# Dataset history — every release, and what changed This file exists because we tell journalists that the dataset has a dated version history, and a claim like that must be checkable. Every entry below corresponds to a commit in the repository that publishes this directory; the dates are the dates those files went live. Entries marked **against ourselves** are corrections that made our own numbers weaker or our claims narrower. They are listed first in each day because they are the ones worth checking. --- ## 2026-09-15 * **Against ourselves.** The public README and the schema described the numerator of the 53.8 % figure as "records where the goal was not reached". That is wrong. On those 13,972 pages the goal WAS reached — the engine assigns that verdict in its `goalReached == true` branch — but the focus ring was removed or the fields had no labels. The number is unchanged; the sentence describing it was misleading and is now corrected in both files. * **Against ourselves.** `research-records-2026-09-01.csv.gz` shipped with the `verdict` column EMPTY in all 95,524 rows, and `focus_visible` empty as well. The builder read a field name that does not exist in the log. The table has been rebuilt: `verdict` is now populated for every row, and `focus_visible` is filled only where the traversal actually observed the focus ring — 31.4 % of rows — and left blank elsewhere rather than guessed. * The summary JSON was recomputed **from the published table** rather than from internal state, and its keys were translated from Russian to English. Seventy values were compared against the previous summary: all matched. The per-state breakdown used to under-count — its site totals summed to 11,654 of 11,902 and its page totals to 93,491 of 95,524. It now sums to the whole. * Added `research-records-SCHEMA.md` — a column reference for the CSV, including why `focus_visible` is blank on most rows and why both 11,902 and 11,923 are correct site counts. * Added `template-defect-evidence-2026-09-15.csv` — 200 CSS selectors that fail on the same node across many states, with the number of states and cities for each. This is the evidence for the claim that the defect is in a vendor template rather than in individual cities. * Added `keyboard-traversal-engine.mjs` — the measurement program itself, so the numbers can be reproduced rather than trusted. * Schema: the limit of the `server_silent` verdict is now stated plainly. ## 2026-09-15 * **Against ourselves, and this one changes a headline number.** Our traversal engine counted the page goal as reached when the focused element matched a keyword either in its text or in its href. The skip link that good sites put first - "Skip to main content" - contains the word *main* and usually points at `#main` or `#content`, so on pages whose keywords include *main* or *navigation* it was accepted as the goal at Tab #1. A skip link is a genuine accessibility feature and a keyboard user really does use it. But it does not take you to the contact details, the payment page or the opening hours - which is what the verdict `reachable_by_a_person` claims. Measured on the published municipal log (95,524 records): **6,184 records (6.5 %) had the goal credited to a skip link** - 5,064 of them carried the verdict "reachable by a person", 1,120 "partial focus or labelling barrier". Recomputed with skip links **not** counted as reaching the goal: reachable by a person 11,994 -> 6,930 formally accessible, barrier 21,305 -> 27,489 partial focus/label barrier 13,972 -> 12,852 barrier among measurable 74.6 % -> 85.3 % (35,277 of 47,271 -> 40,341 of 47,271) The correction moves **against the sites we measured**, not in their favour: our published figure was too kind. Both numbers are stated here so that anyone can reproduce either one - the raw log carries the text of the focused element in the field `где_оборвалось`, so no re-traversal is needed to verify this. The same check on the ongoing (unpublished) survey of US commercial and service organisations shows a much larger share - 30.5 % of goals credited to a skip link - because that engine also matched on the href. That survey is not published, and will not be until the definition of "goal reached" is settled. ## 2026-09-12 * Attribution and dataset licences stated per file. ## 2026-09-10 * **Against ourselves.** The OSM-derived organisation registry was republished under ODbL 1.0 after we found it had been offered under CC BY 4.0, which is incompatible with its source. It had briefly been withdrawn; withdrawing it was the wrong fix, and relicensing it was the right one. ## 2026-09-09 * Two further datasets published: the US organisation registry and the government-site axe audit. ## 2026-09-07 * Four re-verification slices published, 99,811 records in total. ## 2026-09-03 * The evidence base was re-stamped into Bitcoin: 110,094 files. ## 2026-09-01 * The raw traversal log and its schema were published. * The evidence base (screenshot hashes) was stamped into Bitcoin and the anchor confirmed in block 965040. * The mechanism behind the headline figure was identified: in 81.8 % of the disagreement cases the visible focus outline had been removed. * **Against ourselves.** Three claims were softened after review, and the verdict taxonomy was repaired. * **Against ourselves.** The dead-domain count was corrected from 1,650 to 1,498: 152 domains we had listed as dead answered on re-check. This made our own headline worse, not better. * A plain answer was added to the question "what if a screenshot is retaken" — the honest limit of screenshot evidence. ## 2026-08-31 * The dead-domain README was updated; the previous slice was named rather than silently replaced. ## 2026-08-30 * The first raw file was published alongside the research page: 1,208 dead domains. --- ## What is still being corrected, right now A re-traversal of 1,266 pages that carry a verdict but no screenshot is in progress. On the first 186 of them, 25 verdicts moved in the sites' favour (23 went from "barrier" to "reachable by a person") and 13 moved the other way. **Which dataset this affects, stated precisely.** That re-traversal belongs to our ONGOING survey of US commercial and service organisations — pharmacy chains, dental practices and the like — not to the municipal snapshot published here. We checked: none of its 472 corrections match a row of `research-records-2026-09-01.csv.gz`, because the two studies cover different organisations entirely. The municipal figures in this directory are therefore unchanged by it, and we are not promising a recomputation that would not touch them. What the re-traversal does tell you is how this kind of measurement behaves under repetition: roughly two corrections in the measured site's favour for every one against. That ratio is the reason the reproducibility limits stated in the schema are worth reading before quoting any single verdict. We would rather publish that we were slightly too harsh than leave a number standing because it reads better. Licence: CC BY 4.0 · contact@codeofdigitaleternity.com