to a hosting service. The tools stop where a human has to decide, and the checklists stop where a human has to open the file.

What remains is the part that can be checked from outside: a claim set against the document it came from, described closely enough that somebody else can repeat the comparison without us.

Where the numbers come from

Every version, limit, price and pixel size printed on this site carries three things: the page it was read from, the date it was read, and the wording it was taken from. A figure without those three does not go into a guide.

The counts used on this page are the same kind of record, and each one comes from a page you can open:

• The preflight runner works through ten delivery checks, described on its own page as running “ten delivery checks in the browser” (/tools/preflight/, accessed 2026-09-14).
• The claim ledger keeps five column names, which its page states match the published worksheet: “Claim, Source location, Status, Correction, Notes” (/tools/claim-ledger/, accessed 2026-09-14).
• The action register uses “the same six column names as the file at action-register.csv” (/tools/action-register/, accessed 2026-09-14).
• The commitment diff exports six fields: “Position, Original snippet, Revised snippet, Change type, Status and Notes” (/tools/commitment-diff/, accessed 2026-09-14).

When a source page cannot be read, the row says so instead of holding a guess. The OpenAI file-uploads FAQ returned HTTP 403 when the page on checking AI summaries was prepared on 14 September 2026, so that guide carries no figure from it and records the row as Not checked (/guides/check-ai-summaries/, accessed 2026-09-14).

Two habits follow. A figure is never carried over from an older copy of a page, because an old copy is evidence of what used to be true. And a figure that moves quickly — a plan price, a storage quota, an upload ceiling — is dated in the sentence that uses it, not only in a list at the foot of the page.

How a worked example is made

An example is built from three things, in this order, and the order matters.

The input is real. It is either a document anyone can open, named with its URL and the date it was read, or material the operator is allowed to use with names, addresses and account details reduced. Where the input is not public, the page says what kind of document it was and why it is not quoted.

The output is recorded as it came out. What the tool printed is copied without tidying: the rows that came out empty, the rows that needed a second pass, the sentence a reader would have to fix. Nothing is re-run until the example looks better, because a cleaned-up example teaches the wrong thing about how much work the check actually takes.

The correction is written down. A person reads the output against the input and records what changed and why. The correction field holds the source wording, so the fix can be copied rather than guessed, and a reader sees the draft and its replacement in one row.

Two rules keep the example honest. Material invented to demonstrate a method is labelled inside the row it appears in, not in a note at the top of the page: a sentence that was never in any document says so in its own Notes field. And every image is the output of a tool on this site, captioned with the page and the date, never described as another company's product.

Tools on this site

Four tools sit behind the guides, and all four run inside the browser: the preflight runner (/tools/preflight/), the claim ledger (/tools/claim-ledger/), the action register (/tools/action-register/) and the commitment diff (/tools/commitment-diff/). The index at /tools/ lists what each one is for.

They work on text you paste and on nothing else. Nothing is uploaded and there is no file picker and no account: the ledger states plainly that “Nothing is uploaded. There is no file picker and no address book” (/tools/claim-ledger/, accessed 2026-09-14), and the commitment diff records that the two texts are “not written to a cookie or to browser storage” (/tools/commitment-diff/, accessed 2026-09-14). A tool page reaches the network for nothing while you use it. Close the tab and the rows are gone, so the export is the only copy you will have, and taking it is your decision rather than an automatic step.

Every row ends in one of four states, and there are no others:

• Pass — the line in the source was found and the row matches it.
• Fail — the line was found and the row does not match it; the correction field says what the source holds instead.
• Not checked — nobody has compared this row with the source yet, which is where every new row starts.
• Unassigned — nobody has taken the row. It records who owns the row, not a verdict on the claim, and the two are worth keeping apart.

The four words are fixed across the site, and the tools do not grade the work: a table of twelve rows with three Fails says twelve rows and three Fails, and nothing about whether the draft was good enough to send.

Corrections and versions

Two dates appear on a page and they mean different things. Published is the date the page went up in its current shape. Last checked is the date the sources named on the page were read again. Last checked is not a promise that every link still says the same thing: a provider can edit a help page the day after it was read, and the date does not cover that.

When something is wrong, send the page address, the exact sentence or step, what you believe is right, and a source or a reproducible description of what you got instead. The route is the corrections page, and it is the same route for a broken link, a figure read from the wrong page and a step that no longer works. A correction that can be checked against a document is applied and the page's Last checked date moves with it. A request to remove an accurate but unflattering observation is answered, not obeyed.

We do not move a date to make a page look fresher, and we do not edit silently: what changed is described next to the date, so a reader who saw the earlier version can tell what they were reading. The editorial rules behind that are on the editorial policy page.