roadmap
This is an old revision of the document!
Table of Contents
Roadmap
What this wiki has decided to write next, and what it has decided not to write. Both halves are here on purpose: on a site where every page names its own population, “we looked and there is not enough literature yet” is a finding, and it should be recorded once rather than rediscovered by every gap pass.
Paper counts below are candidate sets from the publication corpus — title-and-summary probes over 5,859 extracted papers, so they are floors, not populations. The page that gets written derives its own.
Queued
Nothing is queued: the queue is empty. As of 2026-09-16 every page scoped here has been written, and the last one — Data subject rights — moved to Assessed below. scripts/sitemap.mjs gates on this table: every promised-but-missing page on the wiki must appear here, a red link anywhere else is a defect, and a row here that has turned blue must be removed. The script accepts an empty queue only when this section says so in words, so that a table-parser regression still fails the run rather than passing vacuously.
When a page is queued again, add a ^ Planned page ^ Why, and the evidence ^ table here and delete the first sentence above.
Assessed, and not written
Recorded so the next gap pass does not re-propose them. Counts are candidate sets from the corpus, not populations.
| Proposal | Disposition |
|---|---|
| Data subject rights | Written 2026-09-16 as Privacy:Data subject rights. Every number in the queued row was wrong in the same direction. The 9 title-and-summary candidates all survive — 100% precision, the only queued probe on this wiki to manage that — but the recall is 17.6%: the page's population is 51 papers and the title probe misses 42 of them, including every paper that obtained its research data through the right of access, both people-search removal studies, and the largest opt-out-signal paper before 2024. On this topic the rights work is usually one section of a paper whose title says nothing about rights. The “6 of them 2024–2026” understates the slope rather than overstating it: 43 of the 51 are 2021 or later. The population is a hand audit of 471 full-text candidates (10.8% precision) against a written inclusion rule; nine tenths of the noise is structured, and the largest single class is an internet-scanning paper's ethics section offering an “opt-out request” for its own probes. Two probes were added mid-run, after the first complete audit could not see Do Opt-Outs Really Opt Me Out? (CCS 2022) — a paper the page is partly about — because industry opt-out tools are never called “do not sell”; they added 182 candidates and 10 population papers. The scope held in one important respect and moved in another: the request-as-instrument framing is the page's spine, and the median number of controllers a request-sending paper wrote to is 30, against a crawl median of 10,000 sites. Probe history, all 471 verdicts, the fold residue and eight dated primary sources: data_subject_rights. |
| Age assurance | Written 2026-09-15 as Age assurance. The queued reasoning held — the page is thin, the thinness is the finding, and there is no figure table — and writing it corrected this row's own count twice. The 11 title-and-summary candidates are really 10: gap_probe_roadmap.mjs has no word boundary before age, so F-BLEAU: Fast Black-Box Leakage Estimation entered on Leakage Estimation. And 11 was the wrong instrument as well as the wrong number — the title probe has 27.3% precision (3 of 11 are in the derived population) and 60% recall (3 of the page's 5), because on this topic the measurement is often one section of a paper whose title says nothing about age: A Picture is Worth 500 Labels and Developers Say the Darnedest Things are both invisible to it. A full-text phrase probe over the 5,855 papers with full text returns 38 candidates, every one hand-judged: 1 has age assurance as its research question, 4 more measure it in a section, and 6 ran into it as an obstacle to a measurement about something else — which is now the most common way it appears in the corpus and the thing the queued row did not anticipate. The population is 5 papers, 13.2% of the candidates and 0.1% of the corpus, and only one of the five is about the web. The children's-privacy reading held and is sized: 36 papers name COPPA in legal[] against the 5, and exactly 1 paper is in both. Probe history, the two regex defects, verdicts for all 38 and nine dated primary regulatory sources: age_assurance. |
| Connected TV | Written 2026-09-12 as Connected TV. The queued reasoning held and both of its numbers moved. The 16 title-and-summary candidates have 62.5% precision — 10 are in the page's population — but the recall is the real story: the derived population is 35 papers and the queued probe misses 25 of them, including the corpus's largest TV app analysis (4,745 Android TV APKs, PETS 2022) and a Wi-Fi-positioning paper in which Roku devices are two of the five most common vendor prefixes in a 490-million-row dataset. The “only one candidate is on the web platform” observation survives derivation exactly: 1 of 35 carries web, 26 carry iot, and only 6 are in the corpus-wide crawled population. One paper the row named as spine is not in the population — Endangered Privacy (USENIX 2025) identifies videos from encrypted MPEG-DASH traffic and never measures a television. The page's population is a hand audit of 103 candidates from nine full-text probes (34.0% precision), with a written inclusion rule, and the biggest noise class is the smart-home survey where “smart TV” is an answer option. Audit, verdict map for all 103, and the 191-name fold residue: connected_tv. |
| Anti-adblock and adblock walls | Written 2026-09-16 as a section on Filter Lists, The site may detect the blocker. The disposition held; the count did not. 3 was a probe defect: the anti_adblock family in gap_probe_roadmap.mjs matches anti-?ad, which allows “anti-ad” and “antiad” but not “anti ad”, so it cannot see Detecting Anti Ad-blockers in the Wild (PoPETs 2017) — the paper whose entire subject is this row's. Widening the separator class gives 4 at title level and 64 in full text, of which only 8 mention the topic ten or more times and 5 more than fifty; two of those five are 2025 Acceptable Ads papers, which is the allowlisting side and not this one. The section's load-bearing figures are both 2017 and both published side by side and deliberately not as a range: anti-adblock filter lists trigger on about 9% of Alexa top-5K, while a differential crawl finds visible reactions on 686 of Alexa top-100K (0.7%) — two definitions of “detects my blocker” on populations twenty times apart. The most recent dedicated paper in this corpus is 2021, and nobody in these seven venues has re-measured wall prevalence since Alexa was retired; that negative claim was re-run at a second, disjoint vocabulary before publishing and survived. Probe widths, the paired-crawl candidate set with its five hand-marked false positives, and the quote checks against both PDF renderings: filter_lists. |
AI crawlers, robots.txt, training opt-out | Written 2026-09-16 as a section on Crawler Detection, Measuring who crawls you. The disposition held and the corpus is even thinner than the row said: three papers name an AI-crawler user agent, all 2025, not four. The page's own earlier sentence said five, from a wider pattern that also matches “AI scraper” and “agentic browser” — two of those five are papers about LLM-driven agents, one matching only inside its reference list, and the sentence was corrected in the same sitting. 74 papers mention robots.txt at all, but 44 mention it exactly once, as a crawl-policy declaration or an archive-coverage caveat; no paper mentions llms.txt. The section is built on the separation the row implied but did not state: a directive is a claim measurable at web scale from the client side, compliance is measurable only from the server's own logs, so the studiable population is the set of sites you operate. The two IMC 2025 papers disagree — vendor training crawlers largely comply, user-triggered assistant fetchers largely never fetch the file — and the section publishes the disagreement rather than choosing. Probe widths, the coding of the 44 with its uncoded residue, and the quote checks: crawler_detection. |
| WebRTC leaks | 6 papers. Below the line — and the line matters, or WebAssembly and every other thinly-published web technology follows. |
| SSO, WebAuthn and MFA deployment | Written 2026-09-11 as Authentication, after the audit this row demanded. The gate was passed but not comfortably by the row's own terms: of the queued 46, only 22 measure deployment — the rest are attacks, SDK and protocol analysis, mobile studies, user studies, abuse ecosystems, and three homonyms (MFA as “Made for AdSense”, OpenID inside “OpenIDEO”, an RPKI relying party). Four more probes and a full-text recall pass take the candidate pool to 184, of which 45 are deployment measurements — 24.5% precision — and the branded-mechanism slice this row described is 25, four of which the queued probe could not see. The page was broadened from SSO/WebAuthn/MFA to the whole login surface, because the login-page and login-policy papers are where the denominators come from: “how many sites support SSO” is 6.30%, 7.23%, 27% and 57.8% in five papers that divide by different things. The audit is on authentication. |
| Sock-puppet and personalization audits | 20 papers, the smallest live proposal. Likely a section on Stateful Stateless rather than a page. |
| Per-company platform pages (Facebook, Twitter, TikTok, Amazon) | Assessed against the corpus and deliberately not written; Platforms carries the reasoning and the counts. |
| Blocking and geodifference | Written 2026-09-10 as Blocking and geodifference — a split from Crawling location, which is about choosing a vantage point rather than about the phenomenon; two bullets on that page were rewritten to hand the phenomenon over. The queued “74 papers” is a candidate set and turned out not to be comparable to the page's population in either direction. The recorded alternation re-runs to 68 here, of which 38 are in the population and 30 are not; and 16 of the population's 63 papers are caught by no title probe at all. The page's population is the 63 papers surviving a hand audit of 274 candidates from eight probes — 58.8% of the candidates are off topic, mostly on ordinary English words. Three of the eight probes were added mid-run, each because a later check found a paper the candidate set had missed; the last was prompted by the bibliography's own newest entry, an IMC 2025 outage study that no censorship-shaped probe could see. |
| Internet scanning | Written 2026-09-10 as Internet scanning — an instrument page under programming:, matching Crawler, as decided here on 2026-09-07. Writing it corrected the queued count twice. The 143 that named ZMap/ZGrab/masscan/nmap re-derives to 141 inside the 930 (145 corpus-wide) once the fold is written down; and the page's own population is not that set but the 245 papers naming any active-scan instrument, of which 235 are inside the 930. The queued “only 5 papers mention a hitlist” was a tools[] fold, not a full-text count: 62 SCAN papers say “hitlist” and 39 say it near “IPv6”, while 11 name an IPv6-targeting instrument. Corrected 2026-09-17: a hand audit of the fold's 174 hand verdicts moved the page's population to 241 (232 inside the 930) and the four-scanner set to 139 / 143; the figures in this row are the 2026-09-10 ones. See Follow-up run, 2026-09-17. |
| Email authentication | Written 2026-09-09 as Email authentication. The queued population held: the INFRA verdict of scripts/msg_fold.mjs is still exactly 31 papers and was reused rather than re-derived. A widened full-text recall probe over all 5,859 papers added none. The page became a sixth child of Security. |
| Annotation and validation | Written 2026-09-11 as Annotation. Both queued counts moved. The 42.4% is the share of papers with at least one unvalidated classifier, not the share that validated nothing — that is 29.9% (1,328 of 4,439), and the Queued row this replaces stated the former in words that read as the latter. And the 30-paper LLM-annotator probe is a candidate set with 70% precision and 13% recall: seven of the 30 study attacks on LLMs rather than using one, and the classification.method == “llm” enum finds 175 papers the probe misses all but 23 of. The page's populations are the enum and the 4,439. |
| Ad archives | Written 2026-09-08 as Ad archives. Writing it corrected the count that queued it: the 69 was a homonym artefact (an Android ad library is an advertising SDK), and the hand-audited figure is 11 papers. |
| Ownership resolution | Written 2026-09-11 as Ownership resolution — a carve-out, not a new population: the comparison, the coverage census and the random-sample accuracy rates existed on webXray and moved unchanged, with their two frame dates intact, because re-deriving them would have orphaned the adjudication pinned to 2026-08-17. The queued reasoning held and the 2026-08-17 sitting's decision against a separate page was reversed on the record; webXray is now a 27 KB page about webXray itself rather than the stub that decision feared. The move also corrected a denominator that had been live for three and a half weeks: report_webxray.mjs divided a corpus-wide sweep count by the crawling population, so “12.1% of papers that crawled name an ownership resource” was 136 ÷ 1,120 where the comparable figure is 9.6% (107 ÷ 1,120), and webXray's own 1.3% is 1.2%. No guard could have caught it: the page matched the script and the script was wrong. New corpus figures: 136 papers name an ownership resource, 9.6% of the 1,120; PETS carries 9.6% of its own papers against NDSS's 1.1%; and the share has been flat at 2.5–2.7% since 2023 rather than growing. |
| Privacy policies and terms | Written 2026-09-10 as Policies. The queued count moved: 121 was a four-alternative title probe, and widening it to nine (privacy label, data safety, nutrition label, privacy statement — the store-declaration half of the literature) gives a 123-paper candidate set. The page's actual population is the 102 papers whose classification[].target enum is privacy-policy; the 123 is used for rankings only. The queued scope also held on the tool lineage, and reading the repositories corrected it: PurPliance (2024-03) is more recently touched than PoliGraph (2023-06). |
Adding to this page
A proposal earns a row in Queued when it has an id, a population that can be re-derived from the corpus, and a reason it is not a section on an existing page. It earns a row in Assessed when that check fails — with the count that made it fail. Both are recorded in roadmap when a decision changes.
roadmap.1789612696.txt.gz · Last modified: by karel.kubicek.claude