Differences

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

--- provenance:programming:registration [2026/08/27 12:46] – Query log for programming:registration. Authored by Claude. karel.kubicek.claude
+++ provenance:programming:registration [2026/08/27 13:08] (current) – Review log: three focused GPT 5.6 Sol medium passes then one generic; accept/reject for Cookie Hunter, Kubicek, Tripwire, recaptcha, AZcaptcha, 17-of-81 qualification. Authored by Claude karel.kubicek.claude
@@ Line 6: / Line 6: @@
   * Run date: 2026-08-27 (UTC).
-  * Authoring agent: Cursor Grok 4.6, executing the drain item ''programming:registration (improve; 2.5 KB stub)'' directly rather than via a headless ''claude -p'' session. Review: three focused GPT 5.6 Sol medium passes, then one generic pass with no checklist (see Review log).
+  * Authoring agent: Cursor Grok 4.6, executing the drain item ''programming:registration (improve; 2.5 KB stub)'' directly rather than via a headless ''claude -p'' session. Review: three focused GPT 5.6 Sol medium passes on the frozen first draft (findings applied below), then one generic pass with no checklist after those fixes (see Review log).
   * Corpus at run time: 5,859 extracted papers, 2010–2026, CCS, IMC, NDSS, PETS, USENIX Security, TheWebConf, IEEE S&P. Read-only inputs under ''/workspace/publications_dataset/data/''.
   * Read first: ''data/extract/OVERVIEW.md'', the task spec, live ''programming:registration'' stub (rev 1739892438, 2,544 bytes), live ''programming:interaction'' (rev 1787833760 — underway, **not edited**), [[Practices:Ethics]], Shepherd MADWeb PDF, Kubicek WWW 2024 author PDF.
@@ Line 50: / Line 50: @@
   * PETS 2022 59% / 5.5% from the **body** paragraph (of websites that sent mail / unsolicited marketing with no confirmation), not the abstract.
   * Login-policies 39% from **ground-truth** manual accounts, not of 45.0K.
-  * Cookie Hunter 13.7% of **signup domains**, ~1.6% of **all domains**. Kubicek compares to the latter.
+  * Kubicek 25.7% (169,765) is of **660,202 unique domains**, not of 504,509 loaded (that ratio is 33.7%). The PDF writes the 25.7% "among the loaded websites".
 ===== External and industry verification =====
@@ Line 57: / Line 57: @@
 | Cloudflare Turnstile ''index.md'' | CAPTCHA-free; Managed / Non-interactive / Invisible; WCAG 2.2 AA; updated 2026-08-14 | fetched | Used |
 | Turnstile widget ''index.md'' | ''challenges.cloudflare.com''; updated 2026-04-16 | fetched | Used |
-| Google reCAPTCHA docs | still live | HTTP 200 | "docs still serve", not a share |
+| Google reCAPTCHA docs | landing deprecated; replacement is Cloud Fraud Defense | landing HTTP 200 + "deprecated"; ''cloud.google.com/recaptcha/docs/overview'' HTTP 200 + "Fraud Defense" | Used; do not say only "still serves 200" |
-| hCaptcha docs | still live | HTTP 200 | same |
+| AZcaptcha docs | 2023 papers: OCR not humans | ''azcaptcha.com/docs'' HTTP 200; live text mentions a workers pool | Qualify as paper-era |
+| hCaptcha docs | still live | HTTP 200 | Used; HTTP 200, not a share |
 | Shepherd PDFs + Crossref 10.14722/madweb.2020.23008 | 7,113; 2,759; ≥97% | HTTP 200; title match | Used; out of corpus |
 | Kubicek author PDF + access page | 660,202; 5.9%; 75/20/2/3; farm | HTTP 200 | Used; extraction-gap |
@@ Line 77: / Line 78: @@
 ===== Judgement calls =====
-  * Publish **17**, not 81, as the instrument figure; keep 81 visible because the neighbour already printed it.
+  * Publish **17**, not 81, as the instrument figure; keep 81 visible because the neighbour already printed it. The 17 is a hand split of the 81, not a census of the 776 papers the schema did not mark as logged in.
   * Do not edit ''programming:interaction'' in this sitting.
   * Cite Kubicek from the PDF rather than omit the largest registration crawl.
@@ Line 138: / Line 139: @@
 in extractions.jsonl: NO
 paper.cols.txt: ABSENT. Cited from people.inf.ethz.ch/basin/pubs/www24.pdf
+figures (author PDF, 2026-08-27): 660,202 unique; loaded 504,509;
+  form on 25.7% (169,765) of 660,202 unique domains (PDF writes "among the loaded websites";
+,765/504,509 loaded = 33.7%).
 OK — ROLE covers the schema set both ways; SCHEMA_MISS keys exist and are not schema-logged.
@@ Line 144: / Line 148: @@
 ===== Review log =====
-Three focused GPT 5.6 Sol medium passes (figures vs script; citations/quotes; external currency) and one generic pass with no checklist were run after drafts were frozen. Findings and accept/reject decisions are appended here after those passes. This paragraph is not a review checklist.
+Three focused passes on **GPT 5.6 Sol medium**, as directed for this sitting (the task spec's sonnet/fable split was overridden). They ran in parallel on the frozen first draft (''out/freeze_registration/''). The content page was not edited while those reviewers were running. Findings were applied, then a generic pass with no checklist on ''out/freeze_registration_generic/'' (agent ''5a44afde-2a4d-4fa4-85cb-33c2da9e4358''). Every accept/reject is below. No GENERIC_REVIEW placeholder.
+==== GPT 5.6 Sol medium — figures versus script ====
+^ Finding ^ Disposition ^
+| Cookie Hunter table and report joined 25,242 with 13.7% of 168,594. Arithmetic is 15.0%. The paper's 13.7% is registered-and-logged-in of signup domains; 25,242 is a separate accounts-created count they call "almost 12%". | **Accepted.** Split the table into three rows. Report Z no longer writes "13.7% -> 25,242". |
+| Kubicek "form on 25.7% (169,765) of loaded" is the unique-domain rate: 169,765/660,202 = 25.7%; 169,765/504,509 = 33.7%. The PDF writes the 25.7% "among the loaded websites". | **Accepted.** Page and report now name the 660,202 denominator and the 33.7% loaded ratio. |
+| ''login_diff.py'' substring false positives (''accounting'', ''gravatar'', ''logout-policy''). | **Accepted.** Path-segment equality plus ''-link''/''-btn'' suffixes. New self-tests; documented run re-pasted. |
+| 857, 81, 17, 2.0%, ROLE shares, year buckets, PETS 2022 59%/5.5%, login-policies 39% of ground-truth. | Accepted, no change. |
+==== GPT 5.6 Sol medium — citations and quotes ====
+^ Finding ^ Disposition ^
+| Cookie Hunter rate conflates two results (same as figures pass). | **Accepted** (same split). |
+| The Cookie Hunter farm-refusal quote had no local ''{[drakonakis2020_cookie]}''. | **Accepted.** Citekey on the quote paragraph. |
+| "Tripwire yes" under human CAPTCHA farms. Paper used DeCaptcher; it does not establish that the solvers were human. | **Accepted.** Tripwire is a third-party solver, not a named human farm. |
+| All 20 citation keys unique; quotes otherwise verbatim. | Accepted, no change. |
+==== GPT 5.6 Sol medium — external currency ====
+^ Finding ^ Disposition ^
+| ''developers.google.com/recaptcha'' HTTP 200 but the page is deprecated in favour of Google Cloud Fraud Defense. | **Accepted.** Page names the deprecation and the replacement URL. Probe checks both. |
+| AZcaptcha "OCR not humans" is a 2023 paper claim; live docs mention a workers pool. | **Accepted.** Dated to those papers; live docs recorded. |
+| Turnstile, hCaptcha, Shepherd/Kubicek PDFs, Tranco ''82Q3V'', Crossref records. | Accepted, no change. |
+==== GPT 5.6 Sol medium — generic ====
+Ran after the focused findings were applied. No checklist. The content page was not edited while this reviewer ran.
+^ Finding ^ Disposition ^
+| 17/857 is presented as corpus prevalence; only the 81 schema hits were hand-read. Four SCHEMA_MISS papers prove false negatives exist. | **Accepted as a qualification.** Intro, WRAP, and methodology now say the 17 is a hand split of the 81, not a census of the 776 unlabelled papers. The published 17 is unchanged. |
+| "Logged-out is a design" overstates: the schema records labels, not motives, and the quote cannot be checked. | **Accepted.** Opening now: default practice; 553 are labelled ''none''. |
+| "There is a consensus that almost nobody reads the ToS of 45.0K sites." Five papers cannot establish that. | **Accepted.** Now: those papers wrote they did not check; that is five papers, not a vote of 45.0K. |
+| ''aria-label="Log out"'' is not a marker: segments() does not collapse spaces to the keyword ''log-out''. | **Accepted.** Whole-attribute canonicalisation; eighth self-test. |
+| ''--quotes'' without a regex is skipped; Shepherd DOI HTTP status is printed and not asserted. | **Accepted.** Bare ''--quotes'' now exits 1. DOI must be HTTP 200. |
+| Open-questions box uses a WRAP todo tag the reviewer called lowercase. | **Rejected.** The tag is the required uppercase form; ''check_wrap.mjs'' passed. The defect that guard catches is the lowercase ''wrap'' tag. |
+| Review log promised a generic heading that was not yet written. | **Accepted.** This heading is that log. |
 ====== References ======