Fifty-three items reached Done and four went to Won't Do, and eight of the fifty-seven closed in the hours after the last entry was filed. Entry 019 went out at a quarter past twelve on the Sunday and its counts stop there; what followed that afternoon — a junior chess club's legal and safeguarding epic closing, three open questions retired, a full archive of the old site taken — belongs here instead. For once the arithmetic reconciled to the item: last week's seventy-five and six, plus those eight, is exactly what its window held.
The work worth leading on took four days and could not have been done by a machine. The junior chess club's accessibility has been running on automated gates all summer. This week it was audited by hand — keyboard navigation, zoom and reflow under a reader's own text-spacing, reduced motion and target size, and a screen-reader pass with NVDA on Windows and VoiceOver on iOS — then remediation, then the statement published, and the epic closed behind them.
What the sweeps found is the argument for doing them. Forty-eight focus indicators measured below three-to-one against their backdrop: the board frames, whose focus ring was painted onto the near-white panel behind the board rather than onto the frame. Measured, 1.9:1. Every automated pass had been green on those pages for weeks, because the contrast between a ring and whatever sits behind it is not a property a page can be read for. It has to be measured where it lands.
Two pages then overflowed at three hundred and twenty pixels once a reader's text-spacing was applied — a hostname on one, a reporting address on the other, each a long unbroken string with nowhere left to break. Both were fixed, and then the fix was questioned: it had been verified in Chromium only, and line breaking is one of the few things browser engines genuinely disagree about. Re-checked across engines — a second thought that is cheap before a handover and expensive after one.
— On the record —A record that is wrong is more dangerous than one that is missing — because the missing one gets checked.
That focus-ring finding is also how the week's sharpest lesson surfaced. The sweeps write dated evidence packs, named for the day they were run. The keyboard sweep reported its forty-eight failures; the defect was fixed the same afternoon; the sweep was re-run to confirm the fix; and the pack was rewritten in place, at the same path, now reporting zero. Both numbers are correct. Only one of them is the evidence the audit was closed against, and it is gone.
Version control did not save it either, because the whole day landed in a single commit — the earlier pack was replaced in the working tree and never committed at all. Note which direction that fails in: a sweep is re-run after a fix, so the pack that survives is always the clean one, and the pack destroyed always the one carrying the finding. The bias runs towards a tidier record than the day actually was, which is the exact opposite of what a dated pack exists to do. All three generators now refuse to overwrite, and the refusal names the file it declined to replace.
A second measurement went the other way and ended in changing nothing. A performance audit scores zero on the club's news posts, flagging the lead image as lazily loaded when it is the element the page is judged on — a documented anti-pattern, and with the budget tight it reads as a free win. Measured against a photograph of the size a real post will carry, satisfying it costs seven hundred and fifty-three milliseconds. The delay before the image starts loading falls exactly as the audit intends; the delay before the page renders rises further than it fell, because a large image fetched at high priority competes with the stylesheet and the fonts. It arrives sooner and paints much later.
So nothing changed about how any page loads. What changed is that the measurement now sits beside the exclusion in the gate and beside the image in the post template — the two places where somebody would otherwise act on the report and make the site slower while the score went up. Against the placeholder image the templates ship, the same change measures within noise: the hazard is invisible in exactly the thing you would test it on.
The club also published its first real news post, a season opener replacing the three placeholders that had been standing in so the index and the previous-and-next chain had somewhere to point. It was the first end-to-end run of the weekly publishing runbook, and it did not survive contact intact: one step carried a confidently worded closing sentence that proved false, another specified a directory that had never existed. A runbook step that has only ever been read is not a control.
Two disciplines held on the way through. A chess position supplied as a screenshot was re-authored as a real diagram rather than shipped as a picture of one, and both it and the accompanying game were validated against a chess library rather than by eye — transcription being the one step nothing downstream can check. And the folder incoming material lands in was verified to sit genuinely outside version control, using the command that proves it rather than the one that merely appears to. Material sent in for publication should never reach a repository at all, because a repository cannot be made to forget.
One more measurement, this time of our own working notes. The instruction file loaded before every session on that project had reached four hundred and forty-two kilobytes and nearly five thousand lines, grown sixty-fourfold in twenty-seven days and never once smaller. Every line was justified when it was written; the trouble is that it had taken on four jobs when only two belong in something read before every task. The rule that came out of it matters more than the size — every removal is a move with its destination recorded, never a deletion, and a reviewer confirms that no standing rule or safeguarding decision left the file. A file half the size that loses one trap is a failure.
A local dog-training club's new site was built out across a single evening: design mock-up through to production pages, self-hosted fonts, redirects and a branded 404, structured data, and contact and enrolment routes wired to a managed sending service behind a bot-protection widget, each with a working no-JavaScript fallback. The privacy page was rewritten as a proper information and data-protection policy, and the terms alongside it.
The part worth drawing out is who owns what. The club has its own hosting account, its own repository and its own mirror of the code, populated automatically from ours. That is deliberate, and it is the same principle as the continuity work recorded here a fortnight ago: the test of a handover is not whether it is documented, but whether the client could carry on without us — and an account in our name fails that test however good the documentation.
For the clinical canine massage practice, the week was content and the last of the pre-launch tidying — home-page sections, testimonials and FAQs authored from material already in hand rather than waiting behind the decisions still outstanding, the privacy notice and accessibility statement populated against the specification, the vet consent form wired into the section that asks for it, and the canonical form of the legal-page URLs settled so that sitemap, navigation and canonical tag agree with one another. That last is trivial before a launch and awkward afterwards. For the chess club, authoritative DNS was confirmed to sit with the incumbent host and the nameserver move planned rather than attempted.
Four items were retired as Won't Do — a set of redirect rules and their test harness, a violation-reporting endpoint, and an unattended recovery path — each closed because a decision taken elsewhere had already answered the question, rather than left open to be misread later as live. In progress across every project stands at nine, up from five, most of it the dog-training club's build and a deliverability test still running as the week closed.
The through-line is what a written record is actually worth. An evidence pack that replaced the finding it existed to preserve. An audit whose advice, followed, would have cost three quarters of a second while the score improved. A runbook step worded with total confidence and never once run. Each was written down in good faith, and each was wrong in a way that only measurement could show. Writing something down is not the same as its being true.