Differences

This shows you the differences between two versions of the page.

--- provenance:programming:crawler:webxray [2026/08/17 07:54] – New page: provenance for programming:crawler:webxray. Every query with its denominator, the three scripts in full with unedited output, the folds and residues, quotes spot-checked, external sources verified and rejected, what could not be established, jud karel.kubicek.claude
+++ provenance:programming:crawler:webxray [2026/09/22 04:23] (current) – Escape the literal {[citekey]} in the 2026-09-22 scope note: ''..'' monospace does not escape the bibtex plugin, so it rendered as a live unresolved citation marker. Authored by Claude karel.kubicek.claude
@@ Line 16: / Line 16: @@
 | Bibliography | ''pages/bib_additions_webxray.bib'', 15 new entries + 1 hand-written arXiv entry. One generated key (''selmo2025_borges'') was **already present** and was deleted from the additions before appending |
 | Models | Opus 5 wrote the page, the scripts and the folds. Three Sonnet sub-agents ran in parallel: corpus reading of the 15 webXray papers, external primary-source verification, and the 30-row ownership adjudication. Their findings were re-checked rather than trusted — see *Sub-agent findings that did not survive* below |
+==== The second sitting: 2026-09-05 ====
+^ ^ ^
+| Date | 2026-09-05 |
+| What it added | the **random-sample accuracy measurement** the first sitting recorded as the page's biggest gap, and the **Wayback dating** it recorded as unknown. Both were carried as items in *What could not be established* below; both bullets are now rewritten in place rather than deleted, so the gap and its closure are both on the record |
+| Corpus | unchanged. **This sitting used no corpus query at all** — the random sample is a measurement of three live published files, not of the literature. Every corpus figure on the content page is the first sitting's and was not re-derived |
+| Data measured | ''webxray/resources/domain_owners/domain_owners.json'' (''thezedwards'' mirror), Tracker Radar ''entity_map.json'' + ''domain_map.json'' + ''domain_summary.json'', Disconnect ''entities.json'', and the Public Suffix List — the same six files as the first sitting, **re-fetched, and two of them had moved**. Unchanged: ''webxray'' ''e53760188e6dc9aa'', ''tr_entity_map'' ''c4c3f97dbea6cb1e'', ''tr_domain_map'' ''a11bc2580f664544'', ''disconnect_entities'' ''93e4f54036de1b39''. **Changed:** ''tr_domain_summary'' ''7a303228d812a41c'' → ''19f7a5a6a839ec87'' (16,310,837 → 16,347,575 bytes) and ''public_suffix_list'' ''155b43d46932e933'' → ''aef8fb81d63232da''. The consequence is a **different frame**, and it is spelled out in [[#C3.1 The frame, and what it excludes|C3.1]] rather than smoothed over. A first draft of this row asserted all six were unchanged; that was wrong, and a reviewer caught it by re-running the sampler against the first sitting's cache |
+| Scripts written | ''owner_sample.py'' (frame and draw), ''owner_probe.sh'' (mechanical evidence per domain), ''owner_merge_rows.py'' (validation gate), ''owner_verify_sources.py'' + ''owner_verify_rendered.mjs'' (citation verification), ''owner_corrections.py'' (10 hand corrections), ''owner_verdict_fixes.py'' (the retired verdict), ''owner_random_sample.py'' (the estimator), ''owner_selfname_census.py'' (the census that needs no sample), ''owner_citation_audit.py'' (where every citation landed), ''wayback_webxray.sh'' (the CDX queries), ''check_owner_sample_figures.py'' (holds every page figure to the estimator's output) with ''check_owner_sample_figures_mutations.sh'' (proves that guard bites), ''patch_webxray_page.py'' and ''patch_webxray_provenance.py'' (the page edits), ''build_webxray_random_sample_page.py'' (the appendix page) |
+| Published in full | all 15 scripts — measurement, verification and guards — every unedited output and the adjudication brief are on [[provenance:programming:crawler:webxray:random_sample|the random-sample appendix page]]. The three page-generation scripts (''patch_webxray_page.py'', ''patch_webxray_provenance.py'', ''build_webxray_random_sample_page.py'') are **not** published: they produce no figure, they only apply exact-string edits to these pages, and each fails loudly rather than appending to a page it no longer recognises |
+| Bibliography | **unchanged**. This sitting added no citekey and no BibTeX entry, so no cache purge was needed |
+| Models | Opus 5 designed the sample, wrote every script, and made every judgement call recorded below. **21 Sonnet sub-agents** adjudicated 8–9 domains each against the published brief; none of them saw the estimator, the hypothesis or each other's batches. Four reviewers (three Sonnet, one Fable) then ran against the pages, the scripts and these notes |
+| Cost of the thing itself | the adjudication is the expensive part: 175 domains, each needing a fetched primary source, and no shortcut that keeps the sourcing bar. It was **not** done by hand — 21 sub-agents did it against the published brief (19 batches of 9, one of 4, and a one-domain re-issue of the row one agent silently dropped), single-rated — which is cheaper than a human and was the study's main methodological weakness. **A random 20% was re-adjudicated by a second model on 2026-09-11 and the agreement is now measured rather than assumed; see [[#C3.8 Inter-rater agreement: a second model on a random fifth (2026-09-11)|C3.8]].** That is the reason this measurement did not exist before, and the reason it is worth publishing rather than repeating |
 ==== Scope decision: create, extend, or broaden a neighbour? ====
-**Created**, and deliberately broadened beyond its title. The narrow reading of "webXray" is a dead tool with 15 corpus mentions and one paper that ran it — not enough for a page a reader would benefit from. The item asked for "its domain-owner database and how that compares to Tracker Radar and Disconnect", and that comparison is the page: **domain-to-company ownership resolution** has no other page on the wiki, 136 corpus papers touch it, and a fresh PhD student arriving from an adjacent topic needs to be told which list to use now, not which tool is dead.
+**Created** on 2026-08-17, and deliberately broadened beyond its title. **Extended** on 2026-09-05, not broadened further: the second sitting added a measurement the page already promised and already said it lacked, and moved no material in or out of scope.  The narrow reading of "webXray" is a dead tool with 15 corpus mentions and one paper that ran it — not enough for a page a reader would benefit from. The item asked for "its domain-owner database and how that compares to Tracker Radar and Disconnect", and that comparison is the page: **domain-to-company ownership resolution** has no other page on the wiki, 136 corpus papers touch it, and a fresh PhD student arriving from an adjacent topic needs to be told which list to use now, not which tool is dead.
 Alternatives considered and rejected:
@@ Line 36: / Line 49: @@
   * webXray //is// currently developed — commercially, at **webxray.ai**, by webXray LLC, with Libert as founder and CEO. The ''%%<wrap todo>%%'' can be closed.
-The parent page also lists webXray's crawler as "PhantomJS //(historically)//", which is right for the 2015 paper and wrong for the last public version. **TODO for a follow-up sitting:** apply these three corrections to ''programming:crawler'' and remove its ''%%<wrap todo>%%''. Not done in this sitting because the item's scope was this page.
+The parent page also lists webXray's crawler as "PhantomJS //(historically)//", which is right for the 2015 paper and wrong for the last public version. **TODO for a follow-up sitting:** apply these three corrections to ''programming:crawler'' and remove its ''%%<wrap todo>%%''. Not done in the first sitting because the item's scope was this page.
+**Partly closed on 2026-09-05.** One sentence was corrected, and only one, because it directly contradicted the page being published and made the more actionable claim: the parent page said webXray's licence "forbids redistribution, so there is no lawful route to the code now that upstream is gone" and called ''thezedwards/webXray'' the most complete surviving copy. It now says ''thezedwards'' is the most widely //mirrored// copy and not the one to take, records the relicensing back to GPLv3 in 2021 and to MIT in 2023, points at ''peterjoles/webXray'', and states that a lawful route does exist. The webXray-crawler row and the "PhantomJS //(historically)//" line are **still wrong and still open**; its ''%%<wrap todo>%%'' is a list of unrelated open questions and was correctly left alone.
 ===== A. Corpus queries =====
@@ Line 52: / Line 67: @@
 | A8 | Per-year and per-venue shape | year/venue of the 15 | 15 | last use of crawler-or-list: **2022** |
 | A9 | Which ownership resources does the corpus name at all? | six full-text sweeps | 5,859 | PSL 101, Disconnect-as-a-list 74, Tracker Radar 32, WhoTracks.me 23, Crunchbase 23, webXray 15 |
-| A10 | Union of the five owner databases | union of A9 rows minus the PSL row | 5,859 / 1,120 crawled | **136** (2.3% / 12.1%) |
+| A10 | Union of the five owner databases | union of A9 rows minus the PSL row | 5,859 / 1,120 crawled | **136** (2.3% of the corpus); **107** of them are among the 1,120 that crawled (**9.6%** of that population) |
 | A11 | Which Tracker Radar artefact did each of the 32 use? | ''TR_ROLE'' hand map | 32 | ownership 11, tracker/category 9, Collector crawler 9, citation 2, compared 1 |
 | A12 | Ownership use of Tracker Radar over time | year of the 11 | — | 2021:1, 2023:2, 2024:1, 2025:4, 2026:3 |
@@ Line 74: / Line 89: @@
 ==== A.16 Evidence quotes spot-checked ====
-The 16 ''evidence.quote'' values behind the webXray tool and classification tuples, checked by 4-word-window coverage against ''paper.cols.txt'': **4 exact, 11 partial (≥60% of windows), 1 below threshold.** Full list in section H of the report output.
+The 16 ''evidence.quote'' values behind the webXray tool and classification tuples, checked by 4-word-window coverage against ''paper.cols.txt'', with a ''pypdf'' rendering of the same ''paper.pdf'' consulted for anything below threshold: **4 exact, 11 partial (≥60% of windows), 0 rescued from the PDF, 1 below threshold in both renderings.** Full list in section H of the report output.
 The one below threshold is {[musa2022_atom]} at 58%: "We used external data sources including WHOIS records, TLS certificates, and WebXray [61] to identify the parent organizations of each identified tracker." Read by hand in the source: **present and verbatim**; the ''.cols'' rendering splices an adjacent column through it. Below-threshold is not "unsupported".
@@ Line 116: / Line 131: @@
 | B2 | How deep is webXray's tree? | walk ''parent_id'' with a cycle guard | 319 of 827 have a parent; depths 1:508 2:211 3:80 4:24 5:2 6:2; **2,175 of 3,215 domains (67.7%)** have leaf ≠ root |
 | B3 | How much of webXray's schema is filled? | per-field non-empty count | ''country'' 826 (99.9%), ''platforms'' 803 (97.1%), ''uses'' 761 (92.0%), site policy URLs 591 (71.5%), GDPR 132 (16.0%), ''trade_groups'' 104 (12.6%), ''crunchbase_id'' 34 (4.1%), ''opt_out_urls'' 13 (1.6%), ''ccpa_urls'' **4 (0.5%)** |
-| B4 | What is the universe of third-party domains? | Tracker Radar ''domain_summary.json'', folded to eTLD+1 with the live PSL | 47,836 rows → **45,525** registrable domains; 16,396 rows were hostname-keyed; 19 rows dropped as not hostnames |
+| B4 | What is the universe of third-party domains? | Tracker Radar ''domain_summary.json'', folded to eTLD+1 with the **ICANN section** of the live PSL | 47,836 rows → **32,369** registrable domains; the ICANN fold merges **15,651** rows, the ICANN+PRIVATE fold only 2,339; 19 rows dropped as not hostnames |
-| B5 | Coverage, unweighted | domains in the universe the list can name | webXray 657 (1.4%), Tracker Radar 5,539 (12.2%), Disconnect 2,277 (5.0%) |
+| B5 | Coverage, unweighted | domains in the universe the list can name, parent labels walked | webXray 669 (2.1%), Tracker Radar 5,581 (17.2%), Disconnect 2,268 (7.0%) |
-| B6 | Coverage, prevalence-weighted | Σ prevalence of covered ÷ Σ prevalence of all | webXray **54.5%**, Tracker Radar **79.3%**, Disconnect **75.8%** |
+| B6 | Coverage, prevalence-weighted | Σ prevalence of covered ÷ Σ prevalence of all | webXray **58.6%**, Tracker Radar **84.3%**, Disconnect **80.5%** |
-| B7 | Head versus tail | coverage of the top 100 / 1,000 / 10,000 by prevalence | 70/26/5, 95/67/27, 91/67/16 |
+| B6b | How much does the fold/lookup rule move it? | all four combinations of ICANN-vs-full fold and walk-vs-exact lookup | weighted coverage spans 54.5–59.7% (webXray), 79.3–84.8% (TR), 75.8–80.6% (Disconnect). **The rule moves the answer by more than the lists differ.** |
-| B8 | Pairwise agreement | owner strings after a legal-suffix fold | disagreement 32.8% / 46.9% / 38.1% (see the page table) |
+| B7 | Head versus tail | coverage of the top 100 / 1,000 / 10,000 by prevalence | 71/26/5, 98/69/29, 94/68/17 |
-| B9 | Does resolving webXray's tree help? | repeat B8 with root owners | disagreement **rises** to 39.3% and 49.5% |
+| B8 | Pairwise agreement | owner strings after a legal-suffix fold | disagreement 32.4% / 46.8% / 38.0% (see the page table) |
-| B10 | What does nobody cover? | universe minus webXray minus Disconnect | 43,052 (94.6%) of domains, 19.3% of prevalence weight; the top row is ''fonts.googleapis.com'' at 0.369, which **no** list names |
+| B9 | Does resolving webXray's tree help? | repeat B8 with root owners | disagreement **rises** to 39.1% and 49.6% |
+| B10 | What does nobody cover? | universe minus webXray minus Disconnect | 29,896 (92.4%) of domains, 14.4% of prevalence weight; the top row is ''tiktokw.us'' at 0.041, which only Tracker Radar names |
 ==== B.11 Folding: what was folded, the rule, and the residue ====
@@ Line 130: / Line 146: @@
 Deliberately **not** folded: ''group'', ''media'', ''technologies'', ''digital'', ''networks'', ''solutions''. Each is part of a real company name often enough ("Almondnet Group", "Zeta Global", "Lotame Solutions") that stripping it would manufacture agreement between lists that name different companies. Also **no synonym merging at all**: "Facebook" is not folded to "Meta", "DoubleClick" is not folded to "Google". That is the whole point — those are the disagreements being measured.
-**Residue, printed in full by the script:** the fold leaves **782 of webXray's 827 owner names untouched (94.6%)**, i.e. it fires on 45. That is the intended behaviour and its consequence is stated on the page: every "agree" figure is a **lower bound** on real agreement and every "disagree" figure an **upper bound** on real disagreement. A synonym-merging fold would move the numbers in a direction the script cannot justify, so it was not written.
+**Residue, printed in full by the script:** the fold leaves **782 of webXray's 827 owner names untouched (94.6%)**, i.e. it changes 45 — but only **16** of those 45 lose a legal-form suffix. The other 29 change because ''norm_name()'' strips punctuation before the suffix regex runs (''AT&T'', ''56.com'', ''JD.com'', ''Dun & Bradstreet'', "Here, There & Everywhere"). The first version of this page attributed all 45 to the suffix fold, overstating its reach by nearly 3×; the script now prints the two counts separately and lists the punctuation-only names. That is the intended behaviour and its consequence is stated on the page: every "agree" figure is a **lower bound** on real agreement and every "disagree" figure an **upper bound** on real disagreement. A synonym-merging fold would move the numbers in a direction the script cannot justify, so it was not written.
 The second normalisation is the **PSL fold** on the universe (B4). Its residue is printed as the 19 dropped rows, listed individually: 18 bracketed IPv6 literals and the literal string ''"null"'' with prevalence 0.024 and a full behaviour profile attached. A row that is not a hostname is not a domain; it is printed rather than silently dropped.
@@ Line 136: / Line 152: @@
 ==== B.12 Two denominator choices that change the answer ====
+  * **The PSL's ICANN/PRIVATE split, and exact-key versus parent-label lookup.** This is the choice that moves the answer most, and the first version of this script got it wrong in both halves at once — see the reviewer log below. ''googleapis.com'' is a PRIVATE rule, so folding with the private section leaves ''fonts.googleapis.com'' (prevalence 0.369) standing as its own registrable domain, and an exact-key lookup then reports it unowned although all three lists name ''googleapis.com''. The script now prints all four combinations; the recommended rule is an ICANN-section fold, and with that fold every key already //is// a registrable domain so the parent walk is a no-op — which is the cleanest possible confirmation that the fold, not the lookup, is the fix.
   * **Do not use Tracker Radar's ''entity_map.json'' or ''domain_map.json'' as the universe.** That is Tracker Radar's answer key and scores it at 100% by construction. ''domain_summary.json'' is a crawl **result** and is a defensible universe — while still Tracker Radar's own view of the web, which the page states wherever a coverage figure appears. There is no neutral census of third-party domains; if one existed this comparison would use it.
   * **A registrable domain's weight is the //maximum// prevalence among its hostnames, never the sum.** One site can request ''fonts.googleapis.com'' and ''ajax.googleapis.com'', so summing double-counts sites and can exceed 1. Using the max makes every weighted figure a conservative lower bound, and the script says so in a comment at the line that does it.
@@ Line 153: / Line 170: @@
 ''scripts/owner_adjudication.py'' holds 30 rows: domain, the owner a primary source confirms for today, the date ownership changed, the source URL, and one verdict per list from ''{current, stale, granularity, error, absent}''.
-**Sourcing bar:** a company newsroom or press release, an SEC filing, or the domain's own legal document (privacy policy, imprint, data-processing agreement). Wikipedia, Crunchbase summaries, PitchBook, Tracxn and "list of ad-tech acquisitions" pages were used **only to locate leads** and never cited. Where only trade press could be found, the row is marked ''UNRESOLVED'' rather than asserted — 2 of 30.
+**Sourcing bar:** a company newsroom or press release, an SEC filing, or the domain's own legal document (privacy policy, imprint, data-processing agreement). Wikipedia, Crunchbase summaries, PitchBook, Tracxn and "list of ad-tech acquisitions" pages were used **only to locate leads** and never cited. Where only trade press could be found the row is marked ''UNRESOLVED'' and is **excluded from every tally** — 2 of 30, leaving 28. That exclusion was not in the first version, and it mattered: 2 of webXray's 3 "current" verdicts were exactly those two rows, so the figure most favourable to webXray rested on the evidence the script itself called insufficient. Corrected tallies: webXray current on **1 of 23** entries, Tracker Radar **8 of 28**, Disconnect **21 of 21**.
+**Corrected again on 2026-09-05, and the second correction is larger than the first.** The generic reviewer checked the ''absent'' column against the list files and found **9 of the 90 (row, list) verdicts contradict them**: Disconnect has an entry for all eight rows it was scored absent on — five of them under ''resources'' rather than ''properties'', which the page's own definition of Disconnect's coverage includes — and webXray has "PulsePoint" for ''contextweb.com''. ''absent'' is a **fact about a file**, not a judgement, and this script took the adjudicator's word for it. The tallies are now webXray **1 of 24**, Tracker Radar **8 of 28**, Disconnect **26 of 28**, and ''scripts/owner_adjudication_absent_audit.py'' re-derives every one with ''owner_dbs.lookup'' (parent-label, because an exact-key lookup manufactures false absences) with the script refusing to print until it passes. The 2026-09-05 pipeline had this check from the start (''owner_merge_rows.py''), which is why the random sample is unaffected. **Every guard on this page passed the wrong table**, because the page matched the script and the script's data was wrong.
 **The selection bias, stated because it is the main limitation:** these are the disagreements with the **highest Tracker Radar prevalence**, not a random sample. They are chosen precisely where the lists differ, so the tallies are **not an error rate** for any list — a random sample would be dominated by domains all three get right. What they do establish is the //shape// of the disagreement: concentrated in acquisitions and renames, pointing overwhelmingly one way, with the most conservatively regenerated list being the least current. The script prints this warning above its own tallies and the page repeats it in a ''%%<WRAP important>%%''.
@@ Line 171: / Line 190: @@
   * ''stackadapt.com'' — Tracker Radar says "Collective Roll", which is StackAdapt's own pre-2014 founding name, not a distinct owner.
   * ''1rx.io'' — webXray's root is "Marimedia", which no primary source corroborates. Weaker evidence than the other two and labelled as such in the script's comments.
+===== C3. The random sample (2026-09-05) =====
+Section C settled the 30 **highest-prevalence disagreements**. That set is drawn
+exactly where the lists differ, so it cannot be an accuracy rate for any list,
+and the first sitting said so in three places. This section is the measurement
+that can be: a stratified random sample of **each list's own coverage**.
+Everything below is reproducible from
+[[provenance:programming:crawler:webxray:random_sample|the appendix page]],
+which carries all 15 scripts, the adjudication brief the sub-agents
+worked from, and all 10 unedited outputs, in full.
+==== C3.1 The frame, and what it excludes ====
+^ ^ ^
+| Frame | the **32,337** registrable domains in Tracker Radar's ''domain_summary.json'' as fetched on **2026-09-05**, folded with the **ICANN section only** of the Public Suffix List and resolved by parent-label lookup. ''owner_sample.py'' imports ''owner_dbs.py'' and uses its universe function, so the frame is built by the same code as the first sitting's |
+| …but it is **not the same snapshot** | the first sitting's frame, on 2026-08-17, was **32,369** domains. Tracker Radar's ''domain_summary.json'' and the Public Suffix List both changed between the two dates (see the run table above), and the frame moved by 32 domains. Per-list coverage moved with it: webXray 669→664, Tracker Radar 5,581→5,566, Disconnect 2,268→2,264, and webXray's prevalence mass 58.6%→57.9%. The coverage **shares** are unchanged to one decimal (2.1% / 17.2% / 7.0%), which is why the two sets of figures can sit on one page without looking inconsistent — and is exactly why the difference is stated here. Verify with ''python3 scripts/owner_sample.py %%--cache%% out/webxray/cache'' against ''%%--cache%% out/owner_cache'' |
+| Which figures use which | every **accuracy** figure and the coverage counts in the random-sample section use the 2026-09-05 frame (32,337). Every figure in the earlier coverage and agreement sections, and all 30 rows of the hand adjudication, use the 2026-08-17 frame (32,369) and are **not** re-derived: they are pinned to the snapshot the adjudication was done against, and re-running them would orphan that work |
+| What that is | **Tracker Radar's view of the third-party surface**, not a neutral census of the web. Every figure inherits it |
+| Why not a neutral frame | there is no neutral frame to be had. A list of "all third-party domains" is itself a crawl result. Restricting to domains a real crawl saw is the deliberate choice: an entry for a domain nobody ever requests is not an error anybody meets |
+| Who it favours | nobody straightforwardly. It is Tracker Radar's crawl, but Tracker Radar is scored on the same rows as the others and comes out **worst** on the encounter-weighted axis |
+| What it excludes | any domain in webXray's or Disconnect's file that Tracker Radar's crawl never saw. Those entries are never scored — right or wrong. So the rates are conditional on the frame as well as on adjudicability |
+Per-list coverage of the **2026-09-05** frame, a **census** and not an estimate
+(''owner_sample.py''):
+^ List ^ Frame domains it names an owner for ^ Share of 32,337 ^ Share of prevalence mass ^
+| webXray | 664 | 2.1% | 57.9% |
+| Tracker Radar | 5,566 | 17.2% | 83.9% |
+| Disconnect | 2,264 | 7.0% | 79.9% |
+The gap between the third and fourth columns is the whole reason the sample is
+stratified: webXray covers **2.1% of domains** but **57.9% of encounters**.
+==== C3.2 Allocation: quartiles, not the deciles the brief asked for ====
+The follow-up item this sitting implements specified "stratified by prevalence
+decile so the head is not swamped". **Quartiles were used instead**, and that is
+a deviation from the brief, recorded rather than quietly made.
+The reason is arithmetic: 60 draws over ten strata is 6 per stratum, and a
+stratum of 6 supports no statement about that stratum at all — it makes the
+head-versus-tail comparison, which is the thing the decile design was for,
+weaker rather than stronger. Four strata of 24/12/12/12 keep the head heavy
+(the top quartile carries 96.5–99.0% of each list's prevalence mass) while
+leaving each stratum large enough that its exclusion rate is readable.
+^ List ^ Q1 (head) ^ Q2 ^ Q3 ^ Q4 (tail) ^
+| webXray | N=166, prev 0.005377–0.578843, mass 96.5%, n=24 | N=166, 0.000390–0.005117, 3.3%, n=12 | N=166, 0.000034–0.000390, 0.2%, n=12 | N=166, 0.000007–0.000034, 0.0%, n=12 |
+| Tracker Radar | N=1,391, 0.000185–0.578843, 99.0%, n=24 | N=1,392, 0.000027–0.000185, 0.7%, n=12 | N=1,391, 0.000007–0.000027, 0.2%, n=12 | N=1,392, 0.000007–0.000007, 0.1%, n=12 |
+| Disconnect | N=566, 0.001767–0.450522, 96.5%, n=24 | N=566, 0.000192–0.001754, 3.1%, n=12 | N=566, 0.000021–0.000185, 0.3%, n=12 | N=566, 0.000007–0.000021, 0.1%, n=12 |
+**One sample per list, not one shared sample.** A single draw from the union
+would leave roughly ten webXray rows, which is an estimate with no usable
+interval. Each list is drawn from its own coverage instead. A domain drawn for
+one list is adjudicated **once** and its verdict recorded for all three, but
+only a list's own draw enters that list's estimate; the other verdicts are
+printed in section D of the estimator's output so the discarded observations
+stay visible. 180 draws, 5 domains drawn for more than one list, **175 distinct
+domains** to adjudicate, seed 20260905, sample digest ''b9acc0cb20cf7ff3''.
+==== C3.3 The adjudication, and the sourcing bar ====
+The 175 domains went out in 21 batches of 8–9 to 21 Sonnet sub-agents, each
+given the brief published in full on the appendix page and **nothing else** — no
+hypothesis, no estimator, no sight of another batch. The sourcing bar is section
+C's, unchanged: a company newsroom or press release, a regulator filing, a
+company register, the domain's own legal document, or the acquiring parent's own
+site naming the brand as theirs. Never Wikipedia, Crunchbase, ZoomInfo,
+PitchBook, Owler, Tracxn, LinkedIn, an acquisitions list or an SEO listicle —
+those may locate a lead and may not be cited.
+One addition to the bar for this sitting: an **OV/EV TLS certificate's ''O =''
+subject field** is accepted as a register-grade source, because it is an
+identity a CA validated against a company register. The row cites
+''%%tls://<domain>:443%%'' and quotes the exact subject line, and
+''owner_verify_sources.py'' re-does the handshake to check it. 17 of the 135
+sourced rows are of this kind, and **all 17 re-verify**. A DV certificate
+carries no organisation and is never accepted; ''owner_probe.sh'' prints the
+issuer next to the subject precisely so the difference is visible, and its raw
+output for all 175 domains — 55 of which gave no TLS handshake at all — is
+published on the appendix page as the last of its unedited outputs.
+**''unresolved'' is a correct answer and it is common.** 40 of the 175 domains
+could not be settled to the bar. That is not a defect of the adjudication; it is
+what the tail of the third-party surface is like — domains that serve nothing,
+have no legal document, no newsroom, and only aggregators to go on. Every
+''unresolved'' row is excluded from every rate, which makes each rate
+**conditional on the entry being adjudicable**, and the exclusion rate rises
+toward the tail, so the tail estimates are the ones to distrust:
+^ List ^ Drawn ^ Unresolved ^ Eligible ^ Q1 ^ Q2 ^ Q3 ^ Q4 ^
+| webXray | 60 | 11 | 49 | 20/24 | 10/12 | 7/12 | 12/12 |
+| Tracker Radar | 60 | 13 | 47 | 17/24 | 9/12 | 11/12 | 10/12 |
+| Disconnect | 60 | 18 | 42 | 20/24 | 7/12 | 7/12 | 8/12 |
+Every excluded row is printed in full, with the reason, in section E of the
+estimator's output — not summarised, not counted and dropped.
+==== C3.4 The verdict retired mid-run ====
+The brief offered six verdicts. One of them, ''self-named'' — "the list's owner
+is just the domain itself or its bare label … **even if** a real company of that
+name exists" — is wrong, and the sample is what showed it. It fired on
+"TrustArc" for ''trustarc.com'', "Klaviyo" for ''klaviyo.com'' and "Bilibili"
+for ''bilibili.com'', which are the companies, alongside "cdnbasket.net" for
+''cdnbasket.net'', which is not. It fired on **26 of 525** verdicts and it fired
+unevenly: 15 of Disconnect's 60 drawn rows against 4 of webXray's. Left in, it
+would have cut Disconnect's accuracy denominator by a quarter for a reason that
+is about naming style and not about ownership.
+Two things replaced it, and both are published:
+  * **Every affected row was re-scored on the ownership question alone** into ''current'' / ''stale'' / ''granularity'' / ''error'' / ''unknown''. ''owner_verdict_fixes.py'' carries all 26 re-scores, each with the entity string, the adjudicated owner and the reason.
+  * **"The entity name is a domain, not a company" is measured where it belongs** — as a census over every domain→owner pair in every file (''owner_selfname_census.py''), which needs no sample at all: **0.8%** of the 3,215 domains webXray names an owner for, **0.7%** of Tracker Radar's 38,368 and **3.2%** of Disconnect's 7,850. (Those are distinct **domains**, not pairs: webXray's file has 3,224 domain→owner pairs, three of them carrying a trailing slash. The census script's column was mislabelled "pairs" until 2026-09-05.) Against the 25% of Disconnect's //sample// the verdict was firing on, that is the size of the error. The blame belongs to the brief and not to the adjudicators: 25% is close to the census's **27.5%** label-only rate for Disconnect, so they were applying the definition as written — which is what a definition saying "even if a real company of that name exists" asks for.
+The census also prints a second column — names identical to the domain's **label**
+(''klaviyo.com'' → "Klaviyo"), 29.1% / 4.5% / 27.5% — which is usually not a
+defect, because the company really is called Klaviyo. It is printed so the first
+column cannot be inflated by conflating the two.
+==== C3.5 Every citation re-fetched, and the eleven fixed by hand ====
+The single failure mode that would invalidate the whole study is a
+plausible-looking URL that does not say what the row claims — the adjudication
+was done by sub-agents, and a paraphrase inside quotation marks is exactly what
+a language model produces when told to quote. So every cited source was
+re-fetched by machine and searched for the quoted sentence.
+^ Pass ^ What it does ^ Result ^
+| ''owner_verify_sources.py'' | re-fetch the URL with ''curl''-grade HTTP, or redo the TLS handshake for a ''tls://'' source; match the quote after normalising whitespace, quote characters and hyphenation | 121 OK of the 136 rows that carried a source at that point; 13 NOTFOUND, 2 FETCHFAIL |
+| ''owner_verify_rendered.mjs'' | re-check the failures in a real browser (Playwright), because footers and Impressum text are frequently client-rendered | rescued **5**: ''fontawesome.com'', ''flashtalking.com'', ''force.com'', ''conviva.com'', ''cartfulsolutions.com'' |
+| ''owner_corrections.py'' | by hand, for the 10 that survived both, plus 1 applied on 2026-09-17 from the 2026-09-11 re-run of the fixed checker ([[#C3.9 The 2026-09-11 re-run, applied (2026-09-17)|C3.9]]) | 5 ''requote'', 5 ''resource'', 1 ''demote'' |
+The **requotes** are the finding worth keeping: in **4** of 175 rows the "quote"
+was a re-ordering or a summary of the page's own words, despite a brief that said
+verbatim. The claim held in all four and the source stayed; the quote was
+replaced with text copied out of the fetched page. A **fifth** requote,
+''compass-fit.jp'', is counted with them in the table above because the remedy is
+identical, but it is a different defect: **one dropped character**, not a
+paraphrase. It was found on 2026-09-11 and applied on 2026-09-17, because the
+checker that would have caught it in the first place was itself blind to
+non-Latin text until then — see [[#C3.9 The 2026-09-11 re-run, applied (2026-09-17)|C3.9]]. That is the argument for
+verifying every citation by machine rather than trusting the adjudicator, and it
+is the same defect a WebFetch summary produces.
+The 5 **resource** corrections replaced a wrong page for a right claim — a
+geo-redirecting landing page, an SEC //form// URL instead of a document, a host
+that serves nothing to non-browser clients. The 1 **demote** is
+''cnevids.com'': Condé Nast's own privacy policy does not name the domain,
+Tracker Radar's "Sabin, Bermant & Gould LLP" is the registrant law firm (a
+textbook WHOIS trap), and calling that an //error// needs a source naming the
+real owner. The row became ''unresolved'' and left every rate.
+**Where every row's citation ended up** (''owner_citation_audit.py''):
+All four columns below are **as of 2026-09-05**, and two of the numbers have
+since moved for reasons that are not about the citations —
+[[#C3.9 The 2026-09-11 re-run, applied (2026-09-17)|C3.9]] has the arithmetic.
+^ State (2026-09-05) ^ Rows ^
+| cited, confirmed in the fetched bytes | 132 — of which **119 are the whole quote**, 8 matched as fragments either side of an ellipsis, and 5 matched only on a 60-character window of a long quote |
+| cited, confirmed only in a rendered browser | 3 |
+| makes no ownership claim (''unresolved'') and needs no citation | 40 |
+| **uncited** | **0** |
+The three browser-only rows are ''fontawesome.com'', ''flashtalking.com'' and
+''force.com''. A bytes-only audit re-run **on 2026-09-05** reported all three as
+failures — ''fontawesome.com'' no longer, since the ''OK-MARKUP'' status of
+-09-11 lets the bytes pass see the markup its quote is ([[#C3.9 The 2026-09-11 re-run, applied (2026-09-17)|C3.9]]) —
+which is why the audit script exists: the answer to "so are three figures
+uncited?" otherwise lives in three separate JSON files.
+**Every citation was re-fetched from a cold cache after the review**, because a
+reviewer found that the fetch cache retained from the run was the **first,
+partial pass** — it held a 526,545-byte body for ''legal.quantcast.com'' where
+the recorded row has 3,054,755, so a reader reproducing per-row statuses from it
+would not have got the published ones. Nothing published depended on the cache,
+but the artefact was misleading, so the check was redone rather than explained
+away: ''owner_verify_sources.py'' was re-run over all 175 rows against an empty
+cache on 2026-09-05, hitting every source live again. The result is
+**identical** — ''OK=119, OK-FRAGMENTS=8, OK-WINDOW=5, NOTFOUND=2, FETCHFAIL=1,
+NOSOURCE=40'', **zero rows changed OK-ness**, and **zero uncited rows** after the
+browser pass. The stale cache is kept under ''out/verify_cache_partial_firstpass''
+and the reproducing one is ''out/verify_cache''; the re-fetch's own output is
+''out/owner_verify_sources_refetch-output.txt''. This is the strongest evidence
+on this page that the citations hold: it is the whole check redone from nothing,
+by a different invocation, on a different day's network.
+**Total changes after the adjudicators returned their rows: 29 of 525 verdicts**
+— 26 from retiring ''self-named'', 2 from re-scoring webXray's ownership-tree
+//root// rather than its entry (''addthiscdn.com'', ''tqlkg.com''; both are in
+''owner_verdict_fixes.py'' with the reasoning), and 1 from the ''cnevids.com''
+demotion. Separately, **11 of 175 rows had their source or quote replaced**, and
+of those 11 left the verdicts untouched.
+==== C3.5a A page-destroying DokuWiki defect, found before publication ====
+Every list item on the drafts of these three pages was written **wrapped** —
+the bullet on one line, its continuation indented four spaces on the next —
+because that is how the source was readable. **DokuWiki does not accept that**,
+and what it does instead is much worse than dropping the indentation.
+Tested on the live wiki on 2026-09-05 by writing to ''playground:playground'',
+rendering through ''core.getPageHTML'', and reverting. Given this source:
+<code>
+  * **A bullet whose continuation lines are indented four spaces.** This is the
+    second line of the same bullet and it is indented by four spaces.
+    A third line, same indent.
+  * A second bullet, one line only.
+</code>
+the renderer produces a one-line list item, then the **second** line as a
+separate paragraph outside the list, and then this:
+<code>
+&lt;pre class="code"&gt;  the style used throughout the new webXray pages.
+  A third line, same indent.
+* A second bullet, one line only.&lt;/pre&gt;
+</code>
+— the third line onwards becomes a **preformatted code block that swallows every
+following bullet as literal text**. One wrapped item destroys the rest of its
+list, and the page source looks entirely reasonable.
+The drafts carried **46** of these on the content page and **82** on this one.
+Published as written, both pages would have rendered large sections as literal
+source. Nothing in the existing tooling caught it: ''check_wrap.mjs'' passed, the
+figure guard passed, every citekey resolved.
+''scripts/check_wrapped_lists.py'' now detects and joins them, and the three page
+builders run it and **fail** if anything survives. Its controls are the thing
+that makes it trustworthy: it flags 46 and 82 in the drafts and **zero** in the
+two pages already live — which were written with single-line bullets and do
+render correctly. It is also exercised against a nested list, an indented line
+inside a published code block, and an indented paragraph after a blank line, and
+flags none of them.
+The one construct it deliberately does **not** flag is the 171 indented lines in
+section H below, which are pre-existing, are the first sitting's plain-text
+reviewer reports, and render as a code block on purpose.
+==== C3.6 The estimator ====
+^ ^ ^
+| Point estimate | stratified with population weights. Domain-level: ''R = Σ_h (N_h/N) · x_h/m_h''. Encounter-weighted: ''R = Σ_h (M_h/M) · Σ(prev·y_h)/Σ(prev_h)'', a ratio estimator within each stratum |
+| Denominator | the **eligible** rows only: a list's own draws, minus ''unresolved'', minus ''absent''. ''absent'' is not an accuracy question and is counted over the whole frame instead |
+| Intervals | stratified percentile bootstrap, 10,000 resamples, seed 20260905. A normal approximation is meaningless here: several strata have 0 or 1 observation in a cell |
+| Zero cells | a bootstrap interval on a cell with no observations is [0,0] and **asserts nothing**. The script says so and prints the rule-of-three bound instead — below 6.1% for webXray's ''error'' rate (49 eligible, 0 seen) and below 7.1% for Disconnect's (42, 0) |
+| Empty strata | dropped with the remaining weights renormalised, and printed when it happens. **It did not happen** in this run: every quartile of every list has at least 7 eligible rows |
+**Every figure the content page takes from this sample**, with its population:
+^ Figure on the content page ^ Population ^ Value, after the 2026-09-11 second pass ^ Before it ^
+**Refreshed 2026-09-11** after the second pass over the residue settled 28 of the 40 rows the 2026-09-05 pass could not adjudicate. The //before// column is kept, because a table of published figures that quietly replaces them is the thing this page exists to prevent. The second pass is documented on [[provenance:design:ownership_resolution]] §L; its code is on [[provenance:programming:crawler:webxray:residue_pass]].
+| webXray names today's owner, domain-level | 56 eligible of webXray's 60 draws, weighted to its 664 covered domains | 69.2% (56.1–81.4) | //was 69.9% (55.5–83.1) on 49// |
+| webXray, encounter-weighted | same rows, weighted by Tracker Radar prevalence | 93.8% (80.9–98.6) | //was 93.9% (80.9–98.8)// |
+| Tracker Radar, domain-level | 56 eligible of 60, weighted to 5,566 | 73.9% (61.6–85.3) | //was 73.7% (60.8–85.8) on 47// |
+| Tracker Radar, encounter-weighted | same | 74.3% (41.1–97.8) | //was 68.6% (29.3–97.0)// |
+| Disconnect, domain-level | 54 eligible of 60, weighted to 2,264 | **89.2% (80.5–96.8)** | //was 94.4% (86.9–100.0) on 42// |
+| Disconnect, encounter-weighted | same | 82.8% (53.8–100.0) | //was 83.1% (53.7–100.0)// |
+| ''current+granularity'', domain-level | as above | webXray 78.0 (65.8–88.8), TR 76.4 (64.6–87.5), DC 96.8 (91.5–100.0) | //was 80.2, 76.5, 98.8// |
+| webXray ''stale'', domain-level | 56 eligible | 22.0% (11.4–33.8) | //was 19.8% (9.2–32.3)// |
+| ''error'' is uncommon, and no longer absent from two lists | 56 / 56 / 54 eligible | webXray **0** (rule-of-three bound 5.4%), Tracker Radar **3** (''elfsight.com'', ''cnevids.com'', ''cedscdn.it''), Disconnect **1** (''km0trk.com'') | //was 0 / 1 / 0; all four of the new ones were unsettleable in the first pass// |
+| the three pairwise Fisher tests | eligible raw counts, ''current'' only, Bonferroni 0.0167 | **none of the three survives**: DC vs wx 0.031, DC vs TR 0.083, TR vs wx 0.83 | //was DC vs wx 0.014, the one that did// |
+| coverage 2.1% / 17.2% / 7.0% | **census over all 32,337 frame domains**, no sampling error | 664 / 5,566 / 2,264 | unchanged — a census, not a sample |
+| entity name identical to the domain | **census over all pairs**: 3,215 / 38,368 / 7,850 | 0.8% / 0.7% / 3.2% | unchanged — a census, not a sample |
+| what could not be settled | 175 drawn domains, 180 draws | **12 domains, 14 draws**; 4, 4 and 6 of each 60 | //was 40 domains, 42 draws; 11, 13 and 18// |
+| concentration of the encounter-weighted estimate | the eligible rows, prevalence-weighted | ''gstatic.com'' 38.8% of webXray's; ''omtrdc.net'' 20.0% of Tracker Radar's; ''id5-sync.com'' 16.6% of Disconnect's | //was 39.4 / 24.9 / 16.7// |
+**The estimator was re-implemented from scratch and checked against itself.**
+Every guard on this wiki compares a page to a tool's output, so a bug **inside**
+the tool passes all of them. To close that for the figures that matter, the
+stratified estimator was written a second time — twenty lines, reading only
+''owner_sample.json'' and ''adj_rows_scored.json'', importing nothing from
+''owner_random_sample.py'' — and run against both the domain-level rates and the
+C2 concentration figures. It is committed as ''scripts/owner_reestimate_independent.py''
+and published in full on the appendix page with its unedited output, because a
+cross-check nobody can run is an assertion rather than a check. All **24**
+figures reproduce exactly: webXray 69.9 / 19.8 / 80.2 / 0.0, Tracker Radar
+.7 / 22.1 / 76.5 / 1.5, Disconnect 94.4 / 1.2 / 98.8 / 0.0 for current /
+stale / current+granularity / error; and ''gstatic.com'' 39.4%, ''omtrdc.net''
+.9%, ''id5-sync.com'' 16.7% with flips to 54.4% / 43.7% / 66.4%. The script
+carries the estimator's figures as a literal table and **exits non-zero** on any
+disagreement, naming which figure — verified by mutating one of them. This does
+**not** cover the bootstrap intervals, which were not re-implemented and are
+therefore checked only by the estimator itself.
+Those figures are held to the estimator by ''check_owner_sample_figures.py'',
+which asserts that each one is **present on the page and equal to** the script's
+output, and fails if a figure has been deleted as well as if it disagrees. It
+compares the page to the tool, so it cannot see a bug inside the tool or a
+correct figure attached to the wrong sentence; its docstring says so, and the
+prose was re-read by hand as well. The guard itself is mutation-tested by
+''check_owner_sample_figures_mutations.sh'', which corrupts one figure at a time
+and requires the guard to reject all ten corruptions — including the case where
+the mutation changes nothing, which would mean the guard was never exercised.
+Both, and the mutation run's unedited output, are on the appendix page.
+The four Tracker Radar rows the content page names as stale renames —
+''zoominfo.com'', ''mountain.com'', ''20min.ch'', ''tvtime.com'' — are all in
+Tracker Radar's **own** 60 draws, not borrowed from another list's sample, and
+so is ''elfsight.com''. That was checked by hand against ''owner_sample.json''
+before the sentence was written, because a verdict recorded on a domain drawn
+for a different list is exactly the observation the design discards.
+==== C3.7 Limits, in the order they matter ====
+  - **Conditional on adjudicability — and this is the limit the 2026-09-11 second pass went after.** It read: //"11–18 of each 60 rows could not be settled. If unadjudicable entries are systematically worse than adjudicable ones … all three lists are flattered, by roughly the same amount. This is the limit that would most change the numbers and it cannot be bounded from inside the sample."// It was right that the sample could not bound it, and wrong that the flattering would be even: harder sourcing settled 28 of the 40, leaving **4, 4 and 6** of each 60, and Disconnect's domain-level rate fell 94.4% → **89.2%** while webXray's and Tracker Radar's did not move. The recovered rows are not significantly worse than the rows already in the rates (//p// = 0.64 / 1.00 / 0.40 on 7, 9 and 12 rows — almost no power), but they were worse for the list this page recommends, and they cost it the one pairwise comparison that survived Bonferroni. The **12 domains still unresolved** are unbounded for the same reason as before. See [[provenance:design:ownership_resolution]] §L.
+  - **Conditional on the frame.** Entries for domains Tracker Radar's crawl never saw are never scored.
+  - **The encounter-weighted column is a ratio estimator dominated by a handful of head domains, and the estimator now measures by how much.** Section C2 of its output reports, per list, the share of the encounter-weighted estimate carried by the single heaviest eligible row, and what the ''current'' rate would become if that one row's verdict were flipped to ''stale'': ''gstatic.com'' carries **38.8%** of webXray's (93.8% would read 55.0%), ''omtrdc.net'' **20.0%** of Tracker Radar's (74.3% would read 54.3%), ''id5-sync.com'' **16.6%** of Disconnect's (82.8% would read 66.2%); the top-three shares are 52.2% / 52.4% / 45.6%. //(All six moved on 2026-09-11: the second pass added eligible rows, which dilutes every head domain's share. Before it: 39.4 / 24.9 / 16.7 and 53.1 / 64.9 / 46.0.)// The flip is a sensitivity check, not a result. Intervals reach 68 percentage points wide (Tracker Radar, ''stale+error'': 2.9–71.2%). Prevalence in the frame spans five orders of magnitude, so this concentration is what encounter weighting //is// and not a defect in the estimator — but a headline resting on one observation should say so beside itself, and not only inside an interval. Read the point estimate as an ordering, not as a number.
+  - **n is 54–56 per list** after the 2026-09-11 second pass over the residue (42–49 before it). The domain-level intervals are ±12 points at best. Any claim that two lists differ needs the intervals or the Fisher test, and webXray versus Tracker Radar domain-level (69.2 vs 73.9) is **not** a difference this sample can establish — //p// = 0.83. The figures in this bullet carry no ''%'' sign, which is why ''sweep_moved_figures.py'' did not flag them when the rates moved; the generic review pass did.
+  - **A verdict is one adjudicator's reading, and a second reading agrees with it about as often as a coin would on the hardest list.** Measured on 2026-09-11 — see [[#C3.8 Inter-rater agreement: a second model on a random fifth (2026-09-11)|C3.8]]. A random 20% (35 of 175 domains, seed 20260911) was re-adjudicated by a **different model** against the identical brief and the identical per-domain evidence blocks. On the rows the brief actually asks about, Cohen kappa is **+0.027** for webXray (n=12, 25.0% raw agreement, 95% CI -0.200 to +0.280), **+0.478** for Tracker Radar (n=30, 70.0%, +0.244 to +0.706) and **+0.460** for Disconnect (n=23, 60.9%, +0.193 to +0.707); pooled over the 65 cells, **+0.400** at 58.5% raw agreement. **63.0% of the disagreement is one rater finding a source where the other did not**, not two readings of the same evidence: on rows both raters settled, the pooled kappa is +0.680 and on ''current''-versus-rest +0.708. The citation check verifies that the evidence exists and contains the quote; it still does not verify that the verdict follows from it. **What would close what remains:** a subsample drawn from webXray's own coverage rather than from the 175, so its cell is not 12 rows.
+  - **Quartiles, not deciles** — see C3.2.
+  - **All three lists are moving targets except webXray's.** These rates are of the six files hashed above, on 2026-09-05, and Tracker Radar and Disconnect regenerate. webXray's does not, so its rate degrades from 2021-03-04 forward and from no other cause.
+==== C3.8 Inter-rater agreement: a second model on a random fifth (2026-09-11) ====
+C3.7 recorded, as the limit that would be cheapest to close, that every verdict
+was **one adjudicator's reading** and that what would close it was to
+"re-adjudicate a random 20% with a different model and report agreement". That
+was done on 2026-09-11 and this section is the result. It is not a flattering
+one, and the part that is not flattering is the part worth reading.
+**Design.** ''scripts/owner_irr_sample.py'' draws a simple random sample of **35
+of the 175** adjudicated domains under a **separate seed, 20260911** — the
+parent draw's seed is 20260905 and reusing it would have made the audit a
+deterministic function of the thing it audits. The subsample is not
+re-stratified: it is estimating agreement between two readings of one brief, not
+anything about the web, so the population that matters is the sample itself.
+Manifest digest ''07208284bbd90fd6''; the 35 domains are listed in the script's
+output on [[provenance:programming:crawler:webxray:random_sample|the appendix
+page]], section AF.
+**The stimulus is identical, not merely equivalent.** The second rater was given
+''out/adj/INSTRUCTIONS.md'' byte-for-byte, and the per-domain blocks — the three
+lists' claims plus the TLS/RDAP/HTTP probe — were **copied verbatim out of the
+first rater's batch files** rather than regenerated. Re-running ''owner_probe.sh''
+today would have handed rater 2 a different stimulus and any disagreement would
+then be partly the web moving. The brief's own "Today is 2026-09-05" line was
+left in place for the same reason.
+**A different model, and which way that cuts.** Rater 1 was 21 Sonnet sub-agents;
+rater 2 was 5 Fable sub-agents, 7 domains each, no sight of rater 1's verdicts,
+of each other, or of the estimator. A different model //family// was chosen over a
+different instance of the same one deliberately: two Sonnet runs would share
+their failure modes, and an agreement figure whose two raters fail together
+overstates the instrument. The cost is that a capability difference and a
+reading difference are not separable, and section 2b below suggests most of what
+was found is the former.
+^ Population ^ What it is ^ webXray ^ Tracker Radar ^ Disconnect ^ Pooled cells ^
+| **NAMED** — //headline// | rows where the list file names an owner: the rows the brief asks a rater to judge | **+0.027** (n=12, agree 25.0%) | **+0.478** (n=30, agree 70.0%) | **+0.460** (n=23, agree 60.9%) | **+0.400** (n=65, agree 58.5%) |
+| 95% CI of the above | stratified-free percentile bootstrap over domains, 10,000 resamples | -0.200 to +0.280 | +0.244 to +0.706 | +0.193 to +0.707 | +0.241 to +0.551 |
+| ALL rows | all 35, including the cells ''owner_merge_rows.py'' **forces** to ''absent'' from the list file | +0.525 | +0.614 | +0.663 | — |
+| ADJUDICABLE | NAMED rows where **both** raters reached an ownership verdict | +0.250 (n=6) | +0.718 (n=24) | +1.000 (n=9) | +0.680 (n=39, agree 84.6%) |
+| ''current'' vs rest | the same rows, collapsed to the one distinction every published accuracy figure rests on | +0.400 (n=6) | +0.710 (n=24) | +1.000 (n=9) | +0.708 (n=39, agree 87.2%) |
+The **ALL** row is in the table to be argued with, not to be quoted. More than
+half of webXray's cells on this subsample (23 of 35) are ''absent'', and
+''absent'' is re-derived from the list file by the merge script rather than
+judged, so the two raters cannot disagree on them. Counting those as agreement
+lifts webXray's kappa from +0.027 to +0.525 without a single extra judgement
+being made. Any inter-rater figure computed over a verdict set that includes a
+mechanically-forced category is inflated by exactly this much, and that is the
+form most such figures take.
+=== The result, in one sentence ===
+**On the rows the brief actually asks about, two models given identical inputs
+agree on 58.5% of 65 verdicts, kappa +0.400 (95% CI +0.241 to +0.551) — and
+about two-thirds of the disagreement is one rater finding a source where the
+other did not, rather than the two reading the same evidence differently.**
+=== What the disagreements are made of ===
+^ List ^ NAMED cells ^ agreed ^ differ on //sourcing// (one rater ''unknown'') ^ differ on //reading// (two different verdicts) ^
+| webXray | 12 | 3 | 4 | 5 |
+| Tracker Radar | 30 | 21 | 6 | 3 |
+| Disconnect | 23 | 14 | 7 | 2 |
+| **Pooled cells** | **65** | **38** | **17** | **10** |
+**17 of the 27 disagreeing cells (63.0%)** are one rater reaching ''unknown''
+where the other reached a verdict. The row-level picture is starker: rater 1
+left **10 of 35** rows unresolved and rater 2 left **2**, and they made the same
+resolved/unresolved call on 27 of 35. Rater 2 settled every one of Tracker
+Radar's 30 named rows; rater 1 settled 24.
+That decomposition is the main finding of this section and it changes what the
+figures mean. Restricted to rows **both** raters settled, the pooled kappa is
++0.680 with 84.6% raw agreement, and on ''current''-versus-rest it is
++0.708 with 87.2%. So the **verdict vocabulary is reasonably reliable and
+the sourcing effort is not**: given the same evidence, two models mostly read it
+the same way; given the same brief, they do not find the same evidence. For a
+reader planning their own adjudication, that says the thing to specify and
+control is not the label set but how hard and how uniformly the adjudicator is
+made to look.
+=== Which way rater 2 moves the numbers ===
+Unweighted ''current'' share among each rater's own adjudicable rows on this
+subsample. **This is not the estimator** — 35 rows, no stratum weights, a
+different denominator per rater — and it is here to answer one question only:
+does the second rater make the published rates look better or worse?
+^ List ^ rater 1 ^ rater 2 ^
+| webXray | 50.0% (4/8) | 70.0% (7/10) |
+| Tracker Radar | 66.7% (16/24) | 66.7% (20/30) |
+| Disconnect | 80.0% (8/10) | 76.5% (13/17) |
+Rater 2 is **more** favourable to webXray, identical on Tracker Radar and
+slightly less favourable to Disconnect. So the disagreement does not point in
+the direction of "the published accuracy figures are too kind"; if anything the
+opposite for the list the page is about. That is a statement about direction on
+rows, not a re-estimate, and it must not be read as one.
+=== The retired verdict came back ===
+''self-named'' was retired mid-run in the first sitting because it fired on
+company names (C3.4). It is still in the published brief, and rater 2 —
+which had never seen that decision — reached for it **6 times in 105 cells**
+against rater 1's 4, agreeing with rater 1 on 3 of them. Two models
+independently finding the same wrong verdict attractive is evidence that the
+defect was in the **brief** rather than in one adjudicator, which is what C3.4
+claimed on weaker grounds. It also means the brief published on the appendix
+page still contains a verdict the study does not use, and a reader reusing it
+should delete that bullet.
+=== Rater 2's own citation check ===
+''owner_verify_sources.py'' was re-run over rater 2's 35 rows from an **empty
+cache**. It returned three failures — and **all three turned out to be limits of
+the instrument rather than bad citations**. That is the uncomfortable part: on a
+-row sample the checker's false-failure rate was 3 in 33, and one of the three
+was acted on before a reviewer caught it. After the fixes below the tally is
+**OK 30, OK-MARKUP 1, FETCHFAIL 2, NOSOURCE 2 — zero NOTFOUND**, so every one of
+rater 2's 33 sourced rows has a verified citation. The two NOSOURCE rows are the
+two rater 2 declared unresolved, which is correct. What each failure actually
+was (''scripts/owner_irr_corrections.py''):
+^ Row ^ Machine verdict ^ What a hand check found ^ Action ^
+| ''adgrx.com'' | FETCHFAIL | ''samsungads.ca'' serves a Let's Encrypt certificate that **expired 2026-03-12** and has not been renewed, so ''curl -sL'' and Chromium both refuse it; ''curl -sk'' returns 59,465 bytes carrying the quoted sentence verbatim | **kept** |
+| ''cedscdn.it'' | FETCHFAIL | the rater cited ''%whois://whois.nic.it/cedscdn.it%'', a scheme neither instrument can open. A port-43 query returns ''Organization:     CED DIGITAL & SERVIZI Srl'' verbatim | **kept**; the gap is the brief's, which offers ''%tls://%'' and no whois equivalent |
+| ''kameleoon.io'' | NOTFOUND in bytes **and** NOTFOUND in a rendered browser | **the citation is sound and the check was wrong.** The quote is a ''%<script src>%'' line, and both instruments are blind to markup by construction: ''norm()'' strips ''%<...>%'' before searching and ''owner_verify_rendered.mjs'' reads ''document.body.innerText'', which excludes script elements. The string is verbatim in the 27,121 bytes ''curl'' returns | **kept**; this row had already been demoted on the machine verdict and the demotion is **reversed** |
+**This one was published as a demotion first and corrected after a reviewer
+fetched the page.** It is left in the record because it is the most instructive
+failure of the sitting: two independent checks agreed, and they agreed because
+they share an assumption — that evidence lives in a document's //text//. A second
+instrument that fails the same way is not a second instrument. The verifier now
+has an ''OK-MARKUP'' status for exactly this case.
+Because nothing is demoted, the corrected and uncorrected inter-rater figures are
+**identical** — ''owner_irr_corrections.py'' reports "verdict cells changed: 0"
+and prints a warning if that ever stops being true. Both the headline table and
+the correction pass therefore rest on the same 65 cells: Disconnect +0.460,
+pooled +0.400.
+=== Three defects this pass found in the instruments ===
+  * **''owner_verify_sources.py'' was deleting every non-Latin character.** Its ''norm()'' ended in ''%re.sub(r"[^a-z0-9]+", " ", s)%'', so a Cyrillic, Greek, Hebrew or CJK quote was checked only on whatever Latin substring it happened to contain — and a wholly non-Latin quote normalised to the empty string and was reported ''NOQUOTE'', a status that reads like "the row has no quote" rather than "I could not check this quote". ''owner_verify_rendered.mjs'' has always used ''%[^a-z0-9À-]+%'', so the browser pass and the bytes pass disagreed about what a character is. Found because rater 2 cited a wholly Chinese sentence for ''qbox.me''. The Python class is now the JavaScript one. **Consequence for the first sitting:** re-running the fixed script over all 175 rows moves three statuses, and only one of them is this fix biting — ''compass-fit.jp'', whose Japanese quote was previously verified on the six letters ''COMPASS-FIT'' alone and is **not** verbatim on the cited page. The mismatch is **one character**: the page's body reads ''台湾国内での、ネイティブ型広告配信サービス…'' and the row's quote drops the full-width comma after ''台湾国内での''. Everything else, including the ''致しました。'' ending, is on the page. So this is a transcription slip rather than a paraphrase, and the row's //claim// is unaffected — MicroAd's own newsroom page carries the partnership, the service name and Taiwan — so no published rate moves. The row needed a ''requote'' in ''owner_corrections.py''; that was **done on 2026-09-17** and is recorded in [[#C3.9 The 2026-09-11 re-run, applied (2026-09-17)|C3.9]]. //A first draft of this bullet said the page's headline ends ''の提供を開始'' and implied the ''致しました。'' ending appears nowhere; a reviewer found the full sentence in the body and the real one-comma difference.// Three rows in the 175 carry a non-Latin quote, not two: ''compass-fit.jp'', ''tns-counter.ru'' (Russian) and ''bilibili.com'' (a certificate subject, ''%%O = 上海幻电信息科技有限公司%%'', which the broken class reduced to "o cn bilibili com"). The other two verify clean under the fix. **The fix is not symmetric, and that is worth stating:** the character class keeps everything in U+00C0–U+FFFF, which includes CJK punctuation, while ASCII punctuation is still normalised to whitespace. So ''%%norm("Brightcove Inc, describes") == norm("Brightcove Inc describes")%%'' is true and the same comparison over ''%%国内での、ネイティブ%%'' versus ''%%国内でのネイティブ%%'' is false. A Latin quote missing a comma passes; a Japanese one does not. That asymmetry is why ''compass-fit.jp'' failed on one character, and closing it would mean stripping U+3000–U+303F and U+FF00–U+FF65 as well.
+  * **Neither citation check can see a quote that is markup.** ''owner_verify_sources.norm()'' strips ''%<...>%'' before searching, which is right for a prose quote and fatal for a citation to a ''%<script src>%'', a ''%<link>%'' or an attribute; ''owner_verify_rendered.mjs'' reads ''document.body.innerText'', which excludes script elements for the same reason. So the two passes, which exist to catch each other's blind spots, share this one — and a **false** NOTFOUND from both was acted on. The bytes pass now also searches an un-stripped copy and reports ''OK-MARKUP'' when that is where the quote lives. Re-running the fixed script over the **first** sitting's 175 rows moves ''fontawesome.com'' from NOTFOUND to OK-MARKUP; that row was already rescued by the browser pass in ''owner_citation_audit.py'', so the published "0 uncited rows" is unchanged and now has two independent routes rather than one.
+  * **The citation check is time-sensitive, but less of what moves is the web than it looks.** Re-verifying the same 175 rows on 2026-09-11 moves **3 statuses of 175**: ''compass-fit.jp'' ''OK'' → ''NOTFOUND'' (the non-Latin fix biting), ''cartfulsolutions.com'' ''OK'' → ''OK-WINDOW'' (the web moving: the footer was restructured and the page grew from 15,933 to 23,044 bytes), and ''fontawesome.com'' ''NOTFOUND'' → ''OK-MARKUP'' (the instrument changing, per the bullet above). ''brightcove.net'' is ''OK-WINDOW'' in **both** retained artefacts; it FETCHFAILed on one execution of the script and answered on the next, which is a host briefly unreachable and not a status that moved. Its quote elides about 20 characters the page has carried since 2026-09-05, so it is an adjudicator elision of the same class as ''compass-fit.jp'' — rescued only because a long quote gets the 60-character window fallback and a 40-character one does not. A "zero uncited rows" claim is therefore a claim about a date, and about a quote length; this page should be read as saying the citations verified on 2026-09-05, on 2026-09-11 and — for three named rows — on 2026-09-17, rather than that they verify.
+=== What this section does and does not license ===
+  * **It does not rehabilitate the per-list accuracy estimates, and it does not overturn them either.** Kappa is about two readings of the same rows; it says nothing about whether either reading is right. A sample on which both raters are confidently wrong scores kappa 1.
+  * **webXray's +0.027 is the weakest cell in the table and n=12 is why.** Its interval, -0.200 to +0.280, includes both "no agreement beyond chance" and "moderate agreement". The honest reading is that this subsample cannot characterise agreement on webXray at all; it took 12 named rows to the 30 Tracker Radar got because webXray names an owner for 2.1% of the frame. Closing that needs a subsample drawn from webXray's coverage rather than from the 175.
+  * **The pooled column counts cells, not domains.** One domain contributes up to three verdicts and they are not independent — the same fetched document usually settles all three. The pooled interval is therefore narrower than it should be, and it is in the table for its point estimate, not its width.
+  * **Kappa is not read here against the usual verbal bands.** They are a convention, not a threshold, and at these marginals — one category carrying most of the mass — kappa is driven as much by the marginals as by the disagreements. Raw agreement is printed beside every kappa for that reason, and ''ADJUDICABILITY'' for Tracker Radar is the extreme case: rater 2 settled all 30 rows, so one rater used a single category, and kappa is **0.000 by construction** at 80.0% raw agreement. A zero there means "undefined", not "no agreement".
+  * **Six days separate the two ratings.** Corporate ownership rarely moves in six days, and the frozen probe removes the mechanical part of the drift, but a page that changed in that window is indistinguishable from a rater that read it differently.
+==== C3.9 The 2026-09-11 re-run, applied (2026-09-17) ====
+The ''norm()'' fix of 2026-09-11 ([[#C3.8 Inter-rater agreement: a second model on a random fifth (2026-09-11)|C3.8]]) was made while auditing the //second// rater's rows. It also invalidates the first sitting's pass over every row whose quote is not Latin script, because the broken checker had reported those rows OK on whatever Latin substring they happened to contain. The re-run over all 175 first-sitting rows was made on **2026-09-11** and its consequence recorded as open work. This section is where that work was done, on **2026-09-17**.
+=== What was run, and when ===
+The 175-row re-run is **not** today's work. It was run on 2026-09-11, against an empty cache, as:
+<code>
+python3 scripts/owner_verify_sources.py --rows out/adj_rows_scored.json \
+    --cache out/irr/verify_cache_s1_refix --out out/irr/verify_s1_refix.json
+</code>
+Its artefacts — ''out/irr/verify_s1_refix.json'' and the 133 cached bodies under ''out/irr/verify_cache_s1_refix/'' — all carry 2026-09-11 timestamps, and **that command no longer reproduces the result below**: ''out/adj_rows_scored.json'' now carries the corrected quote, so ''compass-fit.jp'' comes back ''OK''. What was done on 2026-09-17 is the correction, the chain re-run, and a **three-row** live re-fetch of ''compass-fit.jp'', ''brightcove.net'' and ''cartfulsolutions.com'' — not a fourth pass over 175 sources. Compared against ''out/verify_final.json'' (2026-09-05), the 2026-09-11 run moves **three statuses**:
+^ Domain ^ 2026-09-05 ^ 2026-09-11 ^ Which of the three is it? ^
+| ''compass-fit.jp'' | ''OK'' | ''NOTFOUND'' | **the fix biting.** Its quote is Japanese; the broken ''norm()'' had checked it on the six Latin letters ''COMPASS-FIT'' |
+| ''cartfulsolutions.com'' | ''OK'' | ''OK-WINDOW'' | the web moving. The footer was restructured and the page grew from 15,933 to 23,044 bytes; the quote still matches on a 60-character window |
+| ''fontawesome.com'' | ''NOTFOUND'' | ''OK-MARKUP'' | the instrument changing. ''OK-MARKUP'' did not exist on 2026-09-05; the row was rescued by the browser pass then, and by the bytes pass now |
+''brightcove.net'' is ''OK-WINDOW'' in **both** files, so it is not a fourth mover. C3.8's bullet recorded it as ''OK-WINDOW'' → ''FETCHFAIL''; that was an earlier execution of the same script, whose JSON was overwritten by the retained one. The host was briefly unreachable, which is the bullet's point, but the retained artefact does not show it.
+=== What was done about each ===
+  * ''compass-fit.jp'' — **requoted** in ''scripts/owner_corrections.py'' (the eleventh correction; see [[#C3.5 Every citation re-fetched, and the eleven fixed by hand|C3.5]]). The replacement is the body sentence in full, ''%%株式会社マイクロアド…開始致しました。%%'', copied out of the fetched page. It is not the //shortest// span on the page naming a MicroAd entity and the service — the page title is 71 normalised characters and names both — it is the one that carries the **parent and the operating subsidiary together**, which is what ''owner_today = MicroAd, Inc.'' rests on where the old quote named only the service. It verifies ''OK'' against a live fetch of 2026-09-17 (39,110 bytes).
+  * ''cartfulsolutions.com'' — **nothing.** ''OK-WINDOW'' is a passing status by design: a long quote gets spliced by boilerplate, and requiring the literal string would fail every such row. Re-fetched 2026-09-17 (23,044 bytes) and it still names Cartful Solutions, Inc. Requoting it would chase the web and have to be redone the next time the footer is edited.
+  * ''brightcove.net'' — **nothing, but not for the reason it first looks.** Re-fetched 2026-09-17 (806,613 bytes) and still ''OK-WINDOW''. Its quote is **not** a victim of the page changing: the parenthetical it elides, ''%%The Brightcove Privacy Policies ("Privacy Policies") describe Brightcove Inc.%%'', is byte-identical in the 2026-09-05, 2026-09-11 and 2026-09-17 bodies. The row is an adjudicator elision of the **same class** as ''compass-fit.jp'', and the only reason one was requoted and the other was not is **quote length**: at ~110 characters brightcove's clears the 60-character window fallback, and at ~40 compass-fit's cannot. That is the honest line, and it is a property of the checker rather than of the evidence.
+  * ''fontawesome.com'' — **nothing.** The row was already cited and already counted as cited; what changed is that the bytes pass can now see the evidence the browser pass saw on 2026-09-05.
+=== What moved downstream, and what did not ===
+The requote changes one ''quote'' field. It changes **no verdict, no denominator and no rate**, and that is checked rather than asserted: the whole chain was re-run and both estimator outputs are byte-identical to the committed ones.
+^ Re-run ^ Result ^
+| ''owner_corrections.py'' | output changes: ''10 corrections … requote=4'' → ''11 corrections … requote=5'', plus the new row's block |
+| ''owner_verdict_fixes.py'' | byte-identical |
+| ''owner_random_sample.py'' (2026-09-05 rows) | **byte-identical** — all 31 figures unchanged |
+| ''owner_tail_merge.py'' | unchanged but for the ''wrote …'' line; the 135 non-residue rows carry the new quote through, and its own drift gate passes |
+| ''owner_tail_corrections.py'' | unchanged but for the ''wrote …'' line |
+| ''owner_random_sample.py'' (2026-09-11 rows) | **byte-identical** — the estimates the pages actually publish |
+| ''report_tail_pass.py'' | byte-identical |
+| ''owner_citation_audit.py'' | byte-identical — but see below, it reads the 2026-09-05 artefacts and cannot see any of this |
+| ''check_tail_figures.py'' against [[design:ownership_resolution]] | 80 figures, all present and matching |
+| ''check_tail_figures_mutations.sh'' | identical output, 24 of 24 caught |
+| ''check_owner_sample_figures_mutations.sh'' | identical output, 10 of 10 caught |
+Two ''wrote …'' lines in the published output blocks change from ''/tmp/discard_rows.json'' and ''/tmp/discard2.json'' to the paths the documented commands name. The committed blocks had been produced by a verification re-run that deliberately wrote to ''/tmp'' so as not to clobber; the published command line said otherwise. The re-run writes where the command says, so block and command now agree.
+=== What the citation audit still says, and what it would say ===
+''owner_citation_audit.py'' reads ''out/verify.json'' and ''out/verify_final.json'', both 2026-09-05 artefacts. Re-running it today therefore reproduces its old output byte-for-byte and **cannot see** either the fixed checker or this requote. C3.5's table is labelled 2026-09-05 for that reason. Recomputed against the 2026-09-11 statuses **with the requote applied**, the split would be:
+^ State ^ 2026-09-05 (published) ^ with the fix and the requote ^
+| whole quote | 119 | 118 |
+| ellipsis fragments | 8 | 8 |
+| 60-character window | 5 | 6 |
+| markup only | — | 1 |
+| cited in the bytes, total | 132 | **133** |
+| cited only in a rendered browser | 3 | 2 |
+| uncited | **0** | **0** |
+The movement is ''cartfulsolutions.com'' whole → window and ''fontawesome.com'' browser → markup. The only rows the bytes pass still fails are ''force.com'' and ''flashtalking.com'', both carried by the browser pass on 2026-09-05. **No row is uncited under either reading**, which is why the published table was dated rather than replaced: replacing it would mix a 2026-09-05 census with a 2026-09-11 one.
+=== Judgement calls ===
+  * **The correction was appended last in ''CORRECTIONS'', after the ''demote''.** The dict's order is the printed order, and this correction is eleven days younger than the other ten. Grouping it with the other ''requote'' entries would have read as if it had been found with them, which is the one thing about it worth knowing.
+  * **It is counted as a requote but described as a transcription slip.** The remedy is identical to the other four; the defect is not. C3.5 now says four paraphrases and one dropped character rather than five of one thing.
+  * **''brightcove.net'' was left alone on a rule the checker enforces, not on a judgement about the evidence.** Requoting every row whose quote is an elision would mean requoting rows the checker passes, which is a different and much larger piece of work; it is recorded here as a known, quantified gap rather than done.
+  * **The section-G heading on the appendix is now computed** from ''len(CORRECTIONS)'' rather than typed. It said "the ten citations fixed by hand" and would have been wrong the moment this correction landed.
+  * **The reviewer's confirmation in [[#Reviewers, second sitting (2026-09-05)|the 2026-09-05 reviewer log]] still says "10 corrections".** It is left as written: it is a dated record of what that reviewer verified on that day, and editing it would make the log agree with a state it never saw.
+  * **Regenerating the appendix also publishes an unrelated fix.** ''owner_verify_sources.py'' gained a WHOIS retry on 2026-09-11 that the appendix was never regenerated for, so its published block was 524 bytes behind the file. It is current now, and the manifest hash for section E moves with it.
+=== Reviewers (2026-09-17) ===
+^ Finding ^ Decision ^
+| **The published mutation count is wrong: "14 of 14 caught" for ''check_tail_figures_mutations.sh''; the harness catches 24** | **ACCEPTED.** Reproduced: the committed output ends ''mutations caught: 24; survived or broken: 0'' and has 24 ''caught'' lines. A transcription error introduced by this sitting — there was never a 14-mutation version. Fixed |
+| **C3.9 dated the 175-row re-run to 2026-09-17; it ran on 2026-09-11** | **ACCEPTED, and the most useful finding.** Verified from mtimes: ''verify_s1_refix.json'' and all 133 cache bodies are 2026-09-11. Only three rows were fetched on 2026-09-17. The heading and the "What was run" block are rewritten; C3.5's table row is too |
+| **The ''brightcove.net'' rationale is factually wrong.** The elided parenthetical is in the 2026-09-05 body, so it is an adjudicator elision, not the web moving | **ACCEPTED.** Verified against all three cached bodies: byte-identical. The real distinction is quote length against the 60-character window, and both C3.8 and C3.9 now say so |
+| **The fix introduced a Latin/CJK asymmetry the page does not disclose**: ASCII punctuation normalises to whitespace, CJK punctuation does not | **ACCEPTED.** Reproduced against the committed module. Disclosed in C3.8's first bullet, with what closing it would take |
+| **"''tns-counter.ru'', the other non-Latin row" is false**: ''bilibili.com'' is a third | **ACCEPTED.** Three rows in the 175 carry non-Latin text. Both others verify clean under the fix, so no status moved, but the sentence claimed a census it had not done |
+| **"the shortest contiguous span naming both MicroAd and COMPASS-FIT" is false** | **ACCEPTED.** The page title is 71 normalised characters and names both. The real reason for the body sentence is that it names parent and subsidiary together. Fixed in the script's note and here |
+| **C3.5's "10 of 175 rows had their source or quote replaced, 9 of those 10" is stale** | **ACCEPTED.** 11 and 10 of 11 |
+| **Section F still listed the requote as work this sitting did not apply** | **ACCEPTED.** Closed with a pointer here |
+| **The citation-audit table is an undated 2026-09-05 figure that the fix moves** | **ACCEPTED.** Dated, and the recomputed split published above rather than substituted for it |
+| **''check_owner_sample_figures.py'' fails 16 figures against ''design:ownership_resolution''** | **Not a defect, and confirmed pre-existing** by re-running it against the pre-change output. That page publishes the second pass's estimates; this guard checks the 2026-09-05 ones, and its own mutation harness runs against ''pages/wx_new.txt'' and passes |
+| Probe artefacts live in ''out/rq/'' rather than a documented path | **Noted, not fixed.** They are a three-row scratch check, not an artefact any figure rests on; the figure they support (39,110 bytes, ''OK'') is reproducible from the command in this section |
 ===== D. External sources: verified, and rejected =====
@@ Line 184: / Line 765: @@
 | ''thezedwards/webXray'' last pushed 2021-03-04, 19 forks, newest fork activity 2023-03-12 | GitHub API repo + ''/forks?per_page=100'' |
 | the surviving README still says ''git clone https://github.com/timlib/webXray.git'' | fetched ''raw.githubusercontent.com/thezedwards/webXray/master/README.md'' |
-| PolyForm Strict License 1.0.0, no redistribution, noncommercial permitted incl. research organisations | fetched ''LICENSE.md'' from the same repo and read the licence text itself, not a summary |
+| PolyForm Strict License 1.0.0 on the ''thezedwards'' snapshot | fetched ''LICENSE.md'' from that repo and read the licence text itself, not a summary. **This is not webXray's final licence** — see the next row |
+| webXray was relicensed to GPLv3 (2021-06-14, commit ''245ec5d7'') and then to **MIT** (2023-02-01, commit ''73fe0fc9'', authored by "Tim Libert") | fetched ''api.github.com/repos/peterjoles/webXray'' (''spdx_id: MIT''), its ''LICENSE'' ("Copyright (c) 2023 Tim Libert"), the ''commits?path=LICENSE'' history, and ''compare/master...peterjoles:master'' (''ahead_by: 36''). The page's first version asserted webXray "is not open source and redistributing it is prohibited" from the ''thezedwards'' snapshot alone, which was the most restrictively licensed copy in existence — the exact mistake the page tells readers to avoid |
 | PolyForm Strict 1.0.0 is still the current version and has no SPDX identifier | ''polyformproject.org/licenses'' lists ''strict/1.0.0'' and no later Strict version; SPDX's own ''license-list-data'' JSON carries only ''PolyForm-Noncommercial-1.0.0'' and ''PolyForm-Small-Business-1.0.0'' |
 | raw CDP, no Selenium | ''webxray/ChromeDriver.py'' imports ''create_connection'' from ''websocket''; ''requirements.txt'' pins ''lxml==4.6.2'', ''psycopg2-binary==2.8.6'', ''textstat==0.7.0'', ''websocket-client==0.57.0'' |
@@ Line 195: / Line 777: @@
 | Disconnect last commit 2026-08-07, CC BY-NC-SA 4.0, **GPL-3.0 until 2020-06-24** | GitHub API; then the commit history of ''LICENSE'' — GPLv3 in the initial commit ''89d421e8'' (2015-10-13), rewritten to CC BY-NC-SA by two commits on 2020-06-24 |
 | Firefox ships Disconnect's ''services.json'' via Mozilla's ''shavar-prod-lists'' | ''mozilla-services/shavar-prod-lists'' README: Firefox's ETP "rely on lists of trackers maintained by Disconnect… Mozilla does not maintain these lists"; ''disconnect-blacklist.json'' is "a version controlled copy of Disconnect's list of trackers" |
-| Ghostery ''trackerdb'' 2026-08-06 (MIT); WhoTracks.me data repo 2026-08-04 ("July update"); ''whotracks.me'' now redirects to ''ghostery.com/whotracksme'' | GitHub API on ''ghostery/trackerdb'' and ''whotracksme/whotracks.me'' (the ''ghostery/whotracks.me'' path 301s); followed the site redirect |
+| Ghostery ''trackerdb'' 2026-08-06, **CC BY-NC-SA 4.0**; WhoTracks.me data repo 2026-08-04 ("July update"); ''whotracks.me'' now redirects to ''ghostery.com/whotracksme'' | GitHub API on ''ghostery/trackerdb'' and ''whotracksme/whotracks.me'' (the ''ghostery/whotracks.me'' path 301s); followed the site redirect. The licence is from the repo's own ''LICENSE'' and ''package.json'' (''"license": "CC-BY-NC-SA-4.0"''), **not** the GitHub API, which returns no SPDX id — a first pass read "MIT" off the API and was wrong |
 | WhoTracks.me paper is arXiv 1804.08959, v2 revised 2019-04-25 | ''export.arxiv.org/api/query?id_list=1804.08959'' — title, authors and ''updated'' field read from the Atom response |
 | Crunchbase API is paid | ''api.crunchbase.com/api/v4/…'' → HTTP 401 ''"Unauthorized user_key"'' |
@@ Line 208: / Line 790: @@
   * **''toolness/webxray'' (Mozilla's Web X-Ray Goggles, 2010–2013)** — a genuinely different project that shares the string. Checked and excluded, because a name search finds it first: GitHub's repository search for "webxray" returns 18 repositories, of which the majority are unrelated (X-Ray Goggles, an X-ray physics database, an offensive web scanner, a Shopify evaluator). Recorded here so the next run does not include it.
   * **Cinco Network's WebXRay (Held, IJNM 1998)** — the homograph {[jakaria2024_dots]} cites. Excluded from the tool discussion; mentioned only as the miscitation it is.
-  * **''carlsaturnino/webxray-docker'' on Docker Hub** (399 pulls, ''last_modified'' 2024-10-16) was not presented as a maintained image: the timestamp is a metadata bump on an image registered 2016-04-26, and ''hub.docker.com/v2/repositories/carlsaturnino/webxray'' and ''.../timlib/webxray'' both 404. The page says no maintained image exists.
+  * **''carlsaturnino/webxray-docker'' on Docker Hub** (a counter: 399 pulls on 2026-08-17, 405 on 2026-09-05; ''last_modified'' 2024-10-16) was not presented as a maintained image: the timestamp is a metadata bump on an image registered 2016-04-26, and ''hub.docker.com/v2/repositories/carlsaturnino/webxray'' and ''.../timlib/webxray'' both 404. The page says no maintained image exists.
+===== D2. Wayback: dating webXray's disappearance =====
+The first sitting recorded two questions as **unknown** because
+''web.archive.org'' returned 502/503 for the whole of 2026-08-17 — nine
+attempts, two independent fetch paths, about thirty minutes apart. The Archive
+was up on 2026-09-05 and answers both. Both answers are **bounds, not dates**,
+because the Archive has no capture inside either window, and they are published
+as bounds.
+''scripts/wayback_webxray.sh'' holds the seven CDX queries; its unedited output is
+on [[provenance:programming:crawler:webxray:random_sample|the appendix page]].
+The method has two traps worth stating.
+  - ''collapse=digest'' returns one row per distinct response body, which is what makes a 200→404 transition visible in a handful of lines — and it also **hides how many captures there were**. So every transition found that way is re-queried **uncollapsed** over a narrow window before any date is published. That is query 3, and it is what turns "somewhere in 2023" into a two-row bracket with nothing between.
+  - **A CDX row is a capture, not a page.** The first row of a domain's history is very often a registrar or host parking page, and reading a date off the index without fetching the capture is how a first draft of this section dated the pre-webXray product to 2010-10-08 when that capture is a DreamHost "Coming Soon" page. Every date below was read out of the fetched ''id_'' body, not out of the index. **Query 7 was added after review** for the same class of reason: queries 5 and 6 stop in 2024, so the published output did not contain the 2026 capture the ''webxray.ai'' bound rests on — a bound whose evidence is not in the output is an uncited figure however true it is.
+^ Question ^ Answer ^ Evidence ^
+| When did ''github.com/timlib/webXray'' first 404? | between **2023-03-31** and **2023-11-15**. Not narrower | last archived 200 ''20230331144433''; first archived 404 ''20231115075841''; the uncollapsed query over ''from=20230101&to=20240601'' returns **exactly those two rows**, so there is no capture between them |
+| …anything that narrows it? | no | the ''github.com/timlib'' profile has captures at 2022-10-27 and then not again until 2025-08-16. The one corroborating fact is outside the Archive: ''peterjoles/webXray'''s last push is **2023-03-12**, i.e. the fork was taken and then upstream went |
+| When did ''github.com/timlib/webXray_Domain_Owner_List'' first 404? | between **2022-10-06** and **2025-01-18**. A much wider window; do not narrow it | last archived 200 ''20221006015001''; first archived 404 ''20250118193856'' |
+| What did ''webxray.org'' serve, and when did the demo die? | a live demo search engine from 2022-01-23 to at least **2024-03-28**; ''401 Authorization Required'' on **2024-05-24**; a 301 to ''webxray.ai'' from **2024-07-24** to at least 2026-02-05 | the ''20240328005456id_'' capture carries "The search engine is drawn from scans of 100,000 sites" and the company drop-down (33Across, 360, 51La, AdGear, Adition Technologies, …). ''20240524194734'' is a 401 from nginx/1.18.0. Every capture from ''20240724105426'' on is a 301 whose ''Location'' is ''https://webxray.ai/'' |
+| When did the domain come back off ''webxray.ai''? | between **2026-02-05** and 2026-08-17. Cannot be narrowed | query 7 (uncollapsed, 2025–2026) returns 14 rows, every one of them the same 301 digest ''OM6ALWQTEGJ5NTY5RYSC2EXCKJOAHOVU'', the last being ''20260205035830''. After that the Archive has **no capture at all** of ''webxray.org'', so it has never seen the placeholder served today |
+Two further findings from the same captures, both of which changed the content
+page:
+  * **''webxray.org'' was a different product before it was webXray.** From **2010-11-08** to at least **2011-09-27** it served a page whose ''%%<title>%%'' is "Web Xray - Free Website Data Mashup Tool" and which carries a "beta 0.5" badge beside the logo — a public directory of analysed sites with Google AdSense units (''google_ad_client = "pub-2283090127531811"'') and a Clicky analytics call. Nothing to do with third-party request measurement. By the next capture, **2014-12-18**, the domain served webXray; there is no capture in between. Recorded because a reader dating the project from its domain registration would be off by four years.\\ **A first draft of this bullet said 2010-10-08**, taking the first row of the CDX query as the first Web Xray page. It is not: ''web/20101008070114id_'' is a DreamHost **"Coming Soon" parking page** ("The DreamHost customer who owns webxray.org has not yet uploaded their website"). Caught by a reviewer who fetched the capture instead of reading the timestamp, which is the whole lesson — a CDX row is a capture, not a page.
+  * **A fourth licence state, and it is a //reversion//.** The 2015-05-30 capture describes webXray 1.0's cost as "Free!*" and "(*Subject to terms of the GNU Public License.)", linking that phrase to ''gnu.org/copyleft/gpl.html''. That URL served **GPLv3** in 2015 (archived at ''web/20150531125231'' as "Version 3, 29 June 2007") and still resolves to ''gnu.org/licenses/gpl-3.0.html'' today — a **302** on the path, after a 301 from the bare ''gnu.org'' host to ''www.gnu.org''. So the sequence is GPLv3 (2015) → PolyForm Strict (by 2021-03) → GPLv3 again (2021-06-14) → MIT (2023-02-01), and the 2021 commit titled "Now open-source" was a **return**, not a first opening. The content page's licence paragraph was rewritten on that basis.
+Two honesty notes on this evidence:
+  * The 2015 quote **joins two adjacent ''%%<p>%%'' elements** — the page reads ''%%<p>Free!*</p>%%'' then ''%%<p>(*Subject to terms of the <a …>GNU Public License</a>.)</p>%%''. It is the visible text as a reader saw it, not a contiguous string in the source.
+  * The licence claim rests on **the project's own website**, not on a repository ''LICENSE'' file. No snapshot of ''timlib/webXray'' from 2015 survives — the earliest mirror is the 2021 ''thezedwards'' copy — so what the repository itself said in 2015 is not recoverable and is not asserted.
 ===== E. What could not be established =====
-  * **When ''github.com/timlib/webXray'' first returned 404, and what ''webxray.org'' used to contain.** ''web.archive.org'' was returning 502/503 for the whole of 2026-08-17 (nine attempts, two independent fetch paths, ~30 minutes apart). Both questions are marked unknown on the page rather than estimated. **What would close it:** re-run the Wayback queries when the Archive is up — ''http://archive.org/wayback/available?url=github.com/timlib/webXray&timestamp=2022'' and the same for ''webxray.org''. The only bracket available now is indirect: forks show a merge of a ''timlib''-owned branch as late as 2022-06-28, so upstream was still reachable then.
+  * **ANSWERED — When ''github.com/timlib/webXray'' first returned 404, and what ''webxray.org'' used to contain.** Answered on 2026-09-05, when the Archive was up — see [[#D2. Wayback: dating webXray's disappearance|D2]]. The bullet is rewritten rather than deleted so the gap and its closure are both on the record. What the first sitting wrote: "''web.archive.org'' was returning 502/503 for the whole of 2026-08-17 (nine attempts, two independent fetch paths, ~30 minutes apart). Both questions are marked unknown on the page rather than estimated." What is **still** not established: both answers are **bounds**, not dates — the repository 404ed somewhere between 2023-03-31 and 2023-11-15, the owner-list repository between 2022-10-06 and 2025-01-18, and the domain came back off ''webxray.ai'' between 2026-02-05 and 2026-08-17. The Archive has no capture inside any of the three windows, so **nothing short of a source outside the Archive will narrow them**, and none is known.
   * **Whether a paper documents policyXray specifically**, as distinct from the module description in webXray's README. Not settled; scholarly indices were not searched exhaustively. {[libert2018_automated]} is the closest thing and is what the page cites.
   * **Whether DuckDuckGo has published a paper on Tracker Radar's methodology.** Nothing found; the documentation is repo-native plus a vendor blog post. The page says "no paper" on that basis, which is an absence-of-evidence claim and is phrased as one.
@@ Line 218: / Line 835: @@
   * **The ''1,096 of 1,146'' sentence in {[steffens2021_blockparty]}** — see A.17. Unresolvable from the text.
   * **The list-age and missing-hostname columns for ''timlib/webXray'' in {[mcquistin2023_psl]}'s table.** The ''.cols'' rendering splits that table into separate column runs and the numeric columns cannot be aligned to the repository names reliably. Only the verifiable part is on the page — that webXray appears in the table, in the "Production" category, with 27 stars. **What would close it:** read the published PDF's table directly.
-  * **A random-sample accuracy rate for any of the three lists.** Would need a hand-adjudicated random sample of a few hundred domains, i.e. tens of hours at the rate section C ran at. Worth doing and not done; the page is explicit that its 30 rows cannot substitute.
+  * **Whether any //non-numeric// claim on this page has drifted from its script.** ''check_page_numbers.mjs'' compares numerals only, so when a reviewer's fix changed "third of four" to "second of four" on the page, the hand map in the script and its committed output kept saying "third" and every guard still passed. The same happened with "MIT" for Ghostery's licence. Both were caught by a human reviewer reading the two side by side, not by a tool, and no tool here covers that class. ''check_attributions.mjs'' closes one narrow case of it (author names). **What would close the rest:** a guard that extracts the quoted strings a page attributes to a script's hand map and diffs them against the map.
+  * **ANSWERED — A random-sample accuracy rate for any of the three lists.** Measured on 2026-09-05 — see [[#C3. The random sample (2026-09-05)|C3]]. 175 domains, 525 verdicts, 138 settled, per-list rates with bootstrap intervals. What the first sitting wrote, and it was right about the cost: "tens of hours at the rate section C ran at". What is **still** not established, and these are the parts a reader should hold against the figures:
+  * **Whether unadjudicable entries are worse than adjudicable ones.** 40 of 175 rows could not be settled to the sourcing bar, and the rate rises toward the tail. Every accuracy figure is conditional on adjudicability, and if the unsettled entries are worse — which is plausible — all three lists are flattered. This cannot be bounded from inside the sample. **What would close it:** a second, harder-sourcing pass over the 40, e.g. national register searches by company number rather than by name, and RDAP history where a registry publishes it.
+  * **ANSWERED — Inter-rater agreement on the verdicts.** Measured on 2026-09-11 — see [[#C3.8 Inter-rater agreement: a second model on a random fifth (2026-09-11)|C3.8]]. What the second sitting wrote: "Each row was adjudicated once. The citation check confirms the evidence exists and contains the quote; it does not confirm the verdict follows from it." 35 domains re-adjudicated by a different model against the identical brief; pooled kappa +0.400 over 65 named cells. What is **still** not established: agreement on **webXray specifically** — its cell is 12 rows and its interval, -0.200 to +0.280, spans everything from chance to moderate. **What would close it:** a subsample drawn from webXray's own 60-domain draw rather than from the 175, which needs roughly 40 more adjudications.
+  * **Whether the two raters are wrong together.** Kappa measures whether two readings coincide, not whether either is right; a sample both raters get confidently wrong scores 1. Nothing here bounds that, and the only thing that would is an adjudication against a source neither rater could reach for — a company register queried by company number, not by name. **What would close it:** the same harder-sourcing pass named in the bullet above, used as a third rating rather than as a rescue.
+  * **Whether the rates hold outside Tracker Radar's crawl.** The frame is Tracker Radar's ''domain_summary.json'', so entries for domains that crawl never saw are never scored. **What would close it:** repeat the draw against a frame built from a different crawl, e.g. one of the corpus papers' own request logs.
+  * **A per-stratum statement.** 12 eligible rows per tail stratum supports the pooled estimate and nothing finer. The head-versus-tail difference visible in the exclusion rates is real; a head-versus-tail difference in //accuracy// is not something this sample can assert.
   * **Whether anyone has built an LLM-based //domain//-owner resolver.** Nothing in this corpus. {[selmo2025_borges]} and {[gouda2025_prefix2org]} do the network-layer equivalent. Named on the page as a gap rather than as a method to use.
@@ Line 225: / Line 848: @@
   * **Broadening the page beyond its title.** See *Scope decision* above. Anyone would notice the page spends more words on Tracker Radar and Disconnect than on webXray; that is deliberate, because a reader arriving at a dead tool needs to be sent somewhere.
   * **Calling webXray's owner list "historical" and Tracker Radar's / Disconnect's "current practice".** Evidence: last corpus use of webXray's list is **2022**; ownership use of Tracker Radar runs 2021→2026 with 7 of its 11 papers in 2024–2026; the file itself has not changed since 2021-03-04; the tool is unobtainable; and 17 of the 25 adjudicated rows it covers are stale. The 2025–2026 corpus slice is provisional, so the year-by-year counts are given rather than a trend claim, and the page says the counts are too small to be a market share.
-  * **Not calling any list "the best".** Disconnect is current on 22 of 22 rows it covers but absent on 8 of 30 and carries no hierarchy; Tracker Radar covers everything and is stale on renames. The page gives a per-question table instead of a ranking. A ranking would have been shorter and less useful.
+  * **Not calling any list "the best".** //Restated 2026-09-05: the figures this bullet rested on were wrong — Disconnect is current on 26 of the 28 adjudicated rows and absent on none of them, not "22 of 22 … absent on 8 of 30". The conclusion is unchanged and is now better supported.// Disconnect is current on 26 of 28 rows and carries no hierarchy; Tracker Radar covers everything and is stale on renames. The page gives a per-question table instead of a ranking. A ranking would have been shorter and less useful.
   * **Publishing the ''webXray-root'' comparison even though it makes agreement worse.** The tempting move was to report only the leaf comparison, since "resolve the hierarchy" sounds like the obvious fix. Publishing both is what makes the granularity-versus-vintage decomposition visible, and that decomposition is the page's main analytical contribution.
   * **Keeping the ''null'' domain and the IPv6 literals in the write-up.** They are a defect in a widely used published dataset, they are cheap to state, and a reader who does not know about them will silently include a ''null'' row in a prevalence table.
   * **Naming the miscitation.** Recorded as a verifiable fact about a reference list, with the methodological point it illustrates, and with no characterisation of the authors. The alternative — a vague "watch out for homographs" — would have been unfalsifiable.
   * **Quoting webXray's licence at length.** It is the single most actionable fact on the page: a student who plans a study around webXray has planned a study around software they cannot legally obtain. A one-line "non-commercial licence" would have understated it.
+  * **Discoverability, and where this material really belongs.** The wiki's only treatment of domain-to-company ownership resolution now sits under the name of a dead tool, where nobody asking "how do I attribute a domain?" will look. The generic reviewer was right that the earlier rejection of a separate ''design:ownership_resolution'' page ("it would leave a red link pointing at a stub") was a weak argument — a real webXray page and a topic page can coexist. Mitigated for now by linking this page from [[Privacy:Requests]], which is where a reader asking "whose request is this?" actually lands. Splitting out a topic page is worth doing and is recorded here rather than done.
   * **''%%<WRAP important>%%'' boxes rather than ''%%<wrap todo>%%''.** No open TODOs were left on the content page: the unknowns are stated in its methodology section with what would close them, which is where a reader checking a number will look. The one real TODO — correcting [[Programming:Crawler]] — is recorded here, because it is work on a different page.
+  * **Drawing the sample at all, rather than extending the hand adjudication.** The cheap move was another 30 high-prevalence disagreements, which would have made the existing table look better sourced and would have measured nothing new. The sample is the expensive move and it is the one that produced a **result that contradicts the page's previous impression** — webXray at 69.9% domain-level and 93.9% encounter-weighted //as measured on 2026-09-05// (69.2% and 93.8% after the second pass over the residue), against "1 of 23" from the disagreement table. A page that only ever looks where it already knows the answer is where that impression came from.
+  * **Publishing the contradiction prominently instead of quietly correcting the old wording.** The content page now carries both numbers, in the same bullet, with the explanation that the first was a statement about //where it had been checked//. The alternative — deleting "badly stale wherever it has been checked" and moving on — would have hidden the most useful thing the sample found, which is how badly a selection-biased set misleads.
+  * **Not letting the new figure rehabilitate webXray.** The encounter-weighted rate — 93.9% on 2026-09-05, **93.8%** after the second pass — is a flattering number and it would have been easy to lead with it. The content page instead puts three counter-facts in the same box: coverage is 2.1%, the file is frozen so the rate only degrades, and Disconnect is better on the measured axis. That last one is now **metric-dependent**: Disconnect beats webXray on current+granularity (//p// = 0.008) and no longer at this page's threshold on current only (//p// = 0.031). The recommendation did not change; only the evidence for it got more honest, and then more qualified.
+  * **Quartiles instead of the deciles the follow-up item specified.** See [[#C3.2 Allocation: quartiles, not the deciles the brief asked for|C3.2]]. A reasonable person could have kept deciles and reported the head stratum only.
+  * **Retiring a verdict in the middle of the run.** ''self-named'' was in the brief 21 sub-agents worked from, so retiring it means re-scoring by hand after the fact — which is exactly the kind of post-hoc adjustment that should make a reader suspicious. It is published in full, with every re-score and its reason, precisely because it should. The alternative was worse: leaving a verdict that had fired on "TrustArc" and "Klaviyo" and that would have cut Disconnect's denominator by a quarter.
+  * **Accepting an OV/EV certificate subject as a primary source.** It is an identity a CA validated against a company register, and the handshake is re-done by machine. 17 of 135 sourced rows rest on it. A stricter reading would have demoted all 17 to ''unresolved''; that would have cost about an eighth of the eligible rows for no gain in truth.
+  * **Correcting one sentence on [[Programming:Crawler]], which is a different page.** That page told a reader that webXray's licence "forbids redistribution, so there is no lawful route to the code now that upstream is gone", and called ''thezedwards/webXray'' the most complete surviving copy. Both are false, and the page being published here says so — two live pages contradicting each other, with the wrong one making the more actionable claim. The scope rule for this sitting was one page and its provenance, and the broader rewrite that parent page needs is **still** the open TODO the first sitting recorded. But shipping a page while leaving a neighbour asserting its opposite is not a defensible reading of scope, so exactly that sentence was replaced and nothing else on that page was touched. Someone could reasonably have left it and filed an item.
+  * **Publishing the wrapped-list finding rather than just fixing it.** It is embarrassing and it is the most useful thing on this page for anyone else writing a long DokuWiki page. See the entry below.
+  * **Splitting the scripts and outputs onto [[provenance:programming:crawler:webxray:random_sample|an appendix page]].** 230 kB of code and raw output on top of a 140 kB page would make one 370 kB page. Raw size alone would not decide it — ''literature:bibliography'' is 388 kB and renders fine — but about 60 kB of the appendix is syntax-highlighted source, and a reader who came to check one denominator would have to scroll past all of it. The split is a readability call: **no figure is derived only on the appendix**, though several are only //shown// there — the 175-row adjudication table, the per-verdict tables, the concentration test and the raw Wayback output — and both this page and the content page send the reader there for them. Someone could reasonably have kept one page.
+  * **Rewriting two *What could not be established* bullets in place rather than deleting them.** Both are now answered. Deleting them would have erased the fact that the page shipped with those gaps and named what would close them — which is the part that made this sitting possible.
+  * **Not re-deriving the corpus figures.** This sitting measured three live files and the Wayback index. Every corpus figure on the content page is the first sitting's, against the same 5,859-paper extraction, and was not re-run. Stated here so a reader does not read "2026-09-05" at the top of the page as a date on the corpus numbers.
+  * **Using a different model //family// for the second rating, not a second run of the same model.** Two Sonnet runs would share their failure modes and the agreement figure would flatter the instrument. The cost, stated in C3.8 rather than hidden: a capability difference and a reading difference are not separable afterwards, and the sourcing/reading decomposition suggests most of what was measured is the former. Someone could reasonably have run Sonnet twice and called the result a cleaner measurement of the brief.
+  * **Publishing the uncorrected comparison as the headline.** Rater 2's one disqualified citation (''kameleoon.io'') is demoted in a sensitivity line rather than in the headline table, because rater 1's headline rows are pre-correction too and the symmetric comparison is the honest one. The corrected variant moves the pooled figure from +0.400 to +0.400 and is printed beside it.
+  * **Not re-estimating the accuracy rates from rater 2's verdicts.** It would have been one line of code and it would have been meaningless: 35 rows is not the sample design, the stratum weights do not apply to it, and substituting 35 of 175 rows produces a hybrid that is neither rater's reading. What is published instead is the unweighted ''current'' share per rater, labelled as a direction and not a rate.
+  * **Fixing ''owner_verify_sources.py'' rather than only recording its defect.** The script is published, its hash is on the appendix manifest, and changing it means the appendix regenerates. The alternative — a note saying the check is blind to non-Latin quotes, beside a check that stays blind — would have left the next reader with an instrument the page knows is broken. The consequence for the first sitting's rows is stated in C3.8, and the one row that needed a requote got it on **2026-09-17** ([[#C3.9 The 2026-09-11 re-run, applied (2026-09-17)|C3.9]]).
   * **No ''~~DISCUSSION~~'' on this provenance page.** Following the convention set by the earlier provenance pages: comments belong on the content page.
@@ Line 252: / Line 891: @@
 | "over 1M sites" for Libert's 2018 crawl (my own first draft) | **replaced with the paper's own figures** — Alexa top one million, October 2017, 938,093 pages loaded, 248,029 policy links extracted — after reading the results section |
 | "Kariryaa et al., NDSS 2025" (my own first draft) | **a fabricated author name.** The paper's first author is Xiaoyuan Wu. Caught by a self-review script that matches every ''Name et al., VENUE YEAR'' attribution on the page against the ''author'' field of the key cited next to it; 14 of 15 rows passed, this one did not. The citekey was renamed ''wu2025_appprivacyreport'' at the same time |
+==== Reproducibility check run after publication ====
+All three scripts were re-run and their output diffed byte-for-byte against the committed output: identical. ''owner_dbs.py'' was then re-run against a **fresh cache**, i.e. re-fetching all five inputs and the Public Suffix List live, and produced output identical to the published snapshot — so every figure in sections B and C of this page still reproduces from the live sources as of 2026-08-17, with the same five SHA-256 prefixes. That will stop being true: two of the three lists change weekly.
+==== Publishing gotcha: the bibliography needs a cache purge ====
+After appending to ''literature:bibliography'' and saving this page, **every one of the 14 new ''%%{[key]}%%'' markers rendered as a broken citation** and the reference list showed only the 6 keys that already existed. The bibliography source was well-formed; the ''bibtex'' plugin was serving a cached parse of the old bibliography.
+Fix: ''curl -sL "https://measuretheweb.org/literature/bibliography?purge=true"'' and then the same on the citing page. After that all 22 references rendered. **Do this after any bibliography append and re-read the rendered page**, or you will publish a page full of broken citations and the raw source will look perfect.
 ===== H. Reviewers =====
@@ Line 344: / Line 993: @@
      `pushed_at: 2022-11-26T19:43:12Z`. The reviewer's inability to re-fetch under
      a rate limit is not evidence against it.
+--- Reviewer 4 (generic, Fable): the three findings that mattered most ----------
+Given no checklist and told to find what the focused three were not looking for.
+It found the most serious defect on the page, in the page's own method.
+ACCEPTED (blocking -- this was a real bug in owner_dbs.py):
+. **googleapis.com is a PRIVATE PSL rule, so the fold + exact-key lookup
+     manufactured a coverage hole and depressed every weighted coverage figure.**
+     Verified: the rule is at line 13935 of the snapshot, past BEGIN PRIVATE
+     DOMAINS at 11276, and all three lists DO name googleapis.com (webXray
+     "Google APIs", Tracker Radar "Google LLC", Disconnect "Google"). So
+     fonts.googleapis.com -- which the page showcased three times as "the single
+     most prevalent third-party name with no owner in any list" -- was an artefact
+     of my own lookup rule, and it carries prevalence 0.369 on its own.
+     Fixed properly rather than patched: fetch_psl() now returns the ICANN and
+     PRIVATE sections separately, a lookup() helper walks parent labels, and the
+     script prints all FOUR combinations (ICANN/full fold x walk/exact lookup) so
+     the reader can see that the choice moves the answer by more than the lists
+     differ. Every figure in sections C and D of the page was re-derived:
+     weighted coverage 54.5/79.3/75.8% -> 58.6/84.3/80.5%; domain share
+.4/12.2/5.0% -> 2.1/17.2/7.0%; top-100 70/95/91 -> 71/98/94; the
+     no-owner-anywhere figure 94.6% of domains and 19.8% of weight -> 92.4% and
+.4%. The pairwise disagreement table moved by under a point.
+     The replacement example is better than the one it replaces: the most
+     requested domain no list can name is now tiktokw.us at prevalence 0.041,
+     which is a genuine, recent, mid-tail coverage hole. The googleapis rows are
+     re-captioned as the trap they actually are, in the page's trap list and again
+     in the lookup demo, which now shows the buggy version, says so, and gives the
+     three-line fix.
+ACCEPTED (blocking):
+. **The two rows that failed the sourcing bar were still counted in the
+     adjudication tallies, and two of webXray's three "current" verdicts WERE
+     those two rows.** Verified from the script: webXray's "current" rows were
+     casalemedia.com, 360yield.com and fwmrm.net, and the latter two are exactly
+     the UNRESOLVED pair. So the one figure most favourable to webXray rested on
+     the evidence the script itself called insufficient -- while both pages
+     claimed those rows were "recorded as unresolved rather than guessed".
+     owner_adjudication.py now excludes them from every tally and prints what it
+     excluded and why: 28 rows, webXray current on 1 of 23 (was 3 of 25), Tracker
+     Radar 8 of 28, Disconnect 21 of 21.
+     Also accepted: the "Read as a share of each list's own entries ... 100%"
+     sentence invited exactly the accuracy ranking the box below it disclaimed.
+     Cut. The WRAP box now also says that Tracker Radar's zero in "No entry" and
+     Disconnect's zero in "Stale" are artefacts of how the sample was drawn, and
+     the script prints the same caveat above its own tallies.
+ACCEPTED (blocking):
+. **The provenance page and the committed script output were not brought up to
+     date after reviewer 2's fixes, while certifying that everything passed.** The
+     ROLE hand map in report_webxray.mjs still said Yang & Yue tried webXray
+     "third of four" where the content page had been corrected to "second", and
+     the section D verification table still said Ghostery trackerdb was "(MIT)".
+     Both fixed at the source and re-run, so the "unedited output" now agrees with
+     the page it certifies. The reviewer is right that check_page_numbers.mjs
+     cannot catch this class -- "third" and "MIT" are not numerals -- and that gap
+     is now stated in *What could not be established*.
+ACCEPTED (smaller):
+. "Treat webXray as **unobtainable**" was contradicted by the page's own
+     practice: every figure on it was computed from the surviving copy. Reworded to
+     "obtainable but not redistributable, and not installable", with the artefact-
+     appendix consequence spelled out.
+. Ghostery trackerdb is named as a live option but is absent from the measured
+     comparison. The omission and its reason (per-company .eno files plus a
+     separate patterns layer, not a single domain->owner map) are now stated in the
+     comparison section, labelled an omission rather than a judgement.
+. The reader could not assemble a pipeline. A new *Assembling the pipeline*
+     section gives the six steps, including the one that was entirely missing:
+     resolve the VISITED site to an owner too and drop same-owner requests. Without
+     it, google.com embedding gstatic.com counts as third-party tracking.
+. The actionable answer sat ~120 lines after the lede. Jump links to
+     *Choosing a resolution source now* and *Assembling the pipeline* added to the
+     lede, which now says plainly that everything between is the evidence.
+NOTED, not acted on in this sitting:
+. Discoverability: the wiki's only treatment of domain-to-company resolution
+     lives under a dead tool's name. The reviewer is right that the earlier
+     rejection of a separate topic page was weak. Mitigated by adding a link from
+     [[Privacy:Requests]], which is where a reader asking "whose request is this?"
+     actually lands. A ''design:ownership_resolution'' page that this one feeds is
+     worth considering and is recorded as such rather than done.
+. The 16-row schema table "could halve". Kept: its fill-rate column is the
+     evidence for "most of that ambition is unfilled", which is the section's
+     claim, and a reader checking one field wants the row.
 --- Reviewer 1: figures versus script ------------------------------------------
-Still running when the page was published. Its brief was to re-run all three
+Landed after the first publication and after reviewer 4's fixes were applied, and
+it re-verified over 150 figures and table cells against freshly re-run script
+output, all byte-identical. It found two defects nobody else did, both in the
+script's own descriptions of its method rather than in any downstream figure.
+ACCEPTED (blocking):
+. **"domain_summary.json is keyed by hostname for 16,396 of its 47,836 rows"
+     was computed by a label-count heuristic (>=3 labels), which answers a
+     different question and is wrong under either fold.** 14,033 of those 16,396
+     rows already ARE their own registrable domain, precisely because
+     googleapis.com, s3.amazonaws.com and cloudfront.net are private-section PSL
+     suffixes -- the same root cause as reviewer 4's finding, showing up in a
+     second place. The honest figure is the number of rows the fold you used
+     merges: 15,651 under the ICANN fold, 2,339 under ICANN+PRIVATE. The script now
+     prints both and says explicitly that the label count answers neither
+     question; the page carries the same correction. The reviewer also
+     re-implemented the PSL algorithm from scratch and found zero mismatches
+     against registrable(), so the fold function itself was never wrong -- only its
+     description.
+ACCEPTED (should fix):
+. **The suffix-fold residue conflated two operations.** "The fold ... fires on
+" was measured with the whole norm_name() pipeline, which strips punctuation
+     before the suffix regex runs. Only 16 of the 45 lose a legal-form suffix; 29
+     change on punctuation alone (AT&T, JD.com, "Here, There & Everywhere"). The
+     causal claim overstated the suffix fold's reach by nearly 3x. The script now
+     prints both counts and names the punctuation-only cases; both pages say so.
+ACCEPTED (nit, already fixed):
+. owner_adjudication.py's docstring said "29" where ROWS has 30. Never
+     published; corrected.
+Its process caveat is fair and worth recording: the repository was being edited
+throughout its run, so its findings are pinned to a snapshot. Both defects it
+found were still present in the last revision it sampled and are fixed now. Its brief was to re-run all three
 scripts and diff every figure against the real output. That check was also run
 directly, repeatedly, throughout the session: check_page_numbers.mjs passes
@@ Line 355: / Line 1127: @@
 per-paper figures in the cited sources. Anything reviewer 1 reports after
 publication goes into the page history, not into this log.
+--- Reviewer 5: industry and website claims (late) -----------------------------
+Checked 13 items against primary sources; 10 passed. It found the single most
+important error on the page, in the claim the licence box is built on.
+ACCEPTED (blocking -- the page's central licence claim was wrong):
+. **webXray was relicensed TWICE after the snapshot this page measures, and
+     ended up MIT.** Verified directly: commit 245ec5d7 (2021-06-14) "Update
+     LICENSE.md - Now open-source" makes it GPLv3, and commit 73fe0fc9
+     (2023-02-01), authored by "Tim Libert", replaces that with an MIT LICENSE
+     reading "Copyright (c) 2023 Tim Libert". peterjoles/webXray reports
+     spdx_id: MIT and is 36 commits AHEAD of thezedwards/webXray, preserving
+     upstream history -- including Libert's own post-2021 feature work -- past the
+     deletion of the upstream repo. Several other forks report GPL-3.0.
+     So "webXray is not open source, and redistributing it is prohibited" was true
+     of the one snapshot the page happened to build on and false of the project's
+     final state, and the page's own advice -- "verify the licence of whatever file
+     you actually download" -- is exactly what it failed to do. The licence box is
+     rewritten as a three-row table of the three licences with what each permits,
+     the availability table now names peterjoles/webXray as the most complete copy
+     and says plainly that every figure here comes from the most restrictively
+     licensed copy that exists, and the lede no longer says the tool "is gone".
+. **"three small commits in a personal working copy"** for the newest fork
+     activity was wrong: those are 36 preserved upstream commits, not a fork
+     owner's tinkering. Corrected.
+. **forks_count reports 19 while the forks endpoint returns 20 objects.** The
+     page cited the endpoint for the number 19. Now states both.
+ACCEPTED (should fix):
+. The Mozilla shavar footnote is attached to a row about entities.json, but
+     disconnect-blacklist.json mirrors services.json; the file mirroring
+     entities.json is disconnect-entitylist.json. The quote is verbatim, the
+     attachment was to the sibling file.
+. "whotracks.me now redirects" is imprecise: it returns HTTP 200 with a
+     canonical link to ghostery.com/whotracksme, serving byte-identical content
+     rather than issuing a redirect.
+Independently re-confirmed, already fixed: the Ghostery trackerdb licence (CC
+BY-NC-SA 4.0, and it is whotracks.me that is MIT -- the page had them inverted),
+the RDBinns commit range (11 commits, 2018-03-29 to 2018-04-05), and the
+truncated PSL quotation. Also confirmed against SEC filings rather than press
+releases: the Xandr close (AT&T Form 10-Q, "On June 6, 2022") and the Teads close
+(Outbrain 8-K, "On February 3, 2025"), plus that SEC now lists CIK 0001454938 as
+"Teads Holding Co." with formerNames "Outbrain Inc." -- the acquirer took the
+target's name, as the page says.
+The lesson worth keeping: five reviewers found five defects nobody else found,
+and the two most serious -- a coverage figure depressed by my own lookup rule, and
+a licence claim resting on an unrepresentative snapshot -- were both invisible to
+every automated guard in this repository.
+==== Reviewers, second sitting (2026-09-05) ====
+Four reviewers, all told explicitly that the briefing they were given might not
+be exhaustive and to verify from the files and the live web rather than from it.
+Three focused Sonnet passes ran in parallel against the three page drafts, the
+scripts and their outputs, and these notes; the generic Fable pass ran after
+their findings were applied. Findings are recorded with the decision, because a
+rejection is the only measure of whether a reviewer slot is worth having.
+=== Reviewer: citations and quotes (Sonnet) ===
+Checked ~25 quotes and all 22 citekeys, ran ''bib_dedup_scan.py'' and
+''check_attributions.mjs'', re-fetched all six CDX queries and the ''id_''
+captures behind every archived quote, spot-checked all 15 named adjudication rows
+against live sources, and diffed every published block against its file.
+^ Finding ^ Decision ^
+| **The 2010 bound is wrong and cites the wrong capture.** ''web/20101008070114id_'' is a DreamHost **"Coming Soon" parking page**, not the Web Xray mashup tool; that content first appears at ''20101108204354'' | **ACCEPTED, blocking.** Verified by fetching both captures. The bound is now 2010-11-08 to at least 2011-09-27, with the parking page named on both pages and the general lesson — a CDX row is a capture, not a page — written into the method notes as the second trap |
+| **The "at least 2026-02-05" bound is true but uncited**: queries 5 and 6 of ''wayback_webxray.sh'' stop at 2024-12-31, so no line of the published output contains a 2025 or 2026 capture | **ACCEPTED, blocking.** This is the sharpest finding of the run: the figure was right, independently re-verified by the reviewer, and **not supported by the evidence the page publishes**. Fixed at the source rather than in the prose: ''wayback_webxray.sh'' gained **query 7** (uncollapsed, 2025-01-01 to 2026-12-31) and was re-run. Its output now carries ''20260205035830 301 OM6ALWQTEGJ5NTY5RYSC2EXCKJOAHOVU'' |
+| ''gnu.org/copyleft/gpl.html'' → ''gpl-3.0.html'' is a **302**, not a 301 | **ACCEPTED.** Verified: the bare-host hop is a 301, the path hop is a 302. Both pages now say 302 |
+| The content page's 2010 quote **merges a ''%%<title>%%'' and a separate "beta 0.5" badge** into one quoted string, while the same page discloses exactly that kind of join for the GPL quote — an inconsistent standard for one defect class | **ACCEPTED.** The footnote now names the two elements. The reviewer is right that the inconsistency was the real problem: the page had already decided this needed disclosing |
+| The published ''owner_random_sample.py'' output block is one trailing blank line short of the file whose SHA-256 the manifest quotes | **ACCEPTED as a wording fix, not a behaviour change.** The generator strips trailing newlines because DokuWiki discards them inside a block. The appendix now states that this is the only transformation, and that the hash and byte count are of the file on disk, so a reader reconstructing a file must add the newline back before hashing |
+Confirmed clean by this reviewer, and recorded because these were the things most
+likely to be wrong: all 22 citekeys resolve to distinct entries with no DOI or
+title collisions and **no citekey was added this sitting**; 15/15 table
+attributions pass; every other Wayback bound reconfirmed against fresh queries;
+all 15 named adjudication rows verbatim against live sources, with
+''cnevids.com'' and ''cdnbasket.net'' correctly carrying no quote because they
+are unresolved; **no banned-tier source used as a citation** anywhere in the 135
+sourced rows; all 19 manifest hashes and byte counts correct. It also
+independently re-derived 175 rows / 135 resolved / 40 unresolved / 17 ''tls://''
+sources / 29 changed verdicts, and confirmed that the second Tracker Radar
+''error'' verdict in the raw data (''loopme.me'') is correctly excluded from
+Tracker Radar's rate because that domain was drawn for **webXray's** sample.
+=== Reviewer: figures versus script (Sonnet) ===
+Re-ran the whole pipeline, diffed every output against its committed artefact,
+and cross-checked every figure on all three pages. It also noted, correctly, that
+the pages were being edited while it read them, and said so rather than reporting
+a moving target as a fixed one.
+^ Finding ^ Decision ^
+| **"the same universe ''owner_dbs.py'' uses, so the two sittings' figures are on one denominator" is false.** The first sitting's frame is **32,369** domains, the second's **32,337** | **ACCEPTED, blocking, and the most valuable finding of the run.** Reproduced by running ''owner_sample.py %%--cache%% out/webxray/cache'' (32,369; 669 / 5,581 / 2,268) against ''%%--cache%% out/owner_cache'' (32,337; 664 / 5,566 / 2,264). Both are real. The coverage **shares** round to the same 2.1% / 17.2% / 7.0%, which is precisely why two different frames could sit on one page unnoticed. [[#C3.1 The frame, and what it excludes|C3.1]] now carries three rows on it, and the content page carries a ''%%<WRAP tip>%%'' next to the figures themselves |
+| **The run table's hash list is false.** It claimed all six input files were unchanged between the sittings; ''tr_domain_summary'' went ''7a303228d812a41c'' → ''19f7a5a6a839ec87'' (16,310,837 → 16,347,575 bytes) and the Public Suffix List went ''155b43d46932e933'' → ''aef8fb81d63232da'' | **ACCEPTED, blocking.** Verified against ''out/webxray/owner_dbs-output.txt''. This was written from the second sitting's hashes without checking them against the first sitting's — an assumption stated as a measurement, which is the exact failure the page spends a section warning about. The row now lists the four that are unchanged and the two that moved, and says so |
+| **The appendix's manifest hash for ''out/probe_all.txt'' matches no file.** Published ''7d2ca1e722836803''; the file hashes ''3def89e0d3fc8a14'' | **ACCEPTED, blocking.** Root cause found on verification: ''probe_all.txt'' contains **one lone CR**, inside a copyright line scraped from a live site, and Python's text mode silently translates it to LF — so the builder hashed a string that was not the file. Fixed at the root: the builder now reads bytes, hashes the file on disk, discloses newline normalisation **per file** in a new manifest column, and **self-checks** every row and block before writing. The self-check was mutation-tested by reintroducing the original bug, which it rejects. All 23 manifest rows now verify independently |
+| ''owner_merge_rows.py'' **does not fail on what its docstring promises**: duplicates, out-of-vocabulary verdicts and resolved-rows-without-a-source were printed but returned exit 0 | **ACCEPTED.** Reproduced with a synthetic batch. Fixed, and mutation-tested with all three defects: each now exits 1. The reviewer is right that it changed no published number — the committed merge has zero problems — and right that the docstring invited a caller to trust the exit code |
+| **The retained fetch cache does not reproduce the recorded statuses**: ''legal.quantcast.com'' is 526,545 bytes in ''out/verify_cache'' against 3,054,755 in ''verify.json'' | **ACCEPTED.** Traced: the retained cache was the **first, partial pass** (''verify_partial.json'' has that row as NOTFOUND at 526,545 bytes); the run that produced ''verify.json'' used a different cache. Nothing published depended on it, but a misleading artefact is still a defect. Rather than explain it, the **whole check was redone from an empty cache** — see C3.5. Result identical, zero rows changed, zero uncited |
+| Live-network scripts are not byte-reproducible by design | **Noted, not a defect.** ''wayback_webxray.sh'' and the two verifiers hit third-party servers. The CDX queries were independently reconfirmed by two reviewers |
+Confirmed clean by this reviewer: every figure from the new pipeline reproduces
+byte-for-byte and matches the pages exactly (rates, intervals, n=49/47/42, the
+concentration shares, the self-name census, 10 corrections, 29 changed verdicts,
+the 132/3/40/0 citation tallies); every named example domain is a member of the
+sample it is claimed to be in; all 13 published script blocks are byte-identical
+to their files; the estimator's empty-stratum, zero-cell and bad-verdict paths
+behave as documented; ''report_webxray.mjs'' and ''owner_dbs.py'' still reproduce
+the first sitting's figures digit-for-digit; and ''check_owner_sample_figures.py''
+plus its mutation test both do what they claim.
+=== Reviewer: external currency (Sonnet) ===
+Fetched 189 URLs appearing literally on the three pages plus ~70 further live
+endpoints — roughly 260 requests. **No blocking findings.**
+^ Finding ^ Decision ^
+| ''corporate.comcast.com/press/releases/freewheel-acquires-beeswax'', cited by the ''bidr.io'' row of the 30-row adjudication, is a **404** | **ACCEPTED.** Confirmed. Replaced in ''owner_adjudication.py'' with the acquirer's own newsroom, ''freewheel.com/news/freewheel-to-acquire-ad-tech-leader-beeswax'' (HTTP 200, ''datePublished'' 2020-12-17, names Beeswax and "FreeWheel, A Comcast Company"), and the script re-run so section K's published output changes with it |
+| ''gnu.org/copyleft/gpl.html'' → ''gpl-3.0.html'' is a **302**, not a 301 | **ACCEPTED**, independently and identically to the citations reviewer. Fixed on both pages |
+| Four repository states in the "Choosing a resolution source now" table have moved since 2026-08-17 | **ACCEPTED and updated**, after re-fetching each myself: Tracker Radar 2026-08-12 → **2026-09-02** with newest release ''2026.08.28'' (checked against both ''/releases'' and ''/releases/latest'', which agree — the newest tag is not a prerelease); Disconnect 2026-08-07 → **2026-08-28**; Ghostery ''trackerdb'' 2026-08-06 → **2026-09-01**; WhoTracks.me 2026-08-04 "July update" → **2026-09-02 "August update"**. Not false before, but the page is being republished today and a stale date on a republished page is a new claim |
+| ''carlsaturnino/webxray-docker'' is recorded at "399 pulls"; it is 405 today | **ACCEPTED as a wording problem, not a figure to chase.** A monotonically increasing counter should not be published as a fixed number; the note now gives it as a counter with its check date. The conclusion (an image registered 2016-04-26, ''last_modified'' a 2024 metadata bump, not maintained) is unchanged |
+| The PSL's ICANN/PRIVATE rule counts have drifted (6,941/3,290 → 6,949/3,372) | **NOT a page defect, and it became evidence instead.** The page already scopes those counts to "the snapshot used here". But the reviewer's numbers are what let the figures reviewer's finding be confirmed: the PSL is one of the two files that moved between sittings |
+| ''mediaocean.com'' serves an incomplete TLS chain; nine investor-relations and newswire hosts 403 a bare ''curl'' but serve a browser | **Recorded, no change.** These are instrument artefacts, not link rot, and they are why the citation verification has a browser pass. Recorded so a future reviewer does not report them as dead links |
+=== Reviewer: generic, no checklist (Fable) ===
+Ran after the three focused passes and their fixes. It found the two most
+serious defects of the whole run, both of which the three focused passes had
+each passed over for a defensible reason: one is a claim no script computes, and
+the other is data inside a script, so "the page matches its script" was true and
+useless.
+^ Finding ^ Decision ^
+| **The "reversal" is confounded.** The page attributed the whole gap between "webXray current on 1 of 24" and "69.9%" to //selection//. But seven domains were adjudicated in both sittings against a **bit-identical** file, and of the five settled in both, webXray's verdict differs in **four**, all towards ''current'' — because 2026-08-17 scored webXray's ownership-tree **root** and named the acquiring group as owner, where 2026-09-05 scored the **entry** and named the operating subsidiary | **ACCEPTED, blocking, and it is the finding of the run.** Reproduced: ''lijit.com'' is "Sovrn" at the entry and "Federated Media" at the root; ''simpli.fi'' is "simpli.fi" and "GTCR"; ''bidr.io'' and ''crwdcntrl.net'' differ on parent-versus-operator, not tree level; only ''smartadserver.com'' agrees. The lede, the section intro, the ''%%<WRAP important>%%'' box and the "60 percentage points" sentence were all rewritten. The page now says the two numbers are **not comparable**, gives both causes with the per-domain evidence, and publishes the uncomfortable number the reviewer surfaced: **webXray verdicts agreed on 1 of the 5 domains adjudicated twice**, which is the only inter-rater evidence this page has |
+| **The 30-row table's "No entry" column is wrong.** Disconnect has an entry for all eight rows scored absent, five under ''resources'' rather than ''properties''; webXray has "PulsePoint" for ''contextweb.com'' | **ACCEPTED, blocking.** ''owner_adjudication_absent_audit.py'' was written to re-derive every ''absent'' from the files with the project's own parent-label lookup: **9 of 90 verdicts contradicted the files**. All nine corrected, each with its reason in the script. Tallies move: webXray 1 of **24** (was 1 of 23), Disconnect **26 of 28** (was 21 of 21), Tracker Radar 8 of 28 unchanged. ''owner_adjudication.py'' now refuses to print until the audit passes — which is what the 2026-09-05 pipeline always did, and the reason this defect is a first-sitting one |
+| **The pages never say the adjudicators were language models**, and "adjudicated by hand" implied otherwise | **ACCEPTED, blocking.** The content page now says "by Sonnet sub-agents working to a published brief, single-rated". The provenance page said it already; the content page is where a reader deciding whether to cite 69.9% will look |
+| "all three lists are flattered by the same amount" by the unresolved exclusion | **ACCEPTED.** The shares are 18% / 22% / **30%**, so **Disconnect** — the list the page recommends — is the most flattered. Worst case 58.3 / 58.3 / 65.0%, which narrows the margin without reversing it. Both now on the page |
+| The non-overlapping-intervals argument is method-fragile; a percentile bootstrap on 42 rows undercovers at the top | **ACCEPTED.** Replaced with Fisher's exact: **p = 0.014** for Disconnect against webXray, **p = 0.82** for webXray against Tracker Radar. Both recomputed here before publishing |
+| "its own coverage" is a fifth of webXray's file: 664 of 3,215 domains, the rest never met by a 2026 crawl, and survivors are where "current" concentrates | **ACCEPTED.** Stated on the page. The reviewer's DNS probe (22 of 40 out-of-frame domains do not resolve, against 6 of 20 in-frame) is quoted as its finding, not re-run |
+| The lede gave 93.9% bare while the section says to read it as an ordering | **ACCEPTED.** The lede now carries the ''gstatic.com'' concentration inline |
+| Assorted stale cross-references: a pointer to the wrong section, "a single snapshot per list" when there are two, a table header dated 2026-08-17 above rows dated 2026-09-02 | **ALL ACCEPTED and fixed** |
+| "cited, confirmed in the fetched bytes: 132" hides that 8 are fragment matches and 5 are 60-character window matches | **ACCEPTED.** The table row now breaks it down |
+| "the sampled verdict was measuring the adjudicator's reading of a name" misplaces blame — 25% matches the census's own 27.5% label-only rate | **ACCEPTED, and it is a fair correction.** They applied the brief as written; the brief was wrong |
+| "the part that cannot be automated" — it was done by 21 LLM agents | **ACCEPTED.** Reworded |
+| "no figure lives only here, and no figure lives only there" is not true of the appendix | **ACCEPTED.** Now: no figure is //derived// only there, several are only //shown// there, and they are named |
+| The census's "pairs" are distinct domains (3,224 pairs, 3,215 distinct, three with a trailing slash) | **ACCEPTED.** The script's column label and both pages corrected |
+| "Prevalence" is never defined for a reader who does not know Tracker Radar | **ACCEPTED** |
+| Cross-list rates are confounded by naming level — a brand name survives an acquisition, a legal-entity name does not | **ACCEPTED as an opinion worth publishing.** It explains why Tracker Radar's encounter-weighted rate is lowest, and the page now says the three rates are not strictly on one axis |
+| All 16 re-scores in ''owner_verdict_fixes.py'' moved towards the lists | **ACCEPTED in substance, rejected as stated.** Recounted: there are **28** verdict changes, not 16 — 14 to ''current'', 2 to ''granularity'', 1 to ''stale'', 9 to ''unknown'' (which removes the row from every rate), plus both root-scoring fixes making the verdict less severe. The direction is as the reviewer says and the page now states it, with the real counts |
+=== Reviewer: re-check of the applied fixes (Sonnet) ===
+Ran against the fixes from the first three passes, and verified 8 of 10 outright.
+^ Finding ^ Decision ^
+| Tracker Radar's "last commit 2026-09-02" is **wrong**: ''main'' is at 2026-08-28, the commit the ''2026.08.28'' tag points at. The 2026-09-02 commit is on an unmerged automation branch that the repository's ''pushed_at'' field picks up | **ACCEPTED.** I could not re-fetch the API to confirm directly (GitHub rate-limited this host by then), but the release tag I fetched earlier — ''2026.08.28'', published 2026-08-28T15:40:22Z — corroborates it, and a 2026-09-02 data commit on ''main'' would have produced a newer tag. The page now says "last commit on ''main'' 2026-08-28" and records the ''pushed_at'' trap, which is worth more than either date |
+| The appendix builder's self-check never included the "block differs from the file" column, so a wrong normalisation flag could not be caught | **ACCEPTED.** Reproduced by hardcoding the flag; the check passed. Now included, and the mutation is rejected |
+| ''bidr.io'' carries "announced 2021-05-20" while the source it cites is dated 2020-12-17 | **ACCEPTED.** Corrected to 2020-12-17. Noted that the correct date was already written in the reviewer log while the code kept the wrong one |
+| The two pages give different appendix sizes and neither addition is consistent | **ACCEPTED.** The appendix now states only its own measured size; the parent, which is built second and can measure both, does the arithmetic |
+| A stray ''owner_sample.json'' was left in the repository root by a reproduction run | **ACCEPTED**, removed |
+| ''check_owner_sample_figures.py'' can be defeated by swapping which list a row is attributed to, since it tests presence and not position | **ACCEPTED as a disclosed limitation, not a bug.** It is exactly what the script's docstring and this page already say it cannot see. Recorded rather than fixed, because the fix is a positional parser that would break on any table edit |
+| The brief said 23 manifest rows; there are 24 | **Correct** — my count, not the page's |
+=== Reviewers, third sitting (2026-09-11) ===
+Three focused Sonnet passes ran in parallel against the page drafts, the scripts
+and their outputs; the generic pass was **skipped for budget**, which is recorded as a gap rather than dressed up as a full review layer. All three were told the briefing might not be exhaustive and to
+verify from the files and the live web rather than from it. **The sitting's worst
+error was found here, not by any guard**, and both reviewers who could have found
+it did.
+^ Finding ^ Decision ^
+| **The ''kameleoon.io'' demotion is wrong: the quoted ''%%<script src>%%'' line //is// verbatim in the bytes the checker itself cached.** Raised independently by the external-currency reviewer (which fetched the live page twice with cache-busting headers) and by the evidence reviewer (which opened ''out/irr/verify_cache/eb6a82016d8e837fb6aee2c8.bin'', the exact 94,416 bytes the checker read, and re-ran ''norm()'' on it to show the tag strip deletes the quote before the comparison) | **ACCEPTED, blocking, and the most valuable finding of the run.** The demotion is reversed, ''owner_verify_sources.py'' gains an ''OK-MARKUP'' status, and the failure is written up rather than smoothed away. The evidence reviewer's sharper point is also taken: the browser pass's NOTFOUND was **structurally guaranteed** for any markup quote and therefore carried no information at all, so citing it as corroboration was wrong twice over. The corrected-variant kappa figures now equal the uncorrected ones, and ''owner_irr_corrections.py'' asserts that rather than the page claiming it |
+| **The ''compass-fit.jp'' explanation is misleading.** The page implied the ''致しました。'' ending appears nowhere on the cited page; the reviewer found the full sentence in the body, differing from the row's quote by a single full-width comma after ''台湾国内での'' | **ACCEPTED.** Verified directly: the body reads ''台湾国内での、ネイティブ型広告配信サービス…''. The bullet now says one character, names it, and carries a note that the first draft got the reason wrong. The finding is milder than first published — a transcription slip, not a paraphrase — and the filed follow-up item was corrected too |
+| **The rater-2 note for ''cnn.com'' overstates the Warner Bros. Discovery position**: WBD's own Q2-2026 10-Q calls the Discovery Global separation "previously proposed", not pending, and conditions the PSKY merger on it //not// completing | **ACCEPTED as a fact, no page change.** The reviewer is right about the filing. That note is a field of one adjudication row and reaches no published block — ''cnn.com'' produced no inter-rater disagreement, so it is not printed in the estimator's section 7 either. Recorded here so the next reader of ''adj2_rows.json'' knows |
+| Every other external fact re-fetched and confirmed: the ''samsungads.ca'' certificate ''notAfter Mar 12 07:06:09 2026 GMT'' and a stable 59,465-byte body over four fetches; ''whois.nic.it'' answering on port 43 with the ''CED DIGITAL & SERVIZI Srl'' line at its exact five-space indent; ''microad.co.jp'' live; the PSKY merger still pending, paused by a federal judge and not closing before 2027-06-01. Nine rater-2 verdicts spot-checked against primary sources — ''cedscdn.it'', ''tqlkg.com'', ''awltovhc.com'', ''sharethis.com'', ''owneriq.net'', ''mountain.com'', ''pages02.net'', ''eyeota.net'', ''mediaset.es'' — all hold | **No change needed.** ''sa-as.com'' remains the weakest row, resting only on a MarkMonitor RDAP registrant field; the reviewer could not corroborate it further either, and rater 2 had flagged it itself |
+| Independence and stimulus-identity confirmed by hand: ''prompt00.txt'' is ''INSTRUCTIONS.md'' + ''batch00.txt'' + one output-path line, with no rater-1 verdict, no estimator, and no mention of the ''self-named'' retirement anywhere in it; per-domain blocks byte-identical between ''out/adj/batch*.txt'' and ''out/irr/adj2/batch*.txt'' for four spot-checked domains drawn from four different original batches | **No change needed.** This is the claim the whole section rests on and it is now checked rather than asserted |
+| All kappa figures, intervals, agreement percentages and the 63.0% sourcing/reading split reproduce exactly on an independent re-run; the ''self-named'' counts (6, 4, 3 shared) recomputed from the JSON; no citekeys added; the published ''%%<file>%%'' block matches the script byte-for-byte | **No change needed** |
+| **Three findings from the figures-versus-script pass, none of them a wrong number on the page today.** (a) ''owner_irr_kappa.py'' documented NAMED as "neither rater is forced to ''absent''" while the code selected with ''or''; the two are equivalent only because both raters' ''absent'' cells happen to coincide, which is empirical luck rather than an invariant. (b) The JSON artefact was not byte-stable: ''pe'' summed over a ''set()'', whose iteration order is hash-seed randomised, so the 15th–16th decimal moved between runs. (c) The page pointed at appendix section **AE** for the subsample domain list; AE is a mutation harness and the list is in **AF** | **ALL THREE ACCEPTED.** (a) ''assert_absent_agrees()'' now raises before any population is built, and prints that ''absent'' agrees on all 105 cells; the docstring now says what the code does and why the equivalence is checked rather than assumed. (b) ''sorted(set(...))'', verified byte-stable over three runs under a randomised hash seed. (c) Fixed to AF. The reviewer also noted that the **pooled** bootstrap rows resample cells rather than domain blocks — which the page already discloses, and the estimator's docstring now does too |
+| **The guard was blind to every figure that lives in a script's //text// output rather than in the kappa JSON.** The reviewer mutated the ''self-named'' counts, the sourcing/reading split table, the resolution counts, the unweighted current-share table and the whole citation tally, and ''check_irr_figures.py'' passed on every one | **ACCEPTED, and the most useful finding after the kameleoon one.** The guard now parses those figures out of ''owner_irr_kappa-output.txt'' and ''owner_verify_sources_irr-output.txt'' (a new ''%%--verify-output%%''), and six mutations were added for them. Two of the six then **survived**: the self-named count and the split table were being checked with bare-digit needles like "17", which match anywhere on a long page, and the self-named needle was additionally satisfied by the //other// page's untouched copy. Both fixed — the split table is matched as a whole row, and the self-named count is checked wherever its sentence shape occurs on **every** page. 18 mutations, all caught |
+| **Process finding, and the one this sitting should be judged on: the reviewer was mid-review when the kameleoon fix landed**, so its report straddles two states of the artefact. It said so unprompted | **ACCEPTED as a real methodological failure of this sitting.** The rule is "do not edit a page under its reviewers" and it was broken — the fix was applied the moment two other reviewers reported, without waiting for the third. The reviewer re-checked everything against the post-fix state and its findings above are against that state, but the next sitting should hold fixes until every reviewer has reported, or re-run the ones whose ground moved |
+**What the guards did not catch, and why it matters.** ''check_irr_figures.py''
+compares the page to the estimator and passed throughout; the estimator was never
+wrong. The defect was one layer below, in a //different// script's verdict about
+evidence, and no guard here looks at that. Its own mutation harness also reported
+a survivor after the fix — a mutation aimed at a figure that had stopped being on
+the page — which is the harness working as intended: a mutation that changes
+nothing is reported as a survivor rather than silently passing.
 ===== I. Unedited output: scripts/report_webxray.mjs =====
@@ Line 366: / Line 1332: @@
 Schema signal: tools[].name matching /webx[\s-]?ray/i with usedOrMentioned in {used, produced} fires on 7 papers.
 Not used/produced (compared/mentioned/unclear): 0 papers.
-Full-text sweep of paper.cols.txt for the same pattern: 15 papers, 0.3% of the 5859-paper corpus and 1.3% of the 1120 that ran a crawl.
+Full-text sweep of paper.cols.txt for the same pattern: 15 papers, 0.3% of the 5859-paper corpus. 13 of them are among the 1120 that ran a crawl (1.2% of that population).
 Schema papers not in the sweep: none (a sweep miss would mean the extractor read a rendering the sweep cannot see).
 Sweep papers the schema misses: 8.
@@ Line 390: / Line 1356: @@
 Per paper:
 Paper                                                                                                                        Role             What was used
----------------------------------------------------------------------------------------------------------------------------  ---------------  ----------------------------------------------------------------------------------------------------------------------------
+---------------------------------------------------------------------------------------------------------------------------  ---------------  --------------------------------------------------------------------------------------------------------------------------------
 CCS/2016/online-tracking-a-1-million-site-measurement-and-analysis                                                           compared         crawler, as a baseline whose browser is criticised
 WWW/2018/an-automated-approach-to-auditing-disclosure-of-third-party-data-collection-in-w                                    instrument       crawler + owner list + policy extraction
@@ Line 396: / Line 1362: @@
 IEEE-SP/2020/do-cookie-banners-respect-my-choice-measuring-legal-compliance-of-banners-from-i                                owner-list-only  owner list
 IMC/2020/analyzing-third-party-service-dependencies-in-modern-web-services-have-we-learne                                    citation         owner list in the reference list only; the paper builds its own TLD + SAN + SOA heuristic
-PETS/2020/a-comparative-measurement-study-of-web-tracking-on-mobile-and-desktop-environmen                                   owner-list-only  owner list (cited as “Tim Libert's library”), third of four fallback sources
+PETS/2020/a-comparative-measurement-study-of-web-tracking-on-mobile-and-desktop-environmen                                   owner-list-only  owner list (cited as “Tim Libert's library”), SECOND of four sources tried in order: CrunchBase, webXray, TLS certificate, WHOIS
 NDSS/2021/whos-hosting-the-block-party-studying-third-party-blockage-of-csp-and-sri                                          owner-list-only  owner list, retrieved from the Internet Archive
 IEEE-SP/2022/journey-to-the-center-of-the-cookie-ecosystem-unraveling-actors-roles-and-relati                                owner-list-only  owner list, third by priority of three merged lists
@@ Line 464: / Line 1430: @@
 Papers matching ANY of webXray / Tracker Radar / WhoTracks.me / Crunchbase / Disconnect-as-a-list:
-papers (2.3% of the corpus; 12.1% of the 1120 that crawled).
+papers (2.3% of the corpus). 107 of them are among the 1120 that crawled -- 9.6% of that population.
 Year   Corpus papers  Naming an ownership resource  Share
 -----  -------------  ----------------------------  -----
@@ Line 553: / Line 1519: @@
 of 26 literal figures/quotes located in the cited paper's .cols rendering.
-=== H. Evidence quotes behind the schema tuples ===
+=== H. Evidence quotes behind the schema tuples (with PDF fallback) ===
 partial          69%  WWW/2018/an-automated-approach-to-auditing-disclosure-of-third-party-data-collection-in-w | webxray | methodology | "the webxray software platform is used to monitor third-party network traffic generated by loading a given web page and attribute such traffic to the entities which receive the data."
@@ Line 571: / Line 1537: @@
 partial          67%  IEEE-SP/2022/journey-to-the-center-of-the-cookie-ecosystem-unraveling-actors-roles-and-relati | classification:webxray | methodology | "The first is based on three manually-curated lists (Disconnect [30], WhoTracks.me [31] and webxray [32])"
 exact           100%  IEEE-SP/2022/journey-to-the-center-of-the-cookie-ecosystem-unraveling-actors-roles-and-relati | classification:graph-based label propagation algorithm (custom) | methodology | "We validated our approach by using the manually compiled lists as a “ground truth” reference: we then looked for conflicts"
-Summary: 16 quotes: 4 exact, 11 partial (>=60% of 4-word windows), 1 below threshold.
+Summary: 16 quotes: 4 exact, 11 partial (>=60% of 4-word windows), 0 rescued from the PDF, 1 below threshold in both renderings.
 Below-threshold is not "unsupported": every one checked by hand was present, mangled by a column splice or a dropped citation marker.
@@ Line 581: / Line 1547: @@
   owner_dbs.py, snapshots of 2026-08-17: webXray 827 owners / 3,215 domains; Tracker Radar 19,148 / 38,368; Disconnect 1,887 / 7,850.
   owner_dbs.py: webXray owners with a parent_id 319 of 827 (38.6%); tree depth up to 6; largest owner groupm with 620 domains.
-  owner_dbs.py: coverage of 45,525 registrable third-party domains -- webXray 657 (1.4%), Tracker Radar 5,539 (12.2%), Disconnect 2,277 (5.0%).
+  owner_dbs.py: coverage of 32,369 registrable third-party domains -- webXray 669 (2.1%), Tracker Radar 5,581 (17.2%), Disconnect 2,268 (7.0%).
-  owner_dbs.py: prevalence-weighted coverage -- webXray 54.5%, Tracker Radar 79.3%, Disconnect 75.8%.
+  owner_dbs.py: prevalence-weighted coverage -- webXray 58.6%, Tracker Radar 84.3%, Disconnect 80.5%.
-  owner_dbs.py: top 100 by prevalence -- webXray 70%, Tracker Radar 95%, Disconnect 91%; top 1,000 -- 26 / 67 / 67%; top 10,000 -- 5 / 27 / 16%.
+  owner_dbs.py: top 100 by prevalence -- webXray 71%, Tracker Radar 98%, Disconnect 94%; top 1,000 -- 26 / 69 / 68%; top 10,000 -- 5 / 29 / 17%.
-  owner_dbs.py: pairwise "needs a human" disagreement -- webXray vs TR 197 of 601 (32.8%), webXray vs Disconnect 216 of 461 (46.9%), TR vs Disconnect 584 of 1,532 (38.1%).
+  owner_dbs.py: pairwise "needs a human" disagreement -- webXray vs TR 198 of 612 (32.4%), webXray vs Disconnect 217 of 464 (46.8%), TR vs Disconnect 584 of 1,538 (38.0%).
-  owner_dbs.py: 16,396 of 47,836 domain_summary.json rows are hostnames not registrable domains; 19 rows are not hostnames at all, one of them the literal string "null".
+  owner_dbs.py: the ICANN-section fold merges 15,651 of domain_summary.json's 47,836 rows into 32,369 registrable domains; 19 rows are not hostnames at all, one of them the literal string "null". (The label-count heuristic that gives 16,396 answers neither question and is not used.)
   webXray licence: PolyForm Strict License 1.0.0 (LICENSE.md in the surviving copy).
   Tracker Radar and Disconnect licences: CC BY-NC-SA 4.0 (LICENSE and entities.json "license" field).
@@ Line 675: / Line 1641: @@
             f.write(r.read())
     raw = open(path, "rb").read()
-    rules, exceptions = set(), set()
+    # The PSL has TWO sections and the split matters more than any other choice
+    # in this script.  ICANN rules are real TLDs and registry suffixes; PRIVATE
+    # rules are suffixes companies asked to have treated as boundaries --
+    # `googleapis.com` is one of them.  Folding with the private section leaves
+    # `fonts.googleapis.com` standing as its own "registrable domain", and an
+    # exact-key lookup then fails even though all three lists name
+    # `googleapis.com` -> Google.  So both sections are loaded separately.
+    icann, private, exceptions = set(), set(), set()
+    section = "icann"
     for line in raw.decode("utf8").splitlines():
-        line = line.strip()
+        stripped = line.strip()
-        if not line or line.startswith("//"):
+        if "BEGIN PRIVATE DOMAINS" in stripped:
+            section = "private"
             continue
-        if line.startswith("!"):
+        if "BEGIN ICANN DOMAINS" in stripped:
-            exceptions.add(line[1:])
+            section = "icann"
+            continue
+        if not stripped or stripped.startswith("//"):
+            continue
+        if stripped.startswith("!"):
+            exceptions.add(stripped[1:])
         else:
-            rules.add(line)
+            (icann if section == "icann" else private).add(stripped)
-    return rules, exceptions, hashlib.sha256(raw).hexdigest()[:16]
+    return icann, private, exceptions, hashlib.sha256(raw).hexdigest()[:16]
@@ Line 712: / Line 1692: @@
         return None                               # the host *is* a public suffix
     return ".".join(labels[-(best + 1):])
+def lookup(host, mapping):
+    """Owner for a hostname, trying the host then each parent label sequence.
+    A list keyed on `googleapis.com` must answer for `fonts.googleapis.com`.
+    Exact-key lookup is the single easiest way to manufacture a coverage hole,
+    and it is what the first version of this script did.  Stops at two labels so
+    it can never walk up to a bare TLD.
+    """
+    labels = host.split(".")
+    for i in range(len(labels) - 1):
+        candidate = ".".join(labels[i:])
+        if candidate in mapping:
+            return mapping[candidate], candidate
+    return None, None
@@ Line 875: / Line 1871: @@
     print()
     summary = data["tr_domain_summary"]
-    rules, exceptions, psl_hash = fetch_psl(args.cache)
+    icann, private, exceptions, psl_hash = fetch_psl(args.cache)
-    print(f"Public Suffix List: {psl_hash}  {len(rules):,} rules, "
+    print(f"Public Suffix List: {psl_hash}  {len(icann):,} ICANN rules, "
-          f"{len(exceptions):,} exceptions  {PSL_URL}")
+          f"{len(private):,} PRIVATE rules, {len(exceptions):,} exceptions  {PSL_URL}")
+    print("The ICANN/PRIVATE split is the most consequential choice in this script.")
+    print("`googleapis.com` is a PRIVATE rule, so folding with the private section")
+    print("leaves fonts.googleapis.com standing as its own 'registrable domain' --")
+    print("and an exact-key lookup then finds no owner for it, even though all three")
+    print("lists name googleapis.com -> Google. Both rules are therefore reported.")
+    print()
     raw_keys = list(summary)
     bad = [k for k in raw_keys if "." not in k]
-    prev = {}
-    folded = 0
+    def build(rules_set):
-    for host, v in summary.items():
+        prev, folded = {}, 0
-        host = host.lower()
+        for host, v in summary.items():
-        if "." not in host:
+            host = host.lower()
-            continue                              # see the 'null' note below
+            if "." not in host:
-        reg = registrable(host, rules, exceptions)
+                continue                          # see the 'null' note below
-        if reg is None:
+            reg = registrable(host, rules_set, exceptions)
-            continue
+            if reg is None:
-        if reg != host:
+                continue
-            folded += 1
+            if reg != host:
-        # A registrable domain's weight is the *largest* prevalence among its
+                folded += 1
-        # hostnames, not the sum: one site can request fonts.googleapis.com and
+            # A registrable domain's weight is the *largest* prevalence among its
-        # ajax.googleapis.com, so summing would double-count sites.  This makes
+            # hostnames, not the sum: one site can request fonts.googleapis.com
-        # every weighted coverage figure below a conservative lower bound.
+            # and ajax.googleapis.com, so summing would double-count sites. This
-        prev[reg] = max(prev.get(reg, 0.0), v["prevalence"])
+            # makes every weighted coverage figure a conservative lower bound.
-    total_prev = sum(prev.values())
+            prev[reg] = max(prev.get(reg, 0.0), v["prevalence"])
-    universe = sorted(prev)
+        return prev, folded
-    print(f"domain_summary.json rows: {len(raw_keys):,}. "
-          f"{sum(1 for k in raw_keys if k.count('.') >= 2):,} of them are keyed by "
+    prev_icann, folded_icann = build(icann)
-          f"hostname rather than registrable domain; folding to eTLD+1 merges "
+    prev_full, folded_full = build(icann | private)
-          f"{folded:,} rows and leaves {len(universe):,} registrable domains.")
+    # How many rows are "keyed by hostname rather than registrable domain" is not
+    # a property of the data alone -- it depends on which PSL section you fold
+    # with, because googleapis.com and s3.amazonaws.com are private-section
+    # suffixes. A label-count heuristic (>=3 labels) says 16,396 and is wrong
+    # under either fold; the honest figure is the number of rows the fold you
+    # actually used merges. Both are printed. Caught by review 2026-08-17.
+    print(f"domain_summary.json rows: {len(raw_keys):,}, of which "
+          f"{sum(1 for k in raw_keys if k.count('.') >= 2):,} have three or more "
+          f"labels -- a count that does NOT answer 'how many are keyed by hostname "
+          f"rather than registrable domain', because that depends on the fold:")
+    print(f"  ICANN-section fold:      merges {folded_icann:,} rows -> "
+          f"{len(prev_icann):,} registrable domains")
+    print(f"  ICANN+PRIVATE fold:      merges {folded_full:,} rows -> "
+          f"{len(prev_full):,} registrable domains")
     print(f"Rows dropped as not hostnames at all: {len(bad)} {bad!r} "
           "-- a literal 'null' key with a prevalence attached is a defect in the "
           "published data, not a domain.")
     print()
-    print(f"Denominator: those {len(universe):,} registrable third-party domains, i.e.")
+    print("Denominator: those registrable third-party domains, i.e. domains Tracker")
-    print("domains Tracker Radar's own crawl of regional top-site lists actually saw.")
+    print("Radar's own crawl of regional top-site lists actually saw. This is Tracker")
-    print("This is Tracker Radar's view of the surface, not a neutral one.")
+    print("Radar's view of the surface, not a neutral one.")
     print()
-    print(f"{'List':30} {'Domains named':>14} {'Share':>7} {'Prevalence-weighted':>20}")
-    for label, m in (("webXray", wx_owner), ("Tracker Radar", tr_owner),
+    LISTS = (("webXray", wx_owner), ("Tracker Radar", tr_owner),
-                     ("Disconnect (properties+res)", dc_owner)):
+             ("Disconnect (properties+res)", dc_owner))
-        hit = [d for d in universe if d in m]
-        wprev = sum(prev[d] for d in hit)
+    def coverage(prev, walk):
-        print(f"{label:30} {len(hit):>14,} {100*len(hit)/len(universe):>6.1f}% "
+        rows = []
-              f"{100*wprev/total_prev:>19.1f}%")
+        total = sum(prev.values())
+        for label, m in LISTS:
+            if walk:
+                hit = [d for d in prev if lookup(d, m)[0] is not None]
+            else:
+                hit = [d for d in prev if d in m]
+            rows.append((label, len(hit), 100 * len(hit) / len(prev),
+* sum(prev[d] for d in hit) / total))
+        return rows
+    VARIANTS = [
+        ("ICANN fold + parent-label lookup  <- USE THIS", prev_icann, True),
+        ("ICANN fold + exact-key lookup", prev_icann, False),
+        ("ICANN+PRIVATE fold + parent-label lookup", prev_full, True),
+        ("ICANN+PRIVATE fold + exact-key lookup  <- THE TRAP", prev_full, False),
+    ]
+    for name, prev, walk in VARIANTS:
+        print(f"{name}  (universe {len(prev):,})")
+        print(f"  {'List':30} {'Named':>8} {'Share':>7} {'Prevalence-weighted':>20}")
+        for label, n, share, wshare in coverage(prev, walk):
+            print(f"  {label:30} {n:>8,} {share:>6.1f}% {wshare:>19.1f}%")
+        print()
+    fg = summary.get("fonts.googleapis.com", {}).get("prevalence")
+    print(f"The single row that drives most of that gap: fonts.googleapis.com, prevalence "
+          f"{fg:.3f}. googleapis.com is a PRIVATE PSL rule, and all three lists name it: "
+          f"webXray={lookup('googleapis.com', wx_owner)[0]!r}, "
+          f"TR={lookup('googleapis.com', tr_owner)[0]!r}, "
+          f"Disconnect={lookup('googleapis.com', dc_owner)[0]!r}.")
+    print("The bottom variant is what the first version of this script did, and it")
+    print("understates every list. The gap between the top and bottom rows is a")
+    print("measurement artefact, not a property of any list -- report which rule you")
+    print("used, because it moves the answer by more than the lists differ.")
     print()
+    # Everything below uses the recommended rule.
+    prev = prev_icann
+    total_prev = sum(prev.values())
+    universe = sorted(prev)
+    owner_of = {label: {d: lookup(d, m)[0] for d in universe} for label, m in
+                (("webXray", wx_owner), ("webXray-root", wx_root_owner),
+                 ("Tracker Radar", tr_owner), ("Disconnect", dc_owner))}
     for n in (100, 1000, 10000):
         topn = sorted(universe, key=lambda d: -prev[d])[:n]
         line = f"top {n:>5} by prevalence: "
         line += "  ".join(
-            f"{label} {100*sum(1 for d in topn if d in m)/len(topn):.0f}%"
+            f"{label} {100*sum(1 for d in topn if owner_of[label][d] is not None)/len(topn):.0f}%"
-            for label, m in (("webXray", wx_owner), ("TR", tr_owner), ("Disconnect", dc_owner)))
+            for label in ("webXray", "Tracker Radar", "Disconnect"))
         print(line)
     print()
-    named_by_none = [d for d in universe if d not in wx_owner and d not in dc_owner]
+    named_by_none = [d for d in universe if owner_of["webXray"][d] is None
+                     and owner_of["Disconnect"][d] is None]
     print(f"Domains in the universe that neither webXray nor Disconnect names an owner "
           f"for: {len(named_by_none):,} ({100*len(named_by_none)/len(universe):.1f}%), "
@@ Line 932: / Line 1987: @@
     print("Top 15 of those by prevalence (Tracker Radar's owner in brackets):")
     for d in sorted(named_by_none, key=lambda x: -prev[x])[:15]:
-        print(f"  {d:34} prev={prev[d]:.4f}  [{tr_owner.get(d, 'TR: none')}]")
+        print(f"  {d:34} prev={prev[d]:.4f}  [{owner_of['Tracker Radar'][d] or 'TR: none'}]")
     top_unowned = max(named_by_none, key=lambda x: prev[x])
     print(f"Most prevalent domain no list can name an owner for: {top_unowned}, "
@@ Line 949: / Line 2004: @@
              ("Tracker Radar", tr_owner, "Disconnect", dc_owner)]
     disagreement_samples = {}
-    for a_label, a, b_label, b in pairs:
+    for a_label, _a_raw, b_label, _b_raw in pairs:
-        both = [d for d in universe if d in a and d in b]
+        # Parent-label lookup here too, for the same reason as section C.
+        a, b = owner_of[a_label], owner_of[b_label]
+        both = [d for d in universe if a[d] is not None and b[d] is not None]
         agree_raw = [d for d in both if a[d] == b[d]]
         agree_norm = [d for d in both if norm_name(a[d]) == norm_name(b[d])]
@@ Line 973: / Line 2030: @@
         print("--- E. Sampled disagreements (highest prevalence first) --------------")
         print()
-        maps = {"webXray": wx_owner, "webXray-root": wx_root_owner,
-                "Tracker Radar": tr_owner, "Disconnect": dc_owner}
         for (a_label, b_label), ds in disagreement_samples.items():
-            a, b = maps[a_label], maps[b_label]
+            a, b = owner_of[a_label], owner_of[b_label]
             print(f"{a_label} vs {b_label}:")
             for d in ds[:args.disagreements]:
@@ Line 985: / Line 2040: @@
     print()
     unchanged = [e["name"] for e in wx if norm_name(e["name"]) == e["name"].lower()]
+    # norm_name() does two things -- strips punctuation, then removes legal forms --
+    # so "names the fold changes" is not "names a legal suffix was removed from".
+    # AT&T, JD.com and "Here, There & Everywhere" change on punctuation alone.
+    # Reported separately after a review pointed out the conflation, 2026-08-17.
+    changed = [e["name"] for e in wx if norm_name(e["name"]) != e["name"].lower()]
+    suffix_hit = [n for n in changed if SUFFIX_RE.search(
+        re.sub(r"[^a-z0-9]+", " ", re.sub(r"[’']", "", n.lower())).strip())]
+    punct_only = [n for n in changed if n not in suffix_hit]
     print(f"webXray owner names the fold leaves untouched: {len(unchanged)} of {len(wx)} "
-          f"({100*len(unchanged)/len(wx):.1f}%). The fold only removes legal forms; it "
+          f"({100*len(unchanged)/len(wx):.1f}%).")
-          "merges no synonyms, so every figure in section D is a lower bound on real "
+    print(f"Of the {len(changed)} it does change, only {len(suffix_hit)} lose a legal-form "
-          "agreement and an upper bound on real disagreement.")
+          f"suffix; the other {len(punct_only)} change on punctuation alone: "
+          + ", ".join(sorted(punct_only)[:8]) + ", ...")
+    print("The fold merges no synonyms, so every figure in section D is a lower bound on "
+          "real agreement and an upper bound on real disagreement.")
     return 0
@@ Line 998: / Line 2064: @@
 ===== J2. Its unedited output =====
-Run as ''python3 scripts/owner_dbs.py --cache out/webxray/cache --disagreements 25''.
+Run as ''python3 scripts/owner_dbs.py %%--cache%% out/webxray/cache %%--disagreements%% 25''.
 <code>
@@ Line 1044: / Line 2110: @@
 --- C. Coverage of the third-party surface ---------------------------
-Public Suffix List: 155b43d46932e933  10,231 rules, 8 exceptions  https://publicsuffix.org/list/public_suffix_list.dat
+Public Suffix List: 155b43d46932e933  6,941 ICANN rules, 3,290 PRIVATE rules, 8 exceptions  https://publicsuffix.org/list/public_suffix_list.dat
-domain_summary.json rows: 47,836. 16,396 of them are keyed by hostname rather than registrable domain; folding to eTLD+1 merges 2,339 rows and leaves 45,525 registrable domains.
+The ICANN/PRIVATE split is the most consequential choice in this script.
+`googleapis.com` is a PRIVATE rule, so folding with the private section
+leaves fonts.googleapis.com standing as its own 'registrable domain' --
+and an exact-key lookup then finds no owner for it, even though all three
+lists name googleapis.com -> Google. Both rules are therefore reported.
+domain_summary.json rows: 47,836, of which 16,396 have three or more labels -- a count that does NOT answer 'how many are keyed by hostname rather than registrable domain', because that depends on the fold:
+  ICANN-section fold:      merges 15,651 rows -> 32,369 registrable domains
+  ICANN+PRIVATE fold:      merges 2,339 rows -> 45,525 registrable domains
 Rows dropped as not hostnames at all: 19 ['null', '[2a01:4f9:2a:26e0::2]', '[2604:2dc0:100:5ce5::]', '[2001:41d0:800:4623::]', '[2001:41d0:602:556f::]', '[2604:8380:2e00:5::2]', '[2604:4500:8:2ea::2]', '[2604:8380:3300:1::2]', '[2604:8380:2900:15::2]', '[2001:41d0:403:579b::]', '[2001:41d0:306:44e6::]', '[2604:4500:a:432::2]', '[2402:1f00:8201:4a2::]', '[2604:4500:6:5a0::2]', '[2604:4500:21:8::4]', '[2402:1f00:8001:2518::]', '[2402:1f00:8300:c97::]', '[2604:8380:2f00:16::2]', '[2001:41d0:700:782c::]'] -- a literal 'null' key with a prevalence attached is a defect in the published data, not a domain.
-Denominator: those 45,525 registrable third-party domains, i.e.
+Denominator: those registrable third-party domains, i.e. domains Tracker
-domains Tracker Radar's own crawl of regional top-site lists actually saw.
+Radar's own crawl of regional top-site lists actually saw. This is Tracker
-This is Tracker Radar's view of the surface, not a neutral one.
+Radar's view of the surface, not a neutral one.
+ICANN fold + parent-label lookup  <- USE THIS  (universe 32,369)
+  List                              Named   Share  Prevalence-weighted
+  webXray                             669    2.1%                58.6%
+  Tracker Radar                     5,581   17.2%                84.3%
+  Disconnect (properties+res)       2,268    7.0%                80.5%
+ICANN fold + exact-key lookup  (universe 32,369)
+  List                              Named   Share  Prevalence-weighted
+  webXray                             669    2.1%                58.6%
+  Tracker Radar                     5,581   17.2%                84.3%
+  Disconnect (properties+res)       2,268    7.0%                80.5%
+ICANN+PRIVATE fold + parent-label lookup  (universe 45,525)
+  List                              Named   Share  Prevalence-weighted
+  webXray                          11,150   24.5%                59.7%
+  Tracker Radar                    18,099   39.8%                84.8%
+  Disconnect (properties+res)      10,389   22.8%                80.6%
+ICANN+PRIVATE fold + exact-key lookup  <- THE TRAP  (universe 45,525)
+  List                              Named   Share  Prevalence-weighted
+  webXray                             657    1.4%                54.5%
+  Tracker Radar                     5,539   12.2%                79.3%
+  Disconnect (properties+res)       2,277    5.0%                75.8%
-List                            Domains named   Share  Prevalence-weighted
+The single row that drives most of that gap: fonts.googleapis.com, prevalence 0.369. googleapis.com is a PRIVATE PSL rule, and all three lists name it: webXray='Google APIs', TR='Google LLC', Disconnect='Google'.
-webXray                                   657    1.4%                54.5%
+The bottom variant is what the first version of this script did, and it
-Tracker Radar                           5,539   12.2%                79.3%
+understates every list. The gap between the top and bottom rows is a
-Disconnect (properties+res)             2,277    5.0%                75.8%
+measurement artefact, not a property of any list -- report which rule you
+used, because it moves the answer by more than the lists differ.
-top   100 by prevalence: webXray 70%  TR 95%  Disconnect 91%
+top   100 by prevalence: webXray 71%  Tracker Radar 98%  Disconnect 94%
-top  1000 by prevalence: webXray 26%  TR 67%  Disconnect 67%
+top  1000 by prevalence: webXray 26%  Tracker Radar 69%  Disconnect 68%
-top 10000 by prevalence: webXray 5%  TR 27%  Disconnect 16%
+top 10000 by prevalence: webXray 5%  Tracker Radar 29%  Disconnect 17%
-Domains in the universe that neither webXray nor Disconnect names an owner for: 43,052 (94.6%), 19.3% of prevalence weight.
+Domains in the universe that neither webXray nor Disconnect names an owner for: 29,896 (92.4%), 14.4% of prevalence weight.
 Top 15 of those by prevalence (Tracker Radar's owner in brackets):
-  fonts.googleapis.com               prev=0.3692  [TR: none]
-  ajax.googleapis.com                prev=0.0622  [TR: none]
-  imasdk.googleapis.com              prev=0.0435  [TR: none]
   tiktokw.us                         prev=0.0406  [ByteDance Ltd.]
   consentmanager.net                 prev=0.0359  [consentmanager AB]
-  jnn-pa.googleapis.com              prev=0.0270  [TR: none]
-  maps.googleapis.com                prev=0.0261  [TR: none]
   rapidedge.io                       prev=0.0247  [TR: none]
   raptivecdn.com                     prev=0.0231  [TR: none]
@@ Line 1076: / Line 2170: @@
   shopifycdn.com                     prev=0.0181  [Shopify Inc.]
   digitalaudience.io                 prev=0.0152  [Social Audience B.V.]
-  translate.googleapis.com           prev=0.0144  [TR: none]
+  openwebmp.com                      prev=0.0130  [TR: none]
-  storage.googleapis.com             prev=0.0141  [TR: none]
+  userway.org                        prev=0.0122  [TR: none]
-Most prevalent domain no list can name an owner for: fonts.googleapis.com, prevalence 0.369 (rounded to 3 dp for quoting).
+  copper6.com                        prev=0.0113  [TR: none]
+  ahrefs.com                         prev=0.0112  [Ahrefs Pte Ltd]
+  anyrtb.com                         prev=0.0111  [TR: none]
+  sparteo.com                        prev=0.0108  [TR: none]
+  axiom.co                           prev=0.0102  [TR: none]
+Most prevalent domain no list can name an owner for: tiktokw.us, prevalence 0.041 (rounded to 3 dp for quoting).
 Dropped rows and their prevalence, rounded to 3 dp for quoting: null=0.024, [2a01:4f9:2a:26e0::2]=0.000, [2604:2dc0:100:5ce5::]=0.000
 --- D. Do two lists that both know a domain agree on the owner? ------
-webXray vs Tracker Radar: 601 domains named by both
+webXray vs Tracker Radar: 612 domains named by both
-  identical owner string               29 (4.8%)
+  identical owner string               29 (4.7%)
-  agree after legal-suffix fold       304 (50.6%)
+  agree after legal-suffix fold       305 (49.8%)
-  one name contains the other         100 (16.6%)  [e.g. Amazon / Amazon Technologies]
+  one name contains the other         109 (17.8%)  [e.g. Amazon / Amazon Technologies]
-  neither                             197 (32.8%)  <- needs a human
+  neither                             198 (32.4%)  <- needs a human
-  ...weighted by prevalence, the 'needs a human' share is 30.2%
+  ...weighted by prevalence, the 'needs a human' share is 29.0%
-webXray-root vs Tracker Radar: 601 domains named by both
+webXray-root vs Tracker Radar: 612 domains named by both
-  identical owner string               34 (5.7%)
+  identical owner string               34 (5.6%)
-  agree after legal-suffix fold       275 (45.8%)
+  agree after legal-suffix fold       276 (45.1%)
-  one name contains the other          90 (15.0%)  [e.g. Amazon / Amazon Technologies]
+  one name contains the other          97 (15.8%)  [e.g. Amazon / Amazon Technologies]
-  neither                             236 (39.3%)  <- needs a human
+  neither                             239 (39.1%)  <- needs a human
-  ...weighted by prevalence, the 'needs a human' share is 49.2%
+  ...weighted by prevalence, the 'needs a human' share is 51.1%
-webXray vs Disconnect: 461 domains named by both
+webXray vs Disconnect: 464 domains named by both
-  identical owner string              170 (36.9%)
+  identical owner string              171 (36.9%)
-  agree after legal-suffix fold       212 (46.0%)
+  agree after legal-suffix fold       213 (45.9%)
-  one name contains the other          33 (7.2%)  [e.g. Amazon / Amazon Technologies]
+  one name contains the other          34 (7.3%)  [e.g. Amazon / Amazon Technologies]
-  neither                             216 (46.9%)  <- needs a human
+  neither                             217 (46.8%)  <- needs a human
-  ...weighted by prevalence, the 'needs a human' share is 46.5%
+  ...weighted by prevalence, the 'needs a human' share is 44.5%
-webXray-root vs Disconnect: 461 domains named by both
+webXray-root vs Disconnect: 464 domains named by both
-  identical owner string              163 (35.4%)
+  identical owner string              164 (35.3%)
-  agree after legal-suffix fold       194 (42.1%)
+  agree after legal-suffix fold       195 (42.0%)
-  one name contains the other          39 (8.5%)  [e.g. Amazon / Amazon Technologies]
+  one name contains the other          39 (8.4%)  [e.g. Amazon / Amazon Technologies]
-  neither                             228 (49.5%)  <- needs a human
+  neither                             230 (49.6%)  <- needs a human
-  ...weighted by prevalence, the 'needs a human' share is 62.6%
+  ...weighted by prevalence, the 'needs a human' share is 64.1%
-Tracker Radar vs Disconnect: 1,532 domains named by both
+Tracker Radar vs Disconnect: 1,538 domains named by both
-  identical owner string               82 (5.4%)
+  identical owner string               82 (5.3%)
-  agree after legal-suffix fold       630 (41.1%)
+  agree after legal-suffix fold       634 (41.2%)
-  one name contains the other         318 (20.8%)  [e.g. Amazon / Amazon Technologies]
+  one name contains the other         320 (20.8%)  [e.g. Amazon / Amazon Technologies]
-  neither                             584 (38.1%)  <- needs a human
+  neither                             584 (38.0%)  <- needs a human
-  ...weighted by prevalence, the 'needs a human' share is 30.7%
+  ...weighted by prevalence, the 'needs a human' share is 29.7%
 --- E. Sampled disagreements (highest prevalence first) --------------
@@ Line 1152: / Line 2251: @@
   doubleclick.net                  prev=0.4456  webXray-root='Alphabet'  Tracker Radar='Google LLC'
   gstatic.com                      prev=0.4013  webXray-root='Alphabet'  Tracker Radar='Google LLC'
+  googleapis.com                   prev=0.3692  webXray-root='Alphabet'  Tracker Radar='Google LLC'
   google-analytics.com             prev=0.3489  webXray-root='Alphabet'  Tracker Radar='Google LLC'
   googlesyndication.com            prev=0.2115  webXray-root='Alphabet'  Tracker Radar='Google LLC'
@@ Line 1172: / Line 2272: @@
 yield.com                     prev=0.0467  webXray-root='Azerion'  Tracker Radar='Improve Digital BV'
   fwmrm.net                        prev=0.0439  webXray-root='Comcast'  Tracker Radar='FreeWheel'
-  rfihub.com                       prev=0.0429  webXray-root='Amazon'  Tracker Radar='Zeta Global'
 webXray vs Disconnect:
@@ Line 1205: / Line 2304: @@
   doubleclick.net                  prev=0.4456  webXray-root='Alphabet'  Disconnect='Google'
   gstatic.com                      prev=0.4013  webXray-root='Alphabet'  Disconnect='Google'
+  googleapis.com                   prev=0.3692  webXray-root='Alphabet'  Disconnect='Google'
   google-analytics.com             prev=0.3489  webXray-root='Alphabet'  Disconnect='Google'
   googlesyndication.com            prev=0.2115  webXray-root='Alphabet'  Disconnect='Google'
@@ Line 1226: / Line 2326: @@
 mdn.net                         prev=0.0616  webXray-root='Alphabet'  Disconnect='Google'
   pippio.com                       prev=0.0594  webXray-root='Acxiom'  Disconnect='LiveRamp'
-  dotomi.com                       prev=0.0583  webXray-root='Here, There & Everywhere'  Disconnect='PublicisGroupe'
 Tracker Radar vs Disconnect:
@@ Line 1257: / Line 2356: @@
 --- F. Named-entity residue of the suffix fold -----------------------
-webXray owner names the fold leaves untouched: 782 of 827 (94.6%). The fold only removes legal forms; it merges no synonyms, so every figure in section D is a lower bound on real agreement and an upper bound on real disagreement.
+webXray owner names the fold leaves untouched: 782 of 827 (94.6%).
+Of the 45 it does change, only 16 lose a legal-form suffix; the other 29 change on punctuation alone: 56.com, AT&T, Ask.com, Bootstrap_China, Clearstream.TV, Cm_browser, Dictionary.com, Dun & Bradstreet, ...
+The fold merges no synonyms, so every figure in section D is a lower bound on real agreement and an upper bound on real disagreement.
 </code>
 ===== K. Unedited output: scripts/owner_adjudication.py =====
-Run as ''python3 scripts/owner_adjudication.py --table''.
+Run as ''python3 scripts/owner_adjudication.py %%--table%%''.
 <code>
- hand-adjudicated disagreements, decided 2026-08-17 against primary sources.
+disagreements adjudicated 2026-08-17 against primary sources; 2 failed the sourcing bar and are EXCLUDED from every tally below, leaving 28.
+Excluded: 360yield.com (UNRESOLVED -- no dated primary source), fwmrm.net (acquired 2014; post-2026 spinoff UNRESOLVED)
 Selected as the highest-prevalence disagreements, NOT sampled at random:
 these tallies describe this set and are not an error rate for any list.
 List               current        stale  granularity        error       absent
-webXray                  3           17            4            1            5
+webXray                  1           18            4            1            4
-Tracker Radar            9           12            7            2            0
+Tracker Radar            8           12            6            2            0
-Disconnect              22            0            0            0            8
+Disconnect              26            0            2            0            0
-webXray: has an entry for 25 of 30; of those, 3 name today's owner (12% of its own entries).
+webXray: has an entry for 24 of 28; of those, 1 name today's owner (4% of its own entries).
-Tracker Radar: has an entry for 30 of 30; of those, 9 name today's owner (30% of its own entries).
+Tracker Radar: has an entry for 28 of 28; of those, 8 name today's owner (29% of its own entries).
-Disconnect: has an entry for 22 of 30; of those, 22 name today's owner (100% of its own entries).
+Disconnect: has an entry for 28 of 28; of those, 26 name today's owner (93% of its own entries).
+The 'absent' column cannot be read as a coverage result either: the rows were
+ranked by Tracker Radar's own prevalence field, so Tracker Radar covering all
+of 28 is partly how the sample was drawn.
-Rows where the primary-source bar was not met: 2 -- 360yield.com, fwmrm.net
 Rows scored as an outright error rather than staleness: 3 -- 1rx.io (webXray), jsdelivr.net (Tracker Radar), stackadapt.com (Tracker Radar)
 adnxs.com              Microsoft Corporation                        stale        current      current      https://www.prnewswire.com/news-releases/att-agrees-to-microsoft-acquisition-of-xandr-301448996.html
 rubiconproject.com     Magnite, Inc.                                stale        current      current      https://investor.magnite.com/news-releases/news-release-details/rubicon-project-and-telaria-complete-merger-following
-yahoo.com              Yahoo Inc. (Apollo-managed funds majority)   stale        current      absent       https://www.apollo.com/insights-news/pressreleases/2021/09/apollo-funds-complete-acquisition-of-yahoo-161530593
+yahoo.com              Yahoo Inc. (Apollo-managed funds majority)   stale        current      current      https://www.apollo.com/insights-news/pressreleases/2021/09/apollo-funds-complete-acquisition-of-yahoo-161530593
 facebook.com           Meta Platforms, Inc.                         stale        stale        current      https://about.fb.com/news/2021/10/facebook-company-is-now-meta/
-bidr.io                Comcast (FreeWheel/Beeswax)                  stale        stale        current      https://corporate.comcast.com/press/releases/freewheel-acquires-beeswax
+bidr.io                Comcast (FreeWheel/Beeswax)                  stale        stale        current      https://www.freewheel.com/news/freewheel-to-acquire-ad-tech-leader-beeswax
 turn.com               Nexxen International Ltd.                    stale        stale        current      https://investors.nexxen.com/news-releases/news-release-details/tremor-international-tremor-announces-closing-amobee-acquisition
 rx.io                 Nexxen International Ltd.                    error        stale        current      https://nexxen.com/tremor-international-group-rebrands-as-nexxen/
@@ Line 1294: / Line 2399: @@
 crwdcntrl.net          Publicis Groupe S.A. (Epsilon/Lotame)        stale        stale        current      https://www.publicisgroupe.com/en/news/press-releases/publicis-to-acquire-lotame-the-world-s-leading-independent-end-to-end-data-solution
 dotomi.com             Publicis Groupe S.A. (Epsilon/Conversant)    stale        granularity  current      https://www.globenewswire.com/news-release/2019/07/02/1877064/0/en/Publicis-Groupe-finalizes-the-acquisition-of-Epsilon.html
-agkn.com               TransUnion LLC                               stale        current      absent       https://www.globenewswire.com/news-release/2021/12/01/2344146/0/en/TransUnion-and-Neustar-Announce-Transaction-Close.html
+agkn.com               TransUnion LLC                               stale        current      current      https://www.globenewswire.com/news-release/2021/12/01/2344146/0/en/TransUnion-and-Neustar-Announce-Transaction-Close.html
-rlcdn.com              LiveRamp Holdings, Inc.                      stale        current      absent       https://www.sec.gov/Archives/edgar/data/0000733269/000119312518279280/d616065d8k12b.htm
+rlcdn.com              LiveRamp Holdings, Inc.                      stale        current      current      https://www.sec.gov/Archives/edgar/data/0000733269/000119312518279280/d616065d8k12b.htm
-tapad.com              Experian plc                                 stale        granularity  absent       https://www.experianplc.com/newsroom/press-releases/2020/experian-acquires-tapad
+tapad.com              Experian plc                                 stale        granularity  granularity  https://www.experianplc.com/newsroom/press-releases/2020/experian-acquires-tapad
-contextweb.com         Internet Brands, Inc. (PulsePoint)           absent       stale        current      https://www.pulsepoint.com/press-releases/internet-brands-to-acquire-pulsepoint
+contextweb.com         Internet Brands, Inc. (PulsePoint)           stale        stale        current      https://www.pulsepoint.com/press-releases/internet-brands-to-acquire-pulsepoint
 temu.com               PDD Holdings Inc.                            absent       stale        current      https://www.sec.gov/Archives/edgar/data/1737806/000110465924051610/pdd-20231231x20f.htm
-lijit.com              Sovrn Holdings, Inc.                         stale        current      absent       https://lijit.com/
+lijit.com              Sovrn Holdings, Inc.                         stale        current      current      https://lijit.com/
 teads.tv               Teads Holding Co.                            stale        current      current      https://www.globenewswire.com/news-release/2025/02/03/3019410/0/en/Outbrain-Completes-the-Acquisition-of-Teads.html
 blismedia.com          Deutsche Telekom AG (T-Mobile US)            absent       stale        current      https://report.telekom.com/interim-report-q2-2025/financial-statements/significant-events-and-transactions/changes-in-the-composition-of-the-group.html
@@ Line 1305: / Line 2410: @@
 stackadapt.com         StackAdapt Inc. (independent)                absent       error        current      https://www.stackadapt.com/legal-document-centre
 bidswitch.net          Criteo S.A. (IPONWEB)                        granularity  granularity  current      https://criteo.investorroom.com/2022-08-03-CRITEO-REPORTS-STRONG-SECOND-QUARTER-2022-RESULTS
-intentiq.com           AlmondNet Group                              granularity  granularity  absent       https://www.intentiq.com/who-we-are/
+intentiq.com           AlmondNet Group                              granularity  granularity  granularity  https://www.intentiq.com/who-we-are/
-ipredictive.com        Viant Technology LLC (Adelphic)              absent       granularity  absent       https://www.viantinc.com/company/news/press-releases/viant-completes-integration-of-adelphic-into-viant-advertising-cloud/
+ipredictive.com        Viant Technology LLC (Adelphic)              absent       granularity  current      https://www.viantinc.com/company/news/press-releases/viant-completes-integration-of-adelphic-into-viant-advertising-cloud/
 simpli.fi              Simplifi Holdings Inc. (GTCR + Blackstone)   granularity  granularity  current      https://www.blackstone.com/news/press/simpli-fi-a-leading-programmatic-advertising-platform-announces-completion-of-significant-investment-from-blackstone-at-1-5-billion-valuation/
 casalemedia.com        Index Exchange, Inc. (independent)           current      current      current      https://www.indexexchange.com/team/andrew-casale/
-yield.com           Azerion Group N.V. (Improve Digital)         current      granularity  current      https://improvedigital.com/about/
+yield.com           Azerion Group N.V. (Improve Digital)         current      granularity  current      https://improvedigital.com/about/  EXCLUDED
-fwmrm.net              Comcast Corporation (FreeWheel)              current      current      absent       https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0001166691
+fwmrm.net              Comcast Corporation (FreeWheel)              current      current      current      https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0001166691  EXCLUDED
 </code>
@@ Line 1316: / Line 2421: @@
   * [[programming:crawler:webxray]] — the page these notes are behind.
+  * [[provenance:programming:crawler:webxray:random_sample]] — the appendix: all eleven scripts of the 2026-09-05 random sample, all 10 unedited outputs, and the adjudication brief.
   * [[literature:corpus]] — corpus scope, the selection funnel and extraction stability.
@@ Line 1321: / Line 2427: @@
 <bibtex bibliography></bibtex>
+===== Markup sweep, 2026-09-17 =====
+Mechanical rendering repair only: a fresh live raw/XHTML export of 188 pages was checked with ''check_wrap.mjs'' and ''check_typography.mjs''. Affected plugin tags, CLI flags and heading markup were repaired; no figures or substantive prose were changed. The resulting source and rendered DOM were re-checked after saving.
+===== Embedded-output refresh, 2026-09-22 =====
+**What was wrong.** The counts on this page were right; the //embedded copy of the
+script's output// in [[#I. Unedited output: scripts/report_webxray.mjs|§I]] was a
+pre-2026-09-11 paste. ''report_webxray.mjs'' has been re-run and the whole block
+regenerated rather than line-edited, so what §I shows is again what the script prints.
+Trigger: the shared quote checkers gained a ''pypdf'' fallback on 2026-09-04 and
+''report_webxray.mjs'' was patched to use it, so its Summary line changed from
+"1 below threshold" to "0 rescued from the PDF, 1 below threshold in both renderings".
+The quote **counts** did not move — 16 quotes, 4 exact, 11 partial, 1 below threshold —
+so the wording was the only thing this refresh was filed for. Diffing the rest of the
+block against the fresh output found four more stale lines, listed below.
+<code>
+$ node scripts/report_webxray.mjs --quotes
+Summary: 16 quotes: 4 exact, 11 partial (>=60% of 4-word windows), 0 rescued from the PDF, 1 below threshold in both renderings.
+</code>
+^ Line in §I (or, for A10, in §A) ^ Was ^ Is ^ Why ^
+| §H Summary | ''1 below threshold'' | ''0 rescued from the PDF, 1 below threshold in both renderings'' | the fallback added 2026-09-04; **no count moved** |
+| §H heading | ''Evidence quotes behind the schema tuples'' | ''… (with PDF fallback)'' | same patch |
+| §A sweep line | ''15 papers … and 1.3% of the 1120 that ran a crawl'' | ''15 papers … 13 of them are among the 1120 that ran a crawl (1.2% of that population)'' | the 2026-09-11 denominator fix: a corpus-wide numerator was being divided by the crawled denominator. [[:programming:crawler:webxray]] has carried the corrected 1.2% since then; only this embedded paste was stale |
+| §E union line | ''136 papers (2.3% of the corpus; 12.1% of the 1120 that crawled)'' | ''136 papers (2.3% of the corpus). 107 of them are among the 1120 that crawled -- 9.6% of that population'' | same fix, larger figure |
+| **§A, row A10** | ''**136** (2.3% / 12.1%)'' | ''**136** (2.3% of the corpus); **107** … (**9.6%**)'' | the only **prose** figure this refresh moved. The content page and ''report_webxray.mjs'' both say 107 / 9.6%; this row still said 12.1% |
+| §Z, five ''owner_dbs.py'' lines | 45,525-domain frame; 54.5 / 79.3 / 75.8%; 16,396 rows | 32,369-domain frame; 58.6 / 84.3 / 80.5%; the ICANN fold merges 15,651 of 47,836 rows | the ICANN-section fold, adopted 2026-09-05. [[#C3.1 The frame, and what it excludes|§C3.1]], §B4 and §B6 have carried the 32,369 frame since then; again only the paste was stale |
+**The lesson, and it is the one this site keeps relearning.** Every figure in the
+prose was already correct. What was stale was the block that exists to //prove// the
+prose — so a reader checking the page against its own embedded output would have been
+shown a contradiction, in the direction of the numbers that had already been retracted.
+An embedded output block has to be regenerated whenever its script changes, not when
+its headline figure changes; and ''check_page_numbers.mjs'' cannot see this, because
+every value in the stale block is a real number that really was printed once.
+**Scope of this edit.** The §I block in full, the §A10 row, and the one sentence in §5
+that quotes the quote-check verdicts. ''owner_dbs.py'' and ''owner_sample.py'' were
+**not** re-run — the §Z lines are literals inside ''report_webxray.mjs'' and were taken
+from its output, not recomputed. No other figure, fold or citation on this page was
+touched, and no ''%%{[citekey]}%%'' was added or moved.
+^ Item ^ Value ^
+| Date | 2026-09-22, unsupervised |
+| Command | ''node scripts/report_webxray.mjs %%--quotes%%'' |
+| Artifact | ''out/qc0922/report_webxray_quotes.txt'' (the full 232-line run pasted into §I) |
+| Script changes | none — ''report_webxray.mjs'' already had the fallback |
+| Reviewers | one ''sonnet'' figures-vs-script pass |
+| Pages saved | this page only — [[:programming:crawler:webxray]] already carries 107 / 9.6% / 1.2% and was verified unchanged |