Working notes behind website_selection — every query with its population and denominator, the report script and its unedited output, the vendor fold and its residue, the Farsight hand map, the quotes that were checked, the external sources that were verified or rejected, and what could not be established. Corpus-level caveats are on corpus and are not restated here.

Contemporaneous. Written during the run that produced the refresh, 2026-08-27, not reconstructed afterwards. Cursor drain sitting (item 184, run 64), not Claude Code.

The page already existed as a catalogue of ranking services with two borrowed survey figures. The item asked to close remaining TODOs and replace those figures with a corpus “Use in Publications”. sampling already covers how to draw, and already publishes the Alexa → Tranco year table. This sitting:

• filled Farsight (the last TODO)
• added the vendor-fold Use in Publications, keeping Scheitle/Xie as dated historical surveys
• fixed the leftover “discontinued in 2023” sentence (the list bullet already said 1 May 2022; the survey paragraph did not)
• added a methodology section and this provenance page
• did not absorb sampling's method/size/versioning tables

Judgement call: do not create programming:farsight. The ranking is not independently downloadable; the API detail is the Tranco default-list-only trap, already on tranco.

The item brief said “population.sourceList over 4,207 sampling papers (custom seed list 377, Google Play 183, Tranco 119, Alexa 117)”. Those counts are exact strings on the old 4,322-paper corpus. On this run the same method over ALL yields custom seed list 516, Tranco 180, Google Play 143, Alexa 51. Google Play is 281 after folding and 0 on WEB. The page publishes that trap rather than those four numbers as a ranking ranking.

Sentinels are silence. Papers, not tuples. A string may hit several vendors.

cd /workspace/artifacts/wiki
node scripts/rank_fold.mjs                         # self-test; prints Farsight pdns strings and unnamed residue; exit 1 on unknown Farsight spelling
node scripts/report_website_selection.mjs          # every figure
node scripts/report_website_selection.mjs --wiki
node scripts/report_website_selection.mjs --list
node scripts/report_website_selection.mjs --quotes 'tranco'
bash scripts/external_checks_website_selection.sh
python3 pages/compare_ranks.py --wiki              # live Tranco/Umbrella/Majestic ranks; dated snapshot
node scripts/check_page_numbers.mjs pages/design_website_selection.txt out/report_website_selection.txt \
  '===== Use in Publications =====' '===== What to report ====='
node scripts/check_page_numbers.mjs pages/design_website_selection.txt out/report_website_selection.txt
node scripts/check_tables.mjs pages/design_website_selection.txt
node scripts/check_wrap.mjs pages/design_website_selection.txt

Windowed and whole-page number checks passed after the report grew a Y block (sample_fold 764) and Z lines for CC BY 4.0 / Tranco's still-stated CC BY-SA 4.0, rank 5,000, and 1.1.1.1 — those are on the catalogue half of the page, outside the corpus section. The page prints Google's CC BY 4.0 as the licence to use when citing CrUX.

Multi-label on purpose: “Tranco and Cloudflare Radar” is both. sample_fold.mjs is exclusive first-match and answers a different question; do not reuse it as a vendor ranking.

Exact-string undercount on WEB (from the report):

FARSIGHT_PDNS (exact strings, classified as a passive-DNS dataset, not a ranking). 12 strings, 13 papers. The report fails if a new /farsight|dnsdb/i spelling is not in this set:

DNSDB
Farsight passive DNS database
DNSDB and Project Sonar Data Repository
Farsight Security passive DNS
DNSDB and 360 PassiveDNS
Farsight passive DNS
Farsight Security DNSDB
Farsight's DNSDB
Farsight Security Information Exchange (SIE) / Farsight passive DNS database
passive DNS dataset similar to Farsight DNSDB, provided by QiAnXin Company
Farsight PDNS
Farsight DNSDB

Farsight-as-ranking papers: 0. The ranking exists (Tranco methodology, 1M PLD, since 2022-05-01, cache-misses) and is a default Tranco input; nobody in this corpus names it as a sourceList.

Unnamed top-N residue after tightening isUnnamedTop to \btop\b (a first draft matched “desktop website”):

    1  Google AdWords top websites
    1  Google's Top 1,000 Most-Visited Websites
    1  remaining top websites
    1  top 100K websites used in Phase I

Four papers. “Google's Top 1,000 Most-Visited Websites” is a ranking with no vendor family; leaving it in residue is correct. The 764 vs 763 gap is this frame-kind-only paper.

Radar regex requires cloudflare radar or radar (domain )?rank so Tracker Radar does not match. Four WEB papers name Cloudflare Radar, all 2025–2026.

Whitespace-normalised against paper.cols.txt.

Rejected as a publishable finding: counting Alexa listVersion strings dated 2023 or later as “impossible draws”. Sampling already tried this; four hand-checks failed because listVersion is not scoped to the tuple. The published figure is the undated share (34 of 66).

All re-fetched 2026-08-27 by external_checks_website_selection.sh (exit 0).

Rejected:

• SEO listicles ranking “best website traffic tools 2026”.
• Wikipedia as a primary source for Alexa retirement (the page still links it as a name disambiguator; the date comes from Alexa Support).
• DomainTools acquisition URL /resources/blog/domaintools-acquires-farsight-security/ → 404. The Mirror, Mirror post itself says they acquired Farsight; that sentence is used, the dead URL is not.

• A public download URL for the Farsight ranking (as opposed to DNSDB). Tranco's methodology describes it; DomainTools does not publish a CSV equivalent of Umbrella's.
• Whether Quantcast Measure still sells a top-sites file behind a login. The public URL is publisher analytics.
• A re-run of Ruth et al. against the five-provider Tranco list. The 2022 comparison predates Radar-in-Tranco.
• Full-text (as opposed to sourceList) mentions of Farsight-as-ranking. Out of scope; the schema question is what people sampled from.

• Date: 2026-08-27. Agent: Cursor (not claude -p / drain-sandbox).
• Item: design:website_selection: close the remaining TODOs (claimed as id 184 so a night drain could not take design:automated_measurements instead).
• Reviewers: three focused GPT 5.6 Luna medium passes in parallel, then one generic pass with no checklist, after a freeze. Findings in §10.
• No credentials were echoed. Wiki writes go through scripts/dw.mjs (JSON-RPC).
• Deleted the one-off probe scripts/_probe_ws.mjs after folding its questions into the report.

Freeze directory: out/freeze_ws/. Reviewers were handed that snapshot. The content page was not edited while the three focused passes ran. Generic ran on the post-focused-review page.

Model for all four: GPT 5.6 Luna medium (not sonnet/fable as the task spec's default split).

No findings. Report rerun, fold self-test, windowed and whole-page check_page_numbers, tables and wrap: all passed.

Six live citekeys each occur once in bibliography. Scheitle 26.6%/7.84%, Ruth 1,790/70%/27.2%, Galloway $10/top 100,000: verbatim in .cols.

After item 184 published, Karel asked to question the page against later LLM-written pages (the model was ip_classification) and decide what of the human-era catalogue to keep. Not a new drain item.

Kept from the human page: the vendor catalogue (CrUX / Tranco / Radar / Umbrella / Majestic / Alexa / Farsight / SecRank / SimilarWeb / Quantcast), the DNS-list caveats, the Le Pochat manipulation figure, the Scheitle and Xie survey figures as dated history, Farsight default-list-only, the Alexa 1 May 2022 vs Tranco 1 August 2023 distinction, Best Practices folded into What to report, the split with sampling.

Dropped or rewritten: numbered 1–6 Advantages/Limitations that restated the table; “the provider's own example file begins 1,com / 2,net” (the 2026-08-27 top-1m.csv does not; that is top-1m-TLD.csv.zip); ec2.internal as a live Umbrella example (absent from that day's top-1m); drain-internal voice about the work item that commissioned the corpus section; treating CrUX-first catalogue order as if it were current practice rather than Ruth's 2022 accuracy ranking.

Added, matching IP classification: a decision table (claim → list); a live rank-disagreement table and published compare_ranks.py (the analogue of classify_ips.py); blockquotes from Ruth, Scheitle, Galloway; a 2026 status table with paper counts; Papers to read first; Open questions; custom-seed 257 vs ranking 764 pointed at sampling rather than re-derived as a vendor ranking; gstatic.com at Tranco rank 3 as the page's “not one of them is the BBC” moment.

compare_ranks.py run 2026-08-27T14:26:07Z against Tranco 46W9X, Umbrella top-1m (1,000,000 hostnames) and Majestic Million. Umbrella TLD file: 10,905 rows, com=1, net=2, org=6.

Focused freeze: out/freeze_ws_rewrite/ (page MD5 b45a6832…). Content page not edited while the three focused passes ran. Generic freeze: out/freeze_ws_generic/ after the currency wording fix. Model for all four: GPT 5.6 Luna medium.

No findings. Report, fold, windowed-equivalent whole-page check_page_numbers, tables, wrap: passed. Embedded 180 / 1.05 / 1,000 present in Z.

Contemporaneous. Written during the sitting, Claude Opus 5, 2026-09-03. No corpus query was run — this item is a vendor date, and extractions.jsonl cannot adjudicate when Amazon switched a service off. The inputs were the wiki's own page sources and four external primary sources. No corpus figure on any page changed.

The item read: “design:website_selection says Alexa was discontinued 1 August 2023 … Fix the date on design:website_selection (and check whether the same date appears elsewhere).” That sentence was already gone. It was created with the page in January 2025 and survived untouched until the sampling sitting replaced it:

So the content page needed no edit, and did not get one. What the item still bought was the second half — check whether the same date appears elsewhere — which nobody had done site-wide, plus an independent re-verification of the date that replaced it. Both are below.

Not carried over from the pages. Each fetched fresh on 2026-09-03 and matched verbatim against the response body, not against a summary of it.

Every Wayback URL cited for these dates across the wiki was fetched and returned HTTP 200 with the quoted text present: the dated capture web/20221126115049/… used by website_selection, the year-wildcard web/2023/… used by sampling, the bare-article web/2022/…4410503838999 used by biases, and the login-page capture used by website_classification. No primary source for an Alexa retirement in 2023 was found, by this sitting or by the sampling sitting searching from the other direction, or by the external-currency reviewer searching Wikipedia's cited sources, contemporary news coverage and AWS's own full-shutdown listing (which contains no occurrence of “Alexa” at all). That is three independent negatives, not a proof — see §13.9.

Why the union of three page lists — and how the first draft still missed two pages. ?do=sitemap is cached and lagged: on 2026-09-03 it listed 148 ids and omitted 10 live pages, including privacy:ads_txt, privacy:browser_protection and statistics:how_many_sites. A search finds only pages containing the search term. core.listPages is complete — but only at depth 0, and its default is depth 1, which returns 88 of 160 pages and just 8 of the 73 provenance: pages. scripts/dw.mjs called it with no arguments, so the first draft of this audit swept 158 pages and reported that as “every page id”.

The generic reviewer found it. The two missed pages were provenance:privacy:browser_extensions and provenance:privacy:privacy_sandbox; neither mentions Alexa, so the PASS survived on outcome — but the section had been about to claim a site-wide sweep while omitting 1.25% of the wiki. listPages() in scripts/dw.mjs now defaults to depth 0 for every caller, which is the real fix: the next script to ask this wiki “what pages exist” will get all of them.

Why footnotes are separate units. The first version of this script analysed whole lines and passed a mutation it should have caught: replacing the body text of website_selection's Alexa bullet with “Discontinued as of August 1, 2023” still exited 0, because the footnote attached to the same line contains the correct “May 1, 2022” and the rule accepted that as vouching for the sentence. On this wiki the claim is in the body and the source is in the footnote, so a footnote's correctness must never excuse a body claim. Splitting the units fixed it; the mutation output below is the proof.

Why a proximity rule rather than a keyword rule. “Alexa” and “2023” co-occur on 102 lines of the 71 pages that mention Alexa at all (grep -i alexa *.txt | grep -c 2023 over the exported sources), nearly all of them legitimate (66 papers published 2023–2026 still name Alexa; IMC/2023/tracking-profiling-…-alexa-echo-… is a corpus slug; Tranco's swap is a 2023 event). Flagging co-occurrence produces a list nobody reads. The defect is specifically a death verb next to a 2023 date, so that is what is matched, with a ±90-character window around the year checked for Tranco/CrUX/Radar/provider/default-list wording that would make it correct.

Both halves of that rule were wrong in the first draft, and reviewers broke them. The window was 60 characters and the death vocabulary was six verbs, so “Alexa was decommissioned in 2023” passed (the verb was not in the list), an ordinary verbose sentence putting “discontinued” ~76 characters from the year passed, and “switched the service off” — a split phrasal verb that website_classification's own footnote uses — was invisible. Now: 29 death-verb alternatives including split phrasal forms, and PROX = 200. All three phrasings are in the mutation suite (M6, M7, M10) so a future edit to those regexes cannot silently undo the fix.

PROX is measured, not asserted, and its residue is printed. On 2026-09-03 the widest verb–year pair the wiki actually contains inside one unit is 188 characters and the narrowest excluded pair is 215, so 200 sits in a real gap rather than at a guessed bound. Pairs beyond it are not dropped: they are reported as beyond-window with their gap, five on this run, and read by hand. Four are the reworded longitudinal footnote and a design:sampling sentence about papers; one is deployment line 1731, where “Nomad moved to BUSL in 2023” shares a line with an unrelated mention of Alexa. That last one is why the window exists at all: with no distance rule the audit reports one false positive and exits non-zero on a clean wiki.

The general lesson outlives this script: a keyword list and a magic constant are the weakest parts of any audit, and the only way to size either is to have someone attack it.

The first run of the audit swept 160 pages and found 18 units, 0 SUSPECT. Then §13 was published — and §13 is, by construction, a page densely covered in sentences that pair “Alexa” with 2023 and a death verb: the work item quoted, the old error quoted to correct it, ten mutation strings including a literal “Alexa was shut down on 1 August 2023”, and a review log full of what reviewers found. Re-run against the published wiki, the audit reported 16 SUSPECT — every one of them its own documentation.

That is not a nuisance to suppress. It is the shape of the problem, and it forced two changes that make the script better rather than merely quieter:

• Fenced blocks are stripped. <code> and <file> contents are not prose claims; the audit's own published listing and its mutation fixtures are the proof. sitemap.mjs strips the same constructs before hunting red links, for the same reason. Blocks are blanked line-for-line so reported line numbers still point at the right source line.
• provenance: hits are listed, not failed. A provenance page's job includes recording the error verbatim, so on those pages the pattern is expected. They go to a provenance-note bucket which is printed in full and does not set the exit code. SUSPECT is reserved for content pages, which is where a wrong date actually misleads a reader — and where the original bug was.

What this costs, stated plainly: the audit no longer polices provenance prose. If a future run writes a wrong Alexa date into a provenance: page, this script will print it in the provenance-note list and exit 0. The mitigation is that the list is short (11 units) and the script's own summary line says “read them, do not trust the label”. A stronger version would keep failing on provenance pages and maintain an explicit allow-list of the sentences that are meant to quote the error — which is a maintenance burden this item did not justify, and is the honest reason it was not built.

alexa_date_audit.mjs

// Site-wide audit: does any page on measuretheweb.org still date Alexa's
// retirement to 2023?
//
//   node scripts/alexa_date_audit.mjs        # exit 0 = clean, 1 = review needed
//
// Background. design:website_selection once said Alexa was "Discontinued as of
// August 1, 2023". The primary source — Alexa Support's own notice — says
// alexa.com was retired 1 May 2022 and the Top Sites / Web Information Service
// APIs on 15 December 2022. 1 August 2023 is a real date but it belongs to
// Tranco: the day CrUX and Cloudflare Radar replaced Alexa in Tranco's default
// list. Two facts, one year and one organisation apart, and ~30 pages of this
// wiki mention one or the other, so the separation has to be checked site-wide
// rather than on the page that got it wrong.
//
// Method. The page list is the union of three sources because none is complete
// on its own: ?do=sitemap is cached and lagged, and a search only finds pages
// containing the term. core.listPages IS complete, but only at depth 0 — its
// default depth of 1 returned 88 of 160 pages, and the first draft of this
// script trusted that default and swept 158 pages instead of 160. The two it
// missed happened not to mention Alexa. Each page is split into *units* -
// the body with ((footnote)) bodies removed, plus each footnote body on its own —
// because a footnote's correct 2022 date must not vouch for a body claim that
// says 2023 (that mistake let a mutation test through on the first version of
// this script). URLs are stripped before year detection, so an archive
// timestamp such as web/2023/ is not read as a claim.
//
// The defect is a *proximity* pattern: a death verb within PROX characters of a
// 2023 token. A hit is excused only if the 2023 token itself sits next to
// Tranco/CrUX/Radar/provider/default-list wording, or the sentence is this
// wiki's own record of the old error. Every hit is printed with its verdict, so
// the part the rules cannot judge stays visible.
 
import fs from 'node:fs';
import path from 'node:path';
import zlib from 'node:zlib';
import { fileURLToPath } from 'node:url';
import { listPages, searchPages } from './dw.mjs';
 
const BASE = 'https://measuretheweb.org';
const UA = 'measuretheweb-wiki-bot (Claude; contact karel.kubicek@vaultjs.com)';
const CACHE = path.join(path.dirname(fileURLToPath(import.meta.url)), '..', 'out', 'alexafix', 'raw');
const SEARCH_TERMS = ['Alexa', 'alexa.com', 'discontinued', 'retired', 'retiring'];
// PROX was 60 in the first draft. A reviewer broke that with an ordinary
// verbose sentence putting the verb and the year ~76 chars apart. It is now
// 200: measured on 2026-09-03, the widest verb-year pair the wiki actually
// contains inside one unit is 188 chars and the narrowest one excluded is 215.
// Pairs beyond PROX are NOT dropped - they are reported as 'beyond-window' so
// the window is auditable instead of load-bearing.
const PROX = 200;  // chars between a death verb and a 2023 token
const EXCUSE = 90; // chars around the 2023 token searched for Tranco context
 
const toId = (s) => s.trim().toLowerCase().replace(/^:/, '');
 
async function allPageIds() {
  const ids = new Set();
  const res = await fetch(`${BASE}/doku.php?do=sitemap`, { headers: { 'User-Agent': UA } });
  const xml = zlib.gunzipSync(Buffer.from(await res.arrayBuffer())).toString('utf8');
  const fromSitemap = [...xml.matchAll(/<loc>(.*?)<\/loc>/g)]
    .map((m) => toId(m[1].replace(`${BASE}/`, '').replace(/\//g, ':')));
  const fromRpc = (await listPages()).map((p) => toId(p.id));
  const fromSearch = [];
  for (const t of SEARCH_TERMS) for (const h of await searchPages(t)) fromSearch.push(toId(h.id));
  [...fromSitemap, ...fromRpc, ...fromSearch].forEach((id) => ids.add(id));
  console.log(`page ids: sitemap ${fromSitemap.length}, core.listPages ${fromRpc.length}, ` +
    `search(${SEARCH_TERMS.join('/')}) ${new Set(fromSearch).size} distinct -> union ${ids.size}`);
  return [...ids].sort();
}
 
async function source(id) {
  fs.mkdirSync(CACHE, { recursive: true });
  const file = path.join(CACHE, `${id.replace(/:/g, '__')}.txt`);
  if (fs.existsSync(file)) return fs.readFileSync(file, 'utf8');
  const r = await fetch(`${BASE}/doku.php?id=${encodeURIComponent(id)}&do=export_raw`,
    { headers: { 'User-Agent': UA } });
  if (!r.ok) throw new Error(`${id}: HTTP ${r.status}`);
  const text = await r.text();
  fs.writeFileSync(file, text);
  return text;
}
 
// Death vocabulary. The first published version had only retire/discontinue/
// dead/shut down/sunset/end-of-service, and a reviewer walked "Alexa was
// decommissioned in 2023" straight past it. Anything that ends a service in
// English prose belongs here: a false positive costs one line of reading, a
// false negative is the entire defect this script exists to find.
const DEATH = /(retir\w*|discontinu\w*|decommission\w*|deprecat\w*|dead\b|died|death|shut ?down|shut off|switched off|turned off|sunset\w*|end of service|end[- ]of[- ]life|\bEOL\b|out of service|cease\w*|terminat\w*|wound down|wind[- ]down|went dark|taken offline|pulled offline|killed off|defunct|closed down|no longer (?:available|published|exist\w*|operat\w*|maintain\w*|updat\w*|running|live|provided|a frame)|(?:switch|shut|turn|clos|wind|wound|kill|phas)\w*(?:\s+\w+){1,3}\s+(?:off|down|out)|(?:pull|took|take|taken)\w*(?:\s+\w+){0,3}\s+(?:offline|plug))/gi;
const Y2023 = /(1 august 2023|august 1st?,? 2023|2023-08-01|20230801|aug\.? 2023|\b2023\b)/gi;
const TRANCO = /tranco|crux|chrome user experience|cloudflare radar|\bradar\b|default (?:daily )?list|provider|integrated into|removed from/i;
// The 2023 token is a year of *publications*, not a vendor date: a range that
// starts in 2023, or a year sitting next to paper/venue wording. This is what
// keeps the wide window from drowning the report in the wiki's many true
// sentences about papers published 2023-2026 that still name Alexa.
const LITERATURE = /2023\s*[-–—]\s*20\d\d|20\d\d\s*[-–—]\s*2023|2023 or later|\bpapers?\b|\bpublish\w*|\bpublication\w*|\bvenue\b|submission cycle|CCS|IMC|NDSS|PETS|USENIX|WWW|TheWebConf|S&P/i;
const HISTORY = /earlier revision|earlier version|previous version|it said|once said|still says|used to say|wrong|incorrect|corrected|correction|not when|belongs to|no source|confus\w+|reject\w+|contradict\w+|this page/i;
// Corpus bookkeeping: a venue/year/slug line from a report-script dump.
const CORPUS_SLUG = /(?:CCS|IMC|NDSS|PETS|USENIX|WWW|TheWebConf|S&P)\s*\/\s*20\d\d|20\d\d\s*\/\s*(?:CCS|IMC|NDSS|PETS|USENIX|WWW)|alexa-echo|alexaechos/i;
 
// Split a line into units: body without footnote bodies, plus each footnote body.
function units(line) {
  const out = [];
  const fns = [...line.matchAll(/\(\(([\s\S]*?)\)\)/g)].map((m) => m[1]);
  out.push({ kind: 'body', text: line.replace(/\(\([\s\S]*?\)\)/g, ' ') });
  fns.forEach((f, i) => out.push({ kind: `footnote${i + 1}`, text: f }));
  return out;
}
 
// Remove link targets and bare URLs: an archive timestamp is not a date claim.
const stripUrls = (s) => s
  .replace(/\[\[([^\]|]*)\|([^\]]*)\]\]/g, ' $2 ')
  .replace(/\[\[([^\]]*)\]\]/g, ' $1 ')
  .replace(/https?:\/\/\S+/g, ' URL ');
 
// Fenced blocks are not prose claims. This script's own <file> listing and its
// mutation fixtures ("Alexa was shut down on 1 August 2023") are published on
// provenance:design:website_selection, and without this the audit reads its own
// test data as 16 wiki defects. sitemap.mjs strips the same constructs before
// looking for links, for the same reason. Blanked line-for-line so reported
// line numbers still point at the right line of the source.
const stripBlocks = (text) => text.replace(
  /<(code|file)\b[^>]*>[\s\S]*?<\/\1>/gi,
  (m) => m.replace(/[^\n]/g, ' '));
 
const rows = [];
const ids = await allPageIds();
for (const id of ids) {
  const lines = stripBlocks(await source(id)).split('\n');
  for (let i = 0; i < lines.length; i++) {
    if (!/alexa/i.test(lines[i])) continue;
    for (const u of units(lines[i])) {
      const t = stripUrls(u.text);
      if (!/alexa/i.test(t)) continue;
      const deaths = [...t.matchAll(DEATH)];
      const years = [...t.matchAll(Y2023)];
      if (!deaths.length || !years.length) continue;
      for (const y of years) {
        const gaps = deaths.map((d) => Math.abs(d.index - y.index));
        const near = deaths.find((d, k) =>
          gaps[k] <= PROX + Math.max(d[0].length, y[0].length));
        if (!near) {
          // Residue, not silence: a verb and a year in the same unit but further
          // apart than PROX. Printed so nobody has to trust the window, and so
          // raising PROX later is an informed decision rather than a guess.
          rows.push({ id, line: i + 1, unit: u.kind, verdict: 'beyond-window',
            snippet: `gap ${Math.min(...gaps)} chars: ` +
              t.slice(Math.max(0, y.index - EXCUSE), y.index + y[0].length + EXCUSE)
                .replace(/\s+/g, ' ').trim() });
          continue;
        }
        const window = t.slice(Math.max(0, y.index - EXCUSE), y.index + y[0].length + EXCUSE);
        let verdict;
        if (CORPUS_SLUG.test(window)) verdict = 'corpus-slug';
        else if (TRANCO.test(window)) verdict = 'ok-tranco-swap';
        else if (HISTORY.test(window)) verdict = 'ok-history';
        else if (LITERATURE.test(window)) verdict = 'ok-literature-year';
        // A provenance: page's job includes recording the error, quoting the work
        // item, and logging what a reviewer found - all of which look exactly
        // like the defect. Those are listed for reading, not failed, and they do
        // not set the exit code. SUSPECT is reserved for content pages, which is
        // where a wrong date actually misleads a reader.
        else verdict = id.startsWith('provenance:') ? 'provenance-note' : 'SUSPECT';
        rows.push({ id, line: i + 1, unit: u.kind, verdict,
          snippet: window.replace(/\s+/g, ' ').trim() });
        break; // one row per unit is enough to send a human to the line
      }
    }
  }
}
 
const order = ['SUSPECT', 'ok-tranco-swap', 'ok-history', 'ok-literature-year',
  'corpus-slug', 'provenance-note', 'beyond-window'];
console.log(`\npages scanned: ${ids.length}; death verb within ${PROX} chars of a 2023 token, ` +
  `on a line mentioning Alexa: ${rows.length} unit(s)\n`);
for (const v of order) {
  const g = rows.filter((r) => r.verdict === v);
  console.log(`== ${v}: ${g.length}`);
  for (const r of g) console.log(`   ${r.id}:${r.line} [${r.unit}]  ${r.snippet.slice(0, 300)}`);
  console.log('');
}
const bad = rows.filter((r) => r.verdict === 'SUSPECT').length;
const notes = rows.filter((r) => r.verdict === 'provenance-note').length;
console.log(bad === 0
  ? `PASS: no CONTENT page puts a death verb next to a 2023 date without Tranco ` +
    `context. ${notes} provenance-note unit(s) listed above are not failed - ` +
    `read them, do not trust the label.`
  : `FAIL: ${bad} content unit(s) date Alexa's end to 2023 with no Tranco attribution.`);
process.exit(bad === 0 ? 0 : 1);

Read the unit count as “as of a revision”, and the SUSPECT count as the claim. This is the real, unedited run against the wiki as it stood immediately before this section was last saved. Re-run it now and the unit count will be higher, because this section is itself part of the input and every save adds sentences that pair Alexa with 2023 (§13.3.1). That is a fixed point nobody can publish their way out of: the output block can never be a run that includes the page carrying it.

So the published claim is the invariant, not the tally: SUSPECT is 0 on content pages in every run of this sitting — before the section existed (18 units), after it was first published (37), and after the revision that added §13.3.1 (39). The committed scripts/alexa_date_audit-output.txt is regenerated after each save and will disagree with the block below in its counts, by design.

page ids: sitemap 148, core.listPages 160, search(Alexa/alexa.com/discontinued/retired/retiring) 83 distinct -> union 160

pages scanned: 160; death verb within 200 chars of a 2023 token, on a line mentioning Alexa: 37 unit(s)

== SUSPECT: 0

== ok-tranco-swap: 14
   design:longitudinal:73 [footnote1]  ntegrated into the default Tranco list, starting from the daily updated list of August 1, 2023"// (fetched 2026-09-03). It is **Tranco's** date, not Amazon's — ''alexa.com'' was retire
   design:website_classification:228 [footnote1]  age, URL — retrieved 2026-08-07. An earlier revision of Design:Website Selection gave 1 August 2023; that date is when Tranco dropped Alexa from the default list, not when Amazon switched t
   design:website_selection:18 [body]  ad now. Alexa.com was retired on **1 May 2022** (APIs 15 December 2022); **66 papers from 2023–2026 still name it**, 34 of them with no version at all. Do not write "the Tranco top 1M"
   design:website_selection:263 [footnote1]  Support, "We retired Alexa.com on May 1, 2022", archived at web.archive.org . The date **1 August 2023** is when Tranco dropped Alexa from the default list and folded in CrUX and Cloudflare Ra
   design:website_selection:338 [body]  Tranco overtook Alexa in 2023, the first full year after Alexa.com died. The lag is compatible with a submission cycle;
   design:website_selection:346 [body]  f crawling papers ends around 2022 (PAM 2024; outside the seven venues). Neither includes 2023–2026, which is the window in which Tranco became the default and Alexa became residue. **
   programming:tranco:64 [body]  l 2020; Farsight joined on 1 May 2022; Alexa was replaced by CrUX and Cloudflare Radar on 1 August 2023, which is also when the default prefix stopped being ''full'' and became one million. Yo
   provenance:design:sampling:280 [body]  site_selection'' visibly wrong** — that page still says Alexa was discontinued "August 1, 2023" and that Tranco combines "Alexa, Cisco Umbrella, and Majestic", both contradicted here w
   provenance:design:website_selection:16 [body]  orrection | design:website_classification Alexa footnote no longer says this page gives 1 August 2023 as Alexa's death. CrUX licence one-liners on that page and programming:tranco now match
   provenance:design:website_selection:131 [body]  | No Alexa retirement on 1 August 2023 | That date is Tranco's provider swap (CrUX+Radar in, Alexa out). AWS shutdown listing do
   provenance:design:website_selection:264 [body]  anything — added Tranco's input dates and the explicit //"Alexa was not 'discontinued in 2023'"// disclaimer. |
   provenance:design:website_selection:276 [body]  | **1 August 2023** is Tranco's provider swap, not Amazon's shutdown | ''%% URL front page, HTTP 200 | //"
   provenance:design:website_selection:593 [body]  r-set table samples 1 January of each year, and the sentence under it read //"and then in 2023–24 Alexa was replaced by CrUX and Cloudflare Radar"// — the interval between two sampled
   statistics:biases:62 [body]  ntegrated into the default Tranco list, starting from the daily updated list of August 1, 2023."//

== ok-history: 2
   provenance:design:sampling:36 [body]  - It said Alexa was "**Discontinued as of August 1, 2023**". The primary source says Alexa.com was retired **1 May 2022** and the Top Sites / Web
   provenance:design:sampling:233 [body]  | Alexa's retirement dates on this page are correct; the wrong "August 1, 2023" is on ''design:website_selection'', and no source supports an August 2023 date for any A

== ok-literature-year: 0

== corpus-slug: 1
   provenance:design:website_selection:299 [body]  es), nearly all of them legitimate (66 papers published 2023–2026 still name Alexa; ''IMC/2023/tracking-profiling-…-alexa-echo-…'' is a corpus slug; Tranco's swap is a 2023 event). Fla

== provenance-note: 11
   provenance:design:longitudinal:1485 [body]  "2023–24" is the gap between two sampled rows, not the date of the change. The footnote then ga
   provenance:design:website_selection:258 [body]  The item read: //"design:website_selection says Alexa was discontinued 1 August 2023 … Fix the date on design:website_selection (and check whether the same date appears elsew
   provenance:design:website_selection:263 [body]  86570786** | **Removed it.** "Alexa was retired 1 May 2022 (APIs 15 Dec 2022), not August 2023 …" |
   provenance:design:website_selection:278 [body]  e used by design:website_classification . **No primary source for an Alexa retirement in 2023 was found**, by this sitting or by the sampling sitting searching from the other direct
   provenance:design:website_selection:287 [body]  | Hits | **18 units** pair a death verb with a 2023 token inside one unit on a line mentioning Alexa |
   provenance:design:website_selection:291 [body]  | Result | Nothing on the wiki dates Alexa's retirement to 2023. **No content page was edited for this defect.** |
   provenance:design:website_selection:297 [body]  body text of design:website_selection 's Alexa bullet with "Discontinued as of August 1, 2023" still exited 0, because the //footnote// attached to the same line contains the correct
   provenance:design:website_selection:301 [body]  s 60 characters and the death vocabulary was six verbs, so //"Alexa was decommissioned in 2023"// passed (the verb was not in the list), an ordinary verbose sentence putting "discontin
   provenance:design:website_selection:627 [body]  * **Whether an Alexa-adjacent Amazon service was in fact retired in 2023.** Not exhaustively excluded — the sampling sitting checked AWS's own full-shutdown listi
   provenance:design:website_selection:639 [body]  | **Mutation the script missed:** ''**Alexa was decommissioned in 2023.**'' → exit 0, PASS. The ''DEATH'' vocabulary had no entry for decommission / cease / go
   provenance:design:website_selection:657 [body]  | **No primary source for a 2023 Alexa retirement exists.** Searched Wikipedia's article and its cited sources (which give

== beyond-window: 9
   design:longitudinal:73 [body]  gap 508 chars: Farsight came in on 1 May 2022, and Alexa was replaced by CrUX and Cloudflare Radar on **1 August 2023**. That last change is the sharpest form of the problem. ''W9ZN9'' and ''25299'' are **co
   design:longitudinal:73 [footnote1]  gap 571 chars: day's list id and its ''configuration'': the other five rows on 2026-08-27, the 31 July / 1 August 2023 pair on 2026-09-03. ''date/20230731'' gives ''W9ZN9'', ''providers: [alexa, umbrella, maj
   design:longitudinal:73 [footnote1]  gap 419 chars: ZN9'', ''providers: [alexa, umbrella, majestic, farsight]'', ''listPrefix: full''; ''date/20230801'' gives ''25299'', ''providers: [crux, farsight, majestic, radar, umbrella]'', ''listPref
   design:sampling:188 [body]  gap 327 chars: a mirror, a cached CSV, or a list inherited from an earlier paper. **66 papers published 2023–2026 in this corpus still draw from Alexa, and 34 of them (51.5%) give no version or date
   provenance:design:website_selection:278 [body]  gap 211 chars: ture ''web/20221126115049/…'' used by design:website_selection , the year-wildcard ''web/2023/…'' used by design:sampling , the bare-article ''web/2022/…4410503838999'' used by stat
   provenance:design:website_selection:299 [body]  gap 409 chars: **Why a proximity rule rather than a keyword rule.** "Alexa" and "2023" co-occur on **45 lines** of the **71 pages** that mention Alexa at all (''%%grep -i alex
   provenance:design:website_selection:299 [body]  gap 297 chars: lines** of the **71 pages** that mention Alexa at all (''%%grep -i alexa *.txt | grep -c 2023%%'' over the exported sources), nearly all of them legitimate (66 papers published 2023–2
   provenance:design:website_selection:299 [body]  gap 209 chars: c 2023%%'' over the exported sources), nearly all of them legitimate (66 papers published 2023–2026 still name Alexa; ''IMC/2023/tracking-profiling-…-alexa-echo-…'' is a corpus slug; T
   provenance:programming:deployment:1731 [body]  gap 312 chars: on and does not claim: Redis relicensed in 2024 and again in 2025, Nomad moved to BUSL in 2023.

PASS: no CONTENT page puts a death verb next to a 2023 date without Tranco context. 11 provenance-note unit(s) listed above are not failed - read them, do not trust the label.

An audit that reports “0 SUSPECT” is worth exactly what its ability to find a real one is worth, and the first draft of this one could not (§13.3). So the script is mutation-tested: ten mutations, injected one at a time into the cached copy of website_selection. Eight must make the audit exit 1 (M1–M4, M6–M8, M10) and two are controls that must not be flagged (M5, M9). The interesting ones: M3 puts the true 2022 date in the same sentence after the wrong 2023 one; M4 hides the wrong date in a footnote; M6, M8 and M10 use verb forms absent from the first draft's vocabulary; M7 is a verbose sentence that outran its 60-character window; M9 is a true sentence about papers published 2023–2026 that must not be flagged.

alexa_date_audit_mutations.sh

#!/bin/sh
# Mutation test for alexa_date_audit.mjs: prove the PASS is not vacuous.
#
# Eight mutations must be caught (exit 1): M1, M2, M3, M4, M6, M7, M8, M10.
# Two are controls that must NOT be flagged (exit 0): M5 and M9.
# Each is applied to the cached copy of design:website_selection, the audit is
# re-run, its exit code is checked, and the cache is restored.
#
# M1-M5 are the original set: M3 puts the true 2022 date in the same sentence
# after the wrong 2023 one, M4 hides the wrong date in a footnote.
# M6 and M7 are reviewer mutations that broke the first draft of the audit --
# M6 uses a death verb the vocabulary did not have ("decommissioned"), M7 puts
# the verb and the year ~76 characters apart in ordinary prose. M8 is a second
# unlisted-verb phrasing ("ceased operations"), M9 a control for a true sentence
# about papers published 2023-2026, and M10 a split phrasal verb ("switched the
# service off") -- the form a reviewer found in design:website_classification's
# own footnote, which the contiguous verb list could not see. They are kept here
# so a future edit to the DEATH vocabulary or to PROX cannot silently undo it.
set -e
cd "$(dirname "$0")/.."
P=out/alexafix/raw/design__website_selection.txt
cp "$P" /tmp/ws.orig
TRUE_LINE='**Discontinued — alexa.com was retired on 1 May 2022 and the Alexa Top Sites and Web Information Service APIs on 15 December 2022.**'
run() {
  name="$1"; want="$2"; repl="$3"
  python3 - "$P" "$TRUE_LINE" "$repl" <<'PY'
import sys
p,old,new=sys.argv[1],sys.argv[2],sys.argv[3]
s=open(p).read()
assert old in s, "anchor sentence not found; update the mutation script"
open(p,'w').write(s.replace(old,new,1))
PY
  set +e; node scripts/alexa_date_audit.mjs >/tmp/mut.out 2>&1; got=$?; set -e
  cp /tmp/ws.orig "$P"
  if [ "$got" = "$want" ]; then echo "  OK   $name: exit $got (wanted $want)"
  else echo "  BAD  $name: exit $got (wanted $want)"; grep -E '^(== SUSPECT|PASS|FAIL)' /tmp/mut.out; fi
}
echo "mutation test of scripts/alexa_date_audit.mjs"
run "M1 bare wrong date in body     " 1 '**Discontinued as of August 1, 2023.**'
run "M2 wrong year, no day          " 1 '**Alexa was discontinued in 2023.**'
run "M3 wrong date + true date after" 1 '**Discontinued as of August 1, 2023**; alexa.com was retired on 1 May 2022.'
run "M4 wrong date in a footnote    " 1 '**Discontinued.**((Alexa was shut down on 1 August 2023.))'
run "M5 correct Tranco swap sentence" 0 '**Discontinued 1 May 2022.** Tranco removed the Alexa provider from its default list on 1 August 2023.'
run "M6 unlisted death verb          " 1 '**Alexa was decommissioned in 2023.**'
run "M7 verb and year far apart      " 1 '**Alexa, the ranking list that this whole page is about, was discontinued by Amazon after a long and well-publicised wind-down, in the month of August 2023.**'
run "M8 ceased/went dark phrasing    " 1 '**Alexa ceased operations on 1 August 2023.**'
run "M9 correct paper-year sentence  " 0 '**Discontinued 1 May 2022.** 66 papers published 2023-2026 still name it.'
run "M10 split phrasal verb          " 1 '**Amazon switched the service off in 2023.**'
echo "unmutated:"; node scripts/alexa_date_audit.mjs >/dev/null 2>&1 && echo "  OK   clean corpus: exit 0"

mutation test of scripts/alexa_date_audit.mjs
  OK   M1 bare wrong date in body     : exit 1 (wanted 1)
  OK   M2 wrong year, no day          : exit 1 (wanted 1)
  OK   M3 wrong date + true date after: exit 1 (wanted 1)
  OK   M4 wrong date in a footnote    : exit 1 (wanted 1)
  OK   M5 correct Tranco swap sentence: exit 0 (wanted 0)
  OK   M6 unlisted death verb          : exit 1 (wanted 1)
  OK   M7 verb and year far apart      : exit 1 (wanted 1)
  OK   M8 ceased/went dark phrasing    : exit 1 (wanted 1)
  OK   M9 correct paper-year sentence  : exit 0 (wanted 0)
  OK   M10 split phrasal verb          : exit 1 (wanted 1)
unmutated:
  OK   clean corpus: exit 0

The sweep found no wrong date but one imprecise one, on a page that tells the reader the Tranco default list is not a fixed frame. longitudinal's provider-set table samples 1 January of each year, and the sentence under it read “and then in 2023–24 Alexa was replaced by CrUX and Cloudflare Radar” — the interval between two sampled rows, presented as the date of the change. Its footnote gave Amazon's API retirement (15 December 2022) as the cause but never gave the swap date, so the page had the two events adjacent and neither of them dated to the day. That is the shape the original error grew out of.

Fixed in three saves, the second and third prompted by reviewers:

Notes: longitudinal §17.

Structure checked after the second save, counted from the wiki source (unambiguous) and confirmed in the rendered DOM:

Two of those numbers were wrong in this section's draft — “34 headings, 33 footnotes” — and two independent reviewers caught them. 34 came from grep -c '<h[2-5]' over the whole HTML, which includes the sidebar and page-tools chrome; 33 came from counting class=“fn” / fn_top / fn_bot occurrences, and DokuWiki emits three per footnote, so 11 read as 33. Same trap as the bibtex plugin's two spans per citation marker. Rule for the next run: a rendered-DOM count is only a check if you know how many elements the renderer emits per source construct — count from the source, and use distinct fn__N ids to confirm it rendered.

• Where the “1 August 2023” figure entered the wiki. The oldest revision of website_selection that carries it is the January 2025 human version (rev 1735841558 created the page; the sentence survives untouched until 1786570786). Whether the original author read it off Tranco's front page and mis-attributed it, or got it from a third-party listicle, is not recoverable from the revision history. The Tranco-swap explanation is the only 1 August 2023 event anyone has found, so it remains the most likely origin, but it is inference.
• Whether an Alexa-adjacent Amazon service was in fact retired in 2023. Not exhaustively excluded — the sampling sitting checked AWS's own full-shutdown listing and found no Alexa entry, and this sitting found no primary source either, but “no source exists” is a claim a search cannot close. What can be said is what the pages now say: the Alexa ranking service and its APIs ended in 2022.
• Closed during review. This section originally said no page gave the last pre-swap Tranco list id, only the 2023-08-01 one. The external-currency reviewer simply queried api/lists/date/20230731 and got it — W9ZN9, still [alexa, umbrella, majestic, farsight]. It is now on longitudinal (§13.7). Recorded here because “what could not be established” was, in this case, only what nobody had tried: the endpoint was already documented on tranco and cost one curl.

Four reviewers, all told the context might not be exhaustive and all handed the same bundle: the edited longitudinal footnote, alexa_date_audit.mjs with its output, the mutation script with its output, and this section in draft.

The hardest of the four, and the only one that found a defect in the audit's own population.

Three things went wrong in the making of this, and the rules they earn are worth more than the date was:

← back to the content page · the longitudinal amendment this prompted · corpus-level provenance