Working notes behind archives — every query with its population and denominator, the report script and its unedited output, the fold and its full residue, the quotes that were checked, the external sources that were verified or rejected, and what could not be established. Corpus-level caveats that apply to every page on this site are on corpus and are not restated here.

Contemporaneous. Written during the run that produced the content page, 2026-08-13, not reconstructed afterwards.

The judgement call. The prior design:archives was already a good page on which archive to pick and how to query it — HTTP Archive's BigQuery setup, the Wayback Machine's API surface, a short limitations section, and a TODO asking for more on alternative archives. It was not a candidate for a parallel page: a second page called “archive fidelity” would have split one topic into two, and the reader arriving with “should I use an archive” would have found the wrong half.

So this was an extension. What was kept from the prior page, edited for accuracy: the HTTP Archive section (BigQuery access steps, both SQL examples, the cost warning, the HAR-download route), the Wayback Machine section, the archive-selection criteria, and the reference to Wikipedia's archiving-initiatives list. What was added: the Use in Publications section built from this corpus, the whole What an Archive Preserves section, the Doing It Right checklist, The anachronism trap, Recording Your Own Archive, What to Report, and Open Questions. What was corrected: six factual or misleading statements, listed in §7.

The TODO (“Adding more information to alternative archives would be helpful”) is resolved: Arquivo.pt, Archive-It, Common Crawl and the national archives now have text with measured coverage figures rather than a list of names, and the reason the list is short — the aggregator that would have enumerated them is gone — is now stated.

Every figure on the content page comes from one of these. “All 5,859 papers” is never a denominator on that page.

The intersection is not cosmetic and the first draft got it wrong. Shares of empirical were computed as 122 / 5,118 and 1,261 / 5,118 — a numerator drawn from all 5,859 papers over a denominator of 5,118. It inflated the headline from 2.2% to 2.4%, the live-crawl comparison from 24.2% to 24.6%, and shifted four of five rows of the trend table and five of seven venue rows. Caught by the generic review pass (§11, pass D), which noticed it in the script rather than on the page — the figures pass had diffed page against script and both were wrong in the same way. That the corpus overview independently reports 65 of 5,118 for web-archive is the check that should have been run at the start.

The two counts of “papers using an archive” — 68 and 122 — are both published on the content page, neither is presented as the answer, and the reason is in §5.

cd /workspace/artifacts/wiki
node scripts/report_archives.mjs                 # the audit trail below
node scripts/report_archives.mjs --wiki          # DokuWiki tables
node scripts/report_archives.mjs --list          # the 122 papers, by year
node scripts/report_archives.mjs --quotes Wayback   # evidence quotes
node scripts/quote_check.mjs --temporal 'wayback|internet archive|archive\.org|web-archive|common ?crawl|http ?archive' --show 30

The unedited output of the first command, 2026-08-13, is in §12. Every figure on the content page appears in it.

temporal.mode is an enum and enums are the stable part of this extraction, so web-archive = 68 papers looks like the publishable number. It is a floor, and this was established by hand rather than assumed:

• [1Singh, Sachin Kumar; Mahmud, Faisal; Ricci, Robert; Siby, Sandra (2026): "The Empire Strikes Back (at Your Privacy): An Archaeology of Tracking on Government Websites", Proceedings on Privacy Enhancing Technologies 2026(2):108-126. (DOI)] (PETS 2026) takes all of its measured content from the Wayback Machine — 16,500 archived government websites, 1996–2025, two snapshots per site per year. (Its site list comes from a prior IMC 2024 study; the bytes it measures are all archival.) Its temporal.mode is existing-dataset, with source = “Internet Archive Wayback Machine”. It is not in the 68.
• The extraction is not wrong to do this. From the paper's own framing the archive is a pre-existing dataset. The enum answers “where did the bytes come from as a matter of study design”, not “did you touch an archive”.

So the report also sweeps five free-text fields — temporal.source, tools[].name, otherToolsMentioned[].name, population[].sourceList, classification[].resourceName — for archive names, folds them, and takes the union per paper. That gives 122. It is a ceiling: it counts a paper that used the Wayback Machine once to check a dead link the same as one built on it.

• 59 of the 68 enum papers are also caught by the name sweep. The 9 that are not gave a source too vague to match (e.g. an unnamed “archived version”), or named a non-web archive (Debian snapshot, APKPure) that the fold excludes.
• 63 papers are named-but-not-enum: the enum's undercount factor is 1.8×.

The content page gives both, says which is which, and uses 122 for the headline share because the alternative is publishing a number known to be low.

Free-text archive names are the exact case the corpus README warns about: ~20% run-to-run agreement on exact strings. Measured here on 2026-08-13, the Internet Archive alone appears under 27 distinct spellings, and counting exact strings would have split one archive into 27 and reported the largest as “Wayback Machine, 35 papers” instead of 82.

scripts/arch_fold.mjs is a first-match-wins ordered rule list. Three decisions a reasonable person would have made differently:

1. Access APIs and client libraries fold into the archive. Wayback CDX Server API, Wayback Availability API, Wayback Machine Metadata API, WaybackPy, Waybackpack, WayBack Machine Scraper all become Internet Archive (Wayback Machine). A paper naming only the CDX API is a paper using the Internet Archive. Keeping them separate would have produced a spurious “six archives” tail.
2. Archiving software is a separate family from archive services. Heritrix, pywb, Webrecorder and warcprox form their own row (6 papers) and are excluded from the 122. A paper replaying its own WARCs through pywb is not using a public archive, and merging the two would have inflated archive use with self-archiving plumbing.
3. CC-News is separated from Common Crawl. CC-News is a Common Crawl product, but in this corpus it is used as an NLP training corpus, not as an archive of pages. It gets its own family (3 papers) and is excluded from the 122; names that pair it with other NLP datasets (“The Pile and CC-News”, “WikiText-103, XSum, Amazon Reviews, CC-News, and Reddit”) are excluded outright. Folding it into Common Crawl would have put LLM-training papers into an archives page's headline count.

Unmapped residue: 0 names. Every string that matched the candidate regex was either folded into a family or explicitly excluded.

Excluded as not-a-web-page-archive: 21 distinct names, printed here in full so the exclusions are auditable rather than invisible:

    4  CleanMX
    3  APKPure
    1  Spamtrap feed and CleanMX feed
    1  Google Play Store and six online sources: Androidapps, Apkhere, Apkmirror, Apkpure, Uptodown, and Aptoide
    1  archive.org/details/stackexchange
    1  4plebs
    1  Debian snapshot
    1  warc-clueweb2
    1  Google Play, apkpure.com, and apktada.com
    1  apps hosted by several third party app markets including apkpure, an unofficial mirror of Google Play and wandoujia, a major app market in Mainland China
    1  Internet Archive Metadata API
    1  SiteInspire
    1  Google Play, third-party websites (APKPure and 360 Store)
    1  archive.org 1% sampled tweet stream
    1  Google Play store based on the rank of APKPure
    1  CleanMX feed
    1  The Pile and CC-News
    1  Cosmopedia, WikiText-103, XSum, Amazon Reviews, CC-News, and Reddit
    1  WikiText-103, XSum, Amazon Reviews, CC-News, and Reddit
    1  Internet Archive Stack Exchange data dump
    1  Enron, Patient, CC-News, and IAPR-TC-12 datasets

One of those is a judgement call worth flagging: Internet Archive Metadata API is excluded because in that paper it retrieves item metadata from archive.org's general collections, not web-page captures — but the same API is what [2Hantke, Florian; Calzavara, Stefano; Wilhelm, Moritz; Rabitti, Alvise; Stock, Ben (2023): "You Call This Archaeology? Evaluating Web Archives for Reproducible Web Security Measurements", in: Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, pp. 3168–3182. Association for Computing Machinery, New York, NY, USA. (DOI) (Link)] uses to resolve Wayback contributors. The exclusion is right for this corpus and would be wrong in a different one.

The full spelling-by-family listing is in the script output in §12.

Six statements on the previous revision were wrong, stale, or correct-but-unsafe-to-generalise. Each was checked against a primary source before being changed.

The prior page's claim that “at the time of archiving, crawlers typically read and store web resources but do not execute them” was correct and is retained — the extension quantifies it rather than replacing it.

node scripts/quote_check.mjs –temporal '…' '' over the archive-related temporal tuples: 118 quotes checked, 62 exact, 35 partial (≥60% of five-word windows), 21 below threshold, 0 with no full text on disk. Three of the below-threshold cases were then read by hand against ''paper.cols.txt:

Figures quoted on the content page from individual papers were read in their own full text, not taken from the extraction:

• detection.prevalence for archive papers generally. The field is populated and useful, but most of the values are results about the web obtained via an archive (CSP adoption, header-bidding latency, tracker counts), not results about archives. Only the four that measure the instrument itself — Lerner's escape rate, Rewriting History's vulnerability rate, Weinberg's coverage, Singh's coverage — are on the page. Publishing the rest would have made an archives page into a summary of what archive papers happened to find.
• A per-year series ending in 2026. The bucketed table is used instead, with 2025–2026 labelled provisional in the table itself. The unbucketed per-year numbers are in §11 for anyone who wants to cut it differently; they are not on the content page, because a 2026 point that looks like a rise is an artefact of the two venues that have not happened yet.
• classification.target for archive papers. Computed (§11, section 6) and interesting — 35.7% classify website categories — but it describes what these particular papers were about rather than anything a reader should do, so it stayed in the audit trail.

Every external claim on the content page, how it was verified, and what was rejected. All checks 2026-08-13.

Rejected sources. Two SEO-oriented pages (maxintel.org, archivarix.com) returned by search assert a specific Wayback read limit of “60 requests per minute”, an escalating firewall ban, and a causal story tying throttling to the October 2024 breach. None of it is on the content page. No primary source states any of it; the Internet Archive publishes no read-side limit; and the one peer-reviewed source that looked at this [2Hantke, Florian; Calzavara, Stefano; Wilhelm, Moritz; Rabitti, Alvise; Stock, Ben (2023): "You Call This Archaeology? Evaluating Web Archives for Reproducible Web Security Measurements", in: Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, pp. 3168–3182. Association for Computing Machinery, New York, NY, USA. (DOI) (Link)] explicitly declined to probe the limit on ethical grounds. A plausible-sounding number with no primary source is exactly the failure this site is meant to avoid.

Not fetched, deliberately: a fresh HTTP Archive crawl size. The report page's sample-size series is only reachable through a rendered chart, and har.fyi states “over 1 million pages” which is inconsistent with the ~13M the prior page quotes for February 2025. The page therefore keeps its dated February 2025 example and tells readers to check the live report rather than trust either number. See §10.

• The derivation of the “5% crawled with a browser” figure. [3Zhu, Jingyuan; Sun, Huanchen; Madhyastha, Harsha V. (2025): "Toward Better Efficiency vs. Fidelity Tradeoffs in Web Archives", in: Proceedings of the ACM Internet Measurement Conference, pp. 1025-1031. (DOI)] states it in the introduction and attributes it to their Appendix B; the appendix is not present in the extracted full text. The content page attributes it as the authors' estimate, which is what it is, and does not restate the method.
• Current HTTP Archive crawl size. See §9. Two public sources disagree by an order of magnitude and neither is queryable without BigQuery credentials.
• Whether archive use is really flat, or whether the enum's drift makes it look flat. The trend table uses the name sweep, which is uniform across years, so the shape should be sound — but the 68/122 gap is itself a property of how the extractor read papers, and nobody has checked whether that reading changed over the corpus's sixteen years.
• Whether the field's dependence on one archive matters. Raised as an open question on the content page. Nothing in the corpus can answer it: it needs a study that runs the same measurement through two independent archives, and the only viable second archive was rejected on coverage rather than validated on agreement.
• Four duplicate-entry pairs exist in bibliography, all predating this run: lerner2016internet/lerner2016_internet (USENIX Sec 2016), bouhoula2024automated/bouhoula2024_automated (USENIX Sec 2024), fouad2022my/fouad2022_cookie (PETS 2022), and bottger2025_regional/boettger2025_regional (PETS 2025, differing only in umlaut transliteration). The last three were found by the citations reviewer, not by us. None were fixed here: both keys in each pair resolve, nothing breaks, and silently deleting a key another page may cite is a worse failure than an untidy bibliography. Flagged for whoever next edits that file.

Four reviewers, all told explicitly that the author's context might not be exhaustive, all handed the page text, both scripts, the script output and these notes. The three focused ones ran in parallel first; their findings were acted on before the generic pass, which then ran over the corrected pages. 30 findings in total, 29 accepted, 1 rejected. Pass D — the one with no checklist — found the single worst defect, and found it by reading code the other three had already been given.

Re-ran report_archives.mjs, diffed against the committed output, and re-derived every paper-sourced figure from the full texts.

Everything else the reviewer checked — the full Hantke Tables 1–4, the Zhu, Lerner 2016, Lerner 2017, Weinberg and Singh figures, and a byte-for-byte re-run of the script — matched.

The reviewer independently re-verified and confirmed: the Memento shutdown text, RFC 7089's status, the one-trillion-pages announcement, the Save Page Now notice, the Availability/CDX/replay/id_ endpoints, httparchive.crawl.* table names, har.fyi's BigQuery procedure, Arquivo.pt's operator and figures, Common Crawl's latest index, pywb v2.9.1, browsertrix-crawler v1.14.2, Browsertrix, warcio, cispa/webrec, the WARC 1.1 spec, Wikipedia's list, the wayback library defaults, and that HTTP Archive still runs WebPageTest.

Ran last, over both pages, the scripts and the previous published revision. Fifteen findings; the four most serious are recorded here in full because they are the ones the focused passes were structurally unable to catch.

The other eleven, all accepted: the go/no-go passage was buried at the end of a how-to section and is now its own section near the top with a two-column table; “roughly two thirds of endpoints” did not follow from 13 of 64 and was replaced with the paper's own counts; the stabilisation counts had no denominator (11,364 domains, added); Lerner's 16.1% escape rate is from 2016 and is now dated as such; the page demanded escape-blocking without saying how (one interception rule, now given, plus the instruction to report the count); “a snapshot is static so a later reader can fetch the same one” ignored retroactive exclusions (now qualified); “Recording Your Own Archive” claimed “no escapes”, which is a property of the replayer not the WARC (now “escapes you can block because you control the replayer”); the wayback library was buried in a rate-limit footnote and is now in the access-paths list; the 55%/7% coverage figure was stated twice verbatim (deduplicated); Methodology and limitations was structurally a subsection of What to Report and has moved back under Use in Publications; and the page had neither the What to Read First nor the Related Pages section its neighbours carry (both added). One nit was also taken: the REPLAY_ARCHIVES parenthetical under-described its own definition.

Rejected from pass D: nothing. All fifteen were accepted.

scripts/report_archives.mjs, run 2026-08-13 against data/extract/run1 (5,859 papers), after the three bug fixes from review passes A and D. Nothing below is edited; the full spelling-by-family listing that §6 refers to is the block headed Raw spellings folded into each family, and the EXCLUDED block is the same residue printed in §6.

========================================================================
POPULATIONS
========================================================================
corpus                                              5859
empirical (measured the world as it is)             5118
crawled   (ran an automated web crawl)              1120
with >=1 temporal (data-provenance) tuple           5342

========================================================================
1. DATA PROVENANCE (temporal.mode), of papers with >=1 temporal tuple
========================================================================
Multi-valued: one paper can both crawl live and consult an archive, so the
column does not sum to 100%. not-stated is a value, never an answer.

temporal.mode       papers  share of 5342
------------------  ------  -------------
existing-dataset    2534    47.4%
active-probing      1780    33.3%
live-crawl          1261    23.6%
passive-collection  1012    18.9%
not-stated          353     6.6%
web-archive         68      1.3%

mode=web-archive                68 papers
mode=live-crawl                 1261 papers
ratio live-crawl : web-archive  18.5 : 1
(shares of `empirical` are computed further down, on the intersected sets)

========================================================================
2. NAMED A WEB ARCHIVE, folded (5 free-text fields, union per paper)
========================================================================
papers naming a web-page archive service            122
  ... of which a REPLAY archive (IA, Arquivo, ...)  85
  ... i.e. Common Crawl / HTTP Archive only         37

SHARES OF EMPIRICAL — numerator intersected with `empirical`, never the raw count:
  named an archive AND empirical                    115   2.2%  (of 5118)
  of the 122, NOT empirical                         7
  mode=web-archive AND empirical                    65   1.3%
  mode=live-crawl AND empirical                     1236   24.2%

The 7 papers in the 122 that are not empirical (mostly LLM-training uses of Common Crawl):
  2014 USENIX   iSeeYou: Disabling the MacBook Webcam Indicator LED
  2014 USENIX   Automatically Detecting Vulnerable Websites Before They Turn Malicio
  2021 USENIX   Extracting Training Data from Large Language Models
  2022 WWW      WebFormer: The Web-page Transformer for Structure Information Extrac
  2025 USENIX   Private Investigator: Extracting Personally Identifiable Information
  2026 NDSS     Was My Data Used for Training? Membership Inference in Open-Source L
  2026 WWW      MCP vs RAG vs NLWeb vs HTML: A Comparison of the Effectiveness and E
papers with temporal.mode = web-archive             68
  ... of those, also caught by the name sweep       59
papers the mode enum MISSES (named, mode != archive) 63
undercount factor of the enum                       1.8x

Folded families (paper-counted; multi-valued, so shares do not sum).
IMPORTANT: two families are NOT part of the 122 — self-hosted archiving
software, and CC-News used as an NLP corpus. Their shares would be shares of
a population they are excluded from, so the share column is blank for them
and their papers are counted separately.

archive family                                              papers  share of 122  note                                 raw spellings
----------------------------------------------------------  ------  ------------  -----------------------------------  -------------
Internet Archive (Wayback Machine)                          82      67.2%                                              27
Common Crawl                                                28      23.0%                                              12
HTTP Archive                                                16      13.1%                                              5
archiving software (Heritrix, pywb, ...)  [not in the 122]  6       —             2 of these also name a real archive  4
unspecified web archive                                     4       3.3%                                               6
CC-News (text corpus from Common Crawl)  [not in the 122]   3       —             2 of these also name a real archive  5
Memento / Time Travel                                       1       0.8%                                               1

Raw spellings folded into each family — this is what folding bought:
  Internet Archive (Wayback Machine)  (27)
      Archive.org
      Archive.org historical snapshots
      Internet Archive
      Internet Archive / Wayback Machine
      Internet Archive Wayback Machine
      Internet Archive and Google page snapshots
      Internet Archive and other web archives
      Internet Archive's Wayback Machine
      Internet Archive: Wayback Machine
      Internet Wayback Machine
      WayBack Machine Scraper
      Wayback Availability API
      Wayback CDX API
      Wayback CDX Server API
      Wayback Machine
      Wayback Machine / Internet Archive
      Wayback Machine API
      Wayback Machine Availability JSON API
      Wayback Machine archives
      WaybackMachine CDX API
      WaybackPy
      Waybackpack
      archive.org
      authors' archives, community-shared snapshots, and Internet Archive
      legal documents containing web.archive.org URLs
      live web and Internet Archive
      public firmware archives, official update channels, and Wayback Machine
  Common Crawl  (12)
      Citizen Lab lists, full Tranco list, and Common Crawl Project
      Common Crawl
      Common Crawl CC-MAIN-2021-31
      Common Crawl Project
      Common Crawl and Tranco
      Common Crawl dataset
      Common Crawl email-pattern subset
      Common Crawl's Web Crawl Data
      CommonCrawl
      ICANN Centralized Zone Data Service and full Common Crawl host list
      TLD zone files, Citizen Lab test lists, Tranco list, and Common Crawl project
      subset of the Common Crawl dataset
  unspecified web archive  (6)
      Own weekly crawls and Web Archive
      SWIPE web archive
      Web Archive
      Web Archive API
      custom domain pool from historical snapshots
      web archive snapshots and search engine indexes
  HTTP Archive  (5)
      HTTP Archive
      HTTP Archive 1.2
      HTTP Archive and device data
      HTTP Archive dataset
      HTTPArchive
  CC-News (text corpus from Common Crawl)  (5)
      CC-News
      Common Crawl CC-News
      Common Crawl News
      Common Crawl News (CC-News)
      Common Crawl News (English articles after language filtering)
  archiving software (Heritrix, pywb, ...)  (4)
      Heritrix
      Webrecorder
      pywb
      warcprox
  Memento / Time Travel  (1)
      Memento API

========================================================================
2b. RESIDUE — printed so the part that could not be classified stays visible
========================================================================
UNMAPPED raw names (matched the candidate regex, no fold rule): 0

EXCLUDED as not-a-web-page-archive: 21 distinct names
    4  CleanMX
    3  APKPure
    1  Spamtrap feed and CleanMX feed
    1  Google Play Store and six online sources: Androidapps, Apkhere, Apkmirror, Apkpure, Uptodown, and Aptoide
    1  archive.org/details/stackexchange
    1  4plebs
    1  Debian snapshot
    1  warc-clueweb2
    1  Google Play, apkpure.com, and apktada.com
    1  apps hosted by several third party app markets including apkpure, an unofficial mirror of Google Play and wandoujia, a major app market in Mainland China
    1  Internet Archive Metadata API
    1  SiteInspire
    1  Google Play, third-party websites (APKPure and 360 Store)
    1  archive.org 1% sampled tweet stream
    1  Google Play store based on the rank of APKPure
    1  CleanMX feed
    1  The Pile and CC-News
    1  Cosmopedia, WikiText-103, XSum, Amazon Reviews, CC-News, and Reddit
    1  WikiText-103, XSum, Amazon Reviews, CC-News, and Reddit
    1  Internet Archive Stack Exchange data dump
    1  Enron, Patient, CC-News, and IAPR-TC-12 datasets

========================================================================
3. TREND — EMPIRICAL papers naming a web-page archive, by year bucket
========================================================================
Denominator is empirical papers in the same bucket. 2025-2026 is starred:
CCS/IMC 2026 have not been held and IEEE S&P/WWW 2026 selection is partial,
so that bucket is under-represented by construction, not by relevance.

years       named an archive  empirical papers  share  of which mode=web-archive
----------  ----------------  ----------------  -----  -------------------------
2010–2013   1                 460               0.2%   0
2014–2017   14                718               1.9%   11
2018–2021   39                1272              3.1%   19
2022–2024   36                1649              2.2%   22
2025–2026*  25                1019              2.5%   13

Per year, unbucketed (for anyone who wants to cut the series differently):

year  named an archive  empirical  share
----  ----------------  ---------  -----
2010  1                 102        1.0%
2011  0                 103        0.0%
2012  0                 137        0.0%
2013  0                 118        0.0%
2014  3                 152        2.0%
2015  0                 176        0.0%
2016  5                 171        2.9%
2017  6                 219        2.7%
2018  8                 228        3.5%
2019  14                359        3.9%
2020  6                 342        1.8%
2021  11                343        3.2%
2022  12                471        2.5%
2023  14                601        2.3%
2024  10                577        1.7%
2025  16                675        2.4%
2026  9                 344        2.6%

========================================================================
3b. BY VENUE — EMPIRICAL papers naming a web-page archive
========================================================================
venue    named an archive  empirical  share
-------  ----------------  ---------  -----
USENIX   23                1226       1.9%
IMC      21                625        3.4%
WWW      18                740        2.4%
PETS     17                477        3.6%
NDSS     16                586        2.7%
CCS      14                818        1.7%
IEEE-SP  6                 646        0.9%

========================================================================
4. WHAT ARCHIVE PAPERS DO AND DO NOT STATE
========================================================================
Population: the 122 papers that named a web-page archive.
Every row is a share of that population. A sentinel is not an answer.

papers with an archive-attributed temporal tuple    94

field stated on the archive tuple  papers  share of 94
---------------------------------  ------  -----------
spanStart (first archival date)    71      75.5%
spanEnd (last archival date)       71      75.5%
snapshots (how many were fetched)  30      31.9%
cadence (how often / how chosen)   33      35.1%

For comparison, the same fields on live-crawl tuples:

field stated on the live-crawl tuple  papers  share of 1261
------------------------------------  ------  -------------
spanStart                             796     63.1%
spanEnd                               791     62.7%
snapshots                             386     30.6%
cadence                               316     25.1%

========================================================================
5. ARCHIVE INSTEAD OF, OR ALONGSIDE, A LIVE CRAWL
========================================================================
Population: the 122 papers that named a web-page archive.

                                       papers  share of 122
-------------------------------------  ------  ------------
also ran a live crawl (temporal.mode)  71      58.2%
is in the `crawled` population         82      67.2%

sole-source reading                                             papers  share of 122
--------------------------------------------------------------  ------  ------------
strict: every temporal tuple is mode=web-archive                8       6.6%
loose: collected nothing itself (no crawl/probe/passive tuple)  31      25.4%

Papers in the loose count but not the strict one (the enum-coding gap):
  2016 CCS      PREDATOR: Proactive Recognition and Elimination of Domain Abuse at Tim
  2016 USENIX   Measuring PUP Prevalence and PUP Distribution through Pay-Per-Install 
  2017 CCS      Rewriting History: Changing the Archived Web from the Present.
  2018 PETS     An Empirical Analysis of Traceability in the Monero Blockchain
  2019 IEEE-SP  Does Certificate Transparency Break the Web? Measuring Adoption and Er
  2019 IMC      Errors, Misunderstandings, and Attacks: Analyzing the Crowdsourcing Pr
  2020 NDSS     A Practical Approach for Taking Down Avalanche Botnets Under Real-Worl
  2020 WWW      An Empirical Study of the Use of Integrity Verification Mechanisms for
  2021 IMC      Risky BIZness: risks derived from registrar name management.
  2021 USENIX   Extracting Training Data from Large Language Models
  2022 IMC      HTML violations and where to find them: a longitudinal analysis of spe
  2022 WWW      "Way back then": A Data-driven View of 25+ years of Web Evolution.
  (23 total)

========================================================================
6. WHAT THESE PAPERS MEASURE — classification.target (enum), archive papers
========================================================================
Population: the 122 papers that named a web-page archive,
restricted to those that classified anything. Enum, so publishable.

archive papers that classified something            115

classification.target  papers  share of 115
---------------------  ------  ------------
other                  61      53.0%
website-category       41      35.7%
web-request            26      22.6%
domain                 26      22.6%
vulnerability          18      15.7%
privacy-policy         10      8.7%
user-generated-text    7       6.1%
ip-address             7       6.1%
malware                6       5.2%
mobile-app             5       4.3%
website-popularity     4       3.5%
cookie                 4       3.5%

Mistakes made and caught in review, recorded because the clean version of this page would be worth less:

1. A numerator over the wrong denominator (D1) — the exact error this site's rules exist to prevent, in the page's own headline figure. It survived a review pass whose entire remit was checking figures against the script, because the script was wrong in the same direction. data/extract/OVERVIEW.md already publishes the correct 65-of-5,118 for web-archive; comparing against it would have caught this in the first hour.
2. A fabricated quotation written into a footnote (B3). It came from a search-result summary that blended two sources, and was never checked against the page it was attributed to. Every other footnoted quote on both pages was fetched and read; this one was not, and it is the only failure of that discipline in the run.
3. A factually wrong claim of absence — “no maintained replacement aggregator” (C1). DNS failure for timetravel.mementoweb.org plus a shutdown notice on mementoweb.org were treated as sufficient. They established that Time Travel is gone; they said nothing about what else exists. Absence claims need their own search, and this one did not get one.
4. An internal contradiction between the lead paragraph and a table four screens below it (A1), from computing a complement (100 − 6.6) and describing it as something else.
5. A contradiction inside this provenance page (D2): a figure corrected on the content page was left in its wrong form in the very table meant to let a reader check it.
6. A share of a population its members are excluded from (A3), and a filter that was right by accident (A6), both in code that had already been run and whose output had already been read.

Every one of these was found by a reviewer, none by the author. The pattern worth carrying forward: the two most serious defects (D1, C1) were not the kind a checklist finds — one was a category error inside code everyone had been handed, the other was a claim that something does not exist. Both needed a reader who was not told what to look for.