Differences

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

--- provenance:design:algorithm_audits [2026/09/11 11:02] – Reviewer findings from four passes; correct the probe-2 residue (146->158) and the probe-7 narrative; add the recall-repair sweeps and the shared-denominator judgement call. Authored by Claude karel.kubicek.claude
+++ provenance:design:algorithm_audits [2026/09/11 11:21] (current) – Generic review log (9 findings, all accepted); the null adjudication and its definition; verdicts for all 58 shortlisted papers; residue re-derived (157/391). Authored by Claude karel.kubicek.claude
@@ Line 11: / Line 11: @@
 ^ Question ^ Answer ^
 | Does the wiki already cover it? | No. [[Programming:Stateful stateless]] owns the profile axis and lists personalisation as a phenomenon that needs a stateful design; [[Statistics:Hypothesis testing]] owns the test; [[Design:Platforms]] owns sock puppets as an **access route** and says in its own Open Questions that //"Search-engine and ads-ecosystem auditing has no page on this wiki"//. Nobody owned the experimental design. |
-| How large is the in-corpus population? | **32** papers under a written inclusion rule (below), of 5,859. Twelve further candidates were read and rejected. |
+| How large is the in-corpus population? | **34** papers under a written inclusion rule (below), of 5,859. **36** further candidates were read and rejected. The figure was 32 until a generic review pass found that 26 shortlisted papers had been read and dropped without a written verdict, and that one of them satisfied the rule. |
-| Is it growing or historical? | Growing: 11 papers in 2020–2023 and 10 in 2024–2026*, against 6 in 2010–2015. |
+| Is it growing or historical? | Not historical. Per year: 1.00 (2010–2015), 1.25 (2016–2019), 3.00 (2020–2023), 3.67 (2024–2026*). //n// = 34 will not carry a trend claim and the page does not make one; what it says is that more of this work appeared in the last six years than in the first ten. |
 | Could it be a section instead? | It could have been ~3 KB on [[Design:Automated measurements]]. It was not, for two reasons: (a) that page is a **routing** page between crawl / scan / app, and an audit is a fourth instrument that cuts across the crawl branch rather than sitting beside it; (b) the material that makes the page worth writing — control arms, carry-over, the noise floor — is design advice, not routing, and would have doubled the length of a page whose job is to be short. |
 | The counter-argument | The 32 is a **lower bound with a known bias** (see //The screening loss// below), so a reader could reasonably say the page is built on a population the corpus cannot see properly. That is stated on the page itself, in its own box, rather than buried here. |
@@ Line 23: / Line 23: @@
   - **(T) Treatment.** It deliberately varies a property of the measuring identity or request — profile history, declared attribute, location, device, opt-out setting, ad creative — and holds the rest fixed.
-  - **(O) Outcome.** What it measures is the platform's own response: ads served, results ranked, prices quoted, feed or recommendation contents, or an access decision.
+  - **(O) Outcome.** What it measures is the platform's own **discriminating** response to the identity it has built: ads served, results ranked, prices quoted, feed or recommendation contents, or the profile the platform reports back. **Not reachability.** A study whose outcome is whether you can reach the site at all — geoblocking, censorship, Tor-exit refusal — varies a **vantage point** rather than an identity and belongs to [[Design:Blocking and geodifference]]. This clause was tightened on 2026-09-11 after a review pass; it excludes no paper already in the set, and it is what keeps //403 Forbidden: A Global View of CDN Geoblocking// (IMC 2018) out. Without it, every geodifference study in the corpus would be an algorithm audit and the page would duplicate a neighbour.
   - **(C) Contrast.** The result is a difference (or a bounded absence of difference) **between arms**, not a prevalence over a crawl of many sites.
@@ Line 57: / Line 57: @@
 ===== The adjudication =====
-candidates scored ≥3 across the probes were shortlisted with their abstracts (''_aa_short.mjs''). The 58 scoring ≥4, plus 17 lower-scored papers flagged by hand, were passed through ''_aa_adj.mjs'', which prints every sentence in the paper matching an arm/treatment/control/persona pattern. Each verdict was then made by reading those sentences, and where they were ambiguous, by grepping the paper's methods section directly.
+candidates scored ≥3 across the probes were shortlisted with their abstracts (''_aa_short.mjs''). The 58 scoring ≥4 were passed through ''_aa_adj.mjs'', which prints every sentence in the paper matching an arm/treatment/control/persona pattern; a further 12 lower-scored or probe-7 papers were read directly. Each verdict was made by reading those sentences, and where they were ambiguous, by grepping the paper's methods section.
+**A generic review pass found the weak point here.** 26 of the 58 had been read and dropped with no written verdict, and one of them — //Cart-ology// (CCS 2022, shortlist score 7) — satisfies the rule: five browser profiles, four of them unused baselines, //"created and mechanistically measured in the same way, with separation between attacker, victim, and baselines on different machines with different IPs"//, read out as ad distributions. It had been rejected on the strength of ''_aa_adj.mjs'' returning **zero** sentences for it — a 420-character sentence cap meeting a column-spliced PDF. **A probe returning nothing is not a negative result**, and it was treated as one. All 58 now carry a written verdict, and the IMC 2024 filter-bubble poster was promoted at the same time (four bot arms differing only in video-selection strategy). The population moved from 32 to **34**.
 **Every verdict, with the sentence that settled it, is in the script output below** (''--list''). The script refuses to run if any entry lacks an adjudication note.
@@ Line 96: / Line 98: @@
   * **''statistics.method''** is free text and ~20% stable run-to-run. It is folded to an alphanumeric skeleton (lowercase, non-alphanumerics stripped) and **paper-counted**, and it is published as a **ranking**, not as percentages. The fold does not merge synonyms: ''Holm-Bonferroni correction'' (3), ''Holm-Bonferroni'' (1) and ''Holm-Bonferroni method'' (1) are three rows in the raw output. An unfolded reading would publish "Holm–Bonferroni 3"; the true paper count for the Holm–Bonferroni family is **5 of 32**, and for any Bonferroni-family correction **7 of 32**. The report script now prints the hand-folded family counts alongside the raw skeleton ranking, so the page quotes a number the script produced rather than one assembled in prose. Both are in the output below.
-  * **The eight noise-baseline phrasings** are not a fold but a deliberately widened probe (''_aa_noise.mjs''): the narrow term "A/A test" returns 2 papers corpus-wide, so seven further phrasings were added. The count rose from 0 to 11 of 32. Residue is not applicable — the probe reports term presence, and the page states explicitly that a paper can build a null without any of the eight phrasings.
+  * **The eight noise-baseline phrasings** are not a fold but a deliberately widened probe (''_aa_noise.mjs''): the narrow term "A/A test" returns 2 papers corpus-wide, so seven further phrasings were added, taking the count from 0 to 11. **That probe is now published as a //vocabulary// finding only**, not as a practice rate — see the next section.
+**Residue of the candidate probes.** Probe 2 returned **190** candidates. **28** of the final 34 audits and **5** of the 36 rejections lie inside it — the rest were caught by probes 4–7 — leaving a residue of **157** papers dropped at title-and-abstract level without an individual note. The union across all seven probes is **461**, of which **70** were adjudicated in depth, so **391** rest on a title-level read. That is the honest residue of this page.
+===== The null: why it is adjudicated and not counted =====
+The page's central claim is that establishing a null is the skipped step. The first version of this log said that claim rested on //"a term count plus an eight-way concept probe, never as 'nobody does this'"//, and the page then went ahead and wrote **"only 11 of 32 establish a same-treatment baseline at all"** — which is a practice claim, from a term probe, three times over. A generic review pass caught the gap between the caveat and the sentences.
+It was right, and the probe is wrong **in both directions**:
+^ Paper ^ Term that matched ^ What the sentence actually says ^
+| {[robertson2018_auditing]} | ''noise floor'' | //"Hannak et al. found (1) evidence of general personalization above a noise floor"// — a description of somebody else's result. This paper pairs standard and incognito windows; it has no same-treatment pair |
+| {[iqbal2022_left]} | ''same treatment'' | //"whether the SFA of any given email service provided same treatment to similar emails from candidates of different political affiliations"// — the research question, not a baseline |
+| {[oh2022_cartology]} | **none of the eight** | four unused baseline profiles //"created and mechanistically measured in the same way"//, with the divergence of one read as the noise. The clearest A/A in the corpus, invisible to the probe |
+So the question was put to all 34 papers by hand, the same way the population was, with a stated definition and an evidence sentence per verdict, in ''NULLS'' in ''algorithm_audits_set.mjs''. **A paper establishes a null if it measures outcome variation under the same treatment**, by any of three routes: **A/A** (arms differing in nothing), **repeats** (one condition run several times with the spread reported), or **generated** (a null distribution built from the paper's own observations — permutation over arm labels, or a randomised baseline).
+Result: **10 of 34 (29.4%)**, against 11 of 32 from the probe. The two numbers are close and that is a coincidence, not corroboration: the sets differ by five papers.
-**Residue of the candidate probes.** Probe 2 returned **190** candidates. **27** of the final 32 audits and **5** of the 12 rejections lie inside it — the other 5 and 7 were caught by probes 4–7 — leaving a residue of **158** papers dropped at title-and-abstract level without an individual note. The union across all seven probes is **461**, of which **44** were adjudicated in depth, so **417** rest on a title-level read. That is the honest residue of this page.
+What this is not:
+  * **Not mechanical.** It is a judgement per paper. {[mai2025_more]} is counted in on //repeats// — six runs of each condition a week apart, with outliers cut on the country-specific standard deviation — and a second reader could reasonably say that is replication rather than a null. {[roongta2025_sheep]} is counted **out**, because its fresh control browser establishes a baseline set of ad exchanges for a **mediation** analysis rather than a same-treatment outcome spread. Both are arguable. The evidence sentence is published for each so the disagreement can be about a specific paper.
+  * **Not a quality verdict.** Several of the 24 report effects far larger than any plausible noise floor. What they cannot do is show the reader that.
+  * **Not re-measured after the population grew.** ''NULLS'' was adjudicated against the 34-paper set, including both papers added in the same pass.
 ===== Recall repair =====
@@ Line 104: / Line 126: @@
 Two sweeps were run //after// the population was settled, because the probe-7 result above showed the scoring had silently discarded 9.4% of it.
-**1. The dropped tail of probe 2.** ''scripts/_aa_residue.mjs'' takes the 158 probe-2 candidates that were never adjudicated and re-scores them with the apparatus-density probe that recovered the three late finds, at a lower threshold and requiring an outcome term. **34 of the 158 clear it. All 34 were read at title level; none is a differential platform audit.** They fall into three groups: user-study randomised trials where the treatment is applied to a //person// and the outcome is that person's behaviour (phishing warnings, Tor nudges, consent dialogs, personalised cookie banners), platform-side A/B tests run //with// the platform rather than against it (//How Intention Informed Recommendations Modulate Choices//, //Reducing Symbiosis Bias through Better A/B Tests//), and passing mentions. Output below.
+**1. The dropped tail of probe 2.** ''scripts/_aa_residue.mjs'' takes the probe-2 candidates that were never adjudicated and re-scores them with the apparatus-density probe that recovered the three late finds, at a lower threshold and requiring an outcome term. **33 of the 157 clear it. All were read at title level; none is a differential platform audit.** They fall into three groups: user-study randomised trials where the treatment is applied to a //person// and the outcome is that person's behaviour (phishing warnings, Tor nudges, consent dialogs, personalised cookie banners), platform-side A/B tests run //with// the platform rather than against it (//How Intention Informed Recommendations Modulate Choices//, //Reducing Symbiosis Bias through Better A/B Tests//), and passing mentions. Output below.
 **2. The five zero-years.** The audit set has no papers in 2011, 2012, 2013, 2017 or 2021. ''_aa_gap.mjs'' requires ≥4 apparatus matches, which is too strict to prove a **negative** — it returns nothing at all for 2017, so a zero from it is uninformative. ''scripts/_aa_zeroyears.mjs'' drops the threshold to ≥2 and requires no outcome term, and refuses to run if any year it is given is not actually a zero-year. Over **1,002** papers in those five years it finds **2** with ≥2 apparatus terms: //An Army of Me// (TheWebConf 2017), which is an observational study of sockpuppets other people created, and //Spotting fake reviewer groups in consumer reviews// (TheWebConf 2012), which is not an audit either. **The zero-years are real in this corpus** — which is a different claim from their being real in the field, since 2013 is the year of Hannak et al. and the screen dropped it.
@@ Line 136: / Line 158: @@
   * **The size of this literature.** The corpus cannot give it; see //The screening loss//. Closing it needs a pass over FAccT, EuroS&P and the IR venues, which is out of scope for a corpus-backed page.
-  * **Whether the 146 dropped probe-2 candidates contain any true audits.** They were read at title-and-abstract level only. A full-text read of all 146 would settle it and was not done.
+  * **Whether the papers dropped at title level contain any more true audits.** The union across all seven probes is 461; 70 were adjudicated in depth, so **391** rest on a title-and-abstract read. The probe-2 tail was re-swept (//Recall repair//) and a generic review pass found one more audit in the score-≥4 shortlist; both are reasons to expect the 391 is not empty. A full-text read of all of them would settle it and was not done.
   * **Whether papers build a null without naming it.** The eight-term probe is a lower bound. Establishing the real rate needs 32 methods sections read for the //concept//, which is a different and slower exercise than the verdict read that produced the population.
   * **Effect sizes.** Every audit measures a different outcome on a different platform with a different metric. Nothing is poolable, and the page publishes no cross-paper effect size deliberately.
@@ Line 192: / Line 214: @@
 ==== Generic (''fable'') ====
-GENERICREVIEW
+Handed the corrected text after the three focused passes. No checklist; asked for whatever they were not looking for. It returned nine findings, ranked. **All nine accepted**; the first three changed the page materially.
+^ # ^ Finding ^ Verdict and what changed ^
+| 1 | The central claim states a **practice** ("only 11 of 32 establish a same-treatment baseline", "the practice is missing too", and the heading "The null is missing from two thirds of them") on the strength of a **term count** — and the reviewer printed the matching sentence behind all 11 hits, showing the probe over-counts as well as under-counts | **accepted.** All 34 papers were read for the question and hand-adjudicated with an evidence sentence each: **10 of 34**. The term table stays, relabelled as a vocabulary finding. See [[#The null: why it is adjudicated and not counted]] |
+| 2 | The adjudication dropped candidates silently: **26 of the 58** score-≥4 shortlist papers had no verdict, //"every verdict is in the output"// was true only of the 44 that got one, and one of the 26 — //Cart-ology// — satisfies the rule | **accepted, and it was right about Cart-ology.** All 58 now carry a written verdict. Cart-ology and the IMC 2024 poster promoted; population **32 → 34**; rejections **12 → 36**. Root cause recorded above: ''_aa_adj.mjs'' returned zero sentences for Cart-ology and that silence was read as a negative |
+| 3 | **Nothing links to the page** except the two routing pages, and [[Design:Platforms]] still says twice that this topic //"has no page on this wiki"// — the exact sentence used to justify creating it | **accepted.** Back-links added from the neighbour pages, and the two stale sentences on [[Design:Platforms]] repointed |
+| 4 | //"Most audit papers supply the first three and skip the fourth"// contradicts the page's own 43.8% vantage-location figure | **accepted.** Replaced with the actual figures |
+| 5 | //"Twenty profiles per condition ... for the same total cost"// is false — same measurements, ten times the training, and the page says training is the expensive part | **accepted.** Rewritten to say what it actually costs and why the corpus is full of two-profile designs. The unsupported "audits are its most natural habitat" was cut too |
+| 6 | {[sun2026_suitability]} is in neither ''AUDITS'' nor ''REJECTED'' yet gets four mentions | **accepted.** Trimmed to one table row that says plainly it is not one of the 34, plus the link |
+| 7 | **"This method is not historical. It is growing."** is not what a 6/4/4/3-year window table with //n// = 34 from a biased screen supports | **accepted.** Replaced with papers-per-year (1.00, 1.25, 3.00, 3.67), the observation that the second window is barely above the first, and a narrower claim |
+| 8 | //"used continuously 2010 → 2026"// for personas is not script-derived; the first ''persona'' hit is 2016 and there are none in 2017–2019 | **accepted.** Corrected, with the note that {[guha2010_challenges]} does the same thing and spells it //personae// |
+| 9 | Smaller: **146** survived in one place after being corrected to 158 elsewhere; //"the two papers most cited"// with no citation count; //"never drive a browser"// said of studies that run in participants' browsers; //"Chrome 141"// stale in a sentence dated 2026-09 | **all accepted.** The residue figures were re-derived for the new sets (**157** and **391**); the "most cited" clause is gone with the rewritten section; the eight no-config papers are described accurately; Chrome pinned to **154**, fetched from the Chrome version-history API on 2026-09-11 |
+**What it found fine, and said so:** length (33 KB against 50–70 KB neighbours), no duplication of the four neighbour pages, the arms-versus-waves noise-floor distinction, the 26-vs-14 sock-puppet reconciliation, the //What this corpus cannot tell you// box, the re-based reporting table, and the //What to report// list with its one-sentence template.
+**One thing it flagged that was not a defect:** the literal token ''GENERICREVIEW'' was live on this page at the time of review. That was this section's own placeholder.
 ===== The report script =====
@@ Line 208: / Line 245: @@
 //       -- profile history, declared attribute, location, device, opt-out
 //       setting, ad creative -- and holds the rest fixed;
-//   (O) the outcome it measures is the platform's own response: ads served,
+//   (O) the outcome it measures is the platform's own DISCRIMINATING response to
-//       results ranked, prices quoted, feed or recommendation contents, or an
+//       the identity it has built: ads served, results ranked, prices quoted, feed
-//       access/labelling decision;
+//       or recommendation contents, or the profile the platform reports back.
+//       NOT reachability. A study whose outcome is whether you can reach the site
+//       at all -- geoblocking, censorship, Tor-exit refusal -- varies a vantage
+//       point rather than an identity and belongs to design:blocking_and_geodifference.
+//       This clause was tightened on 2026-09-11; it excludes no paper already in
+//       the set, and it is what keeps "403 Forbidden: A Global View of CDN
+//       Geoblocking" (IMC 2018) out.
 //   (C) the result is a difference (or a bounded absence of difference)
 //       BETWEEN arms, not a prevalence over a crawl of many sites.
@@ Line 242: / Line 285: @@
   ['WWW/2020/stop-tracking-me-bro-differential-tracking-of-user-demographics-on-hyper-partisa',
     '"We create 9 carefully crafted personas representing different genders and age groups"'],
+  ['CCS/2022/cart-ology-intercepting-targeted-advertising-via-ad-network-identity-entanglemen',
+    'five profiles -- two blank baselines, two account baselines, one entangled attack profile: "All profiles are created and mechanistically measured in the same way, with separation between attacker, victim, and baselines on different machines with different IPs"'],
   ['IMC/2022/measurement-and-analysis-of-implied-identity-in-ad-delivery-optimization',
     '"We ran 200 versions of this ad at the same time, all from the same account and with the same budget"'],
@@ Line 262: / Line 307: @@
   ['PETS/2024/opted-out-yet-tracked-are-regulations-enough-to-protect-your-privacy',
     '"We then visit the filtered websites with user interest personas" -- 16 interest personas, two regions'],
+  ['IMC/2024/poster-identifying-filter-bubble-based-on-feed-level-embedding-similarity-analys',
+    'four bot arms differing only in the video-selection strategy: "The bot selected a video having (a) the minimum JCC, (b) JCC larger than the minimum (random selection), (c) JCC larger than the mean, and (d) JCC larger than WCC"; outcome is the recommended feed'],
   ['WWW/2024/tiktok-and-the-art-of-personalization-investigating-exploration-and-exploitation',
     '"validate our results using a baseline generated from automated TikTok bots, as well as a randomized baseline"'],
@@ Line 291: / Line 338: @@
   ['IEEE-SP/2023/collaborative-ad-transparency-promises-and-limitations', 'explicitly the non-persona alternative: "One method that does not use fake personas"'],
   ['USENIX/2023/problematic-advertising-and-its-disparate-exposure-on-facebook', 'donated ad exposure from real users; disparity is observational'],
-  ['CCS/2022/cart-ology-intercepting-targeted-advertising-via-ad-network-identity-entanglemen', 'attack on ad-network identity linking; no persona arms'],
   ['CCS/2022/privacy-limitations-of-interest-based-advertising-on-the-web-a-post-mortem-empir', 'randomised control is a data permutation over a browsing panel, not a live-platform arm'],
   ['IEEE-SP/2024/targeted-and-troublesome-tracking-and-advertising-on-childrens-websites', 'fresh profile per page visit; the paper itself says "Future work could extend our method to incorporate personas"'],
@@ Line 297: / Line 343: @@
   ['WWW/2026/when-ads-become-profiles-uncovering-the-invisible-risk-of-web-advertising-at-sca', '"Random Control group" is a model ablation, not a platform arm'],
   ['IEEE-SP/2022/deployment-of-source-address-validation-by-network-operators-a-randomized-contro', 'an RCT, but the treatment is a notification to operators; outcome is not platform output'],
+  // The rest of the score->=4 shortlist. A generic review found these had been read
+  // and dropped without a written verdict, and that one of the papers so dropped
+  // (Cart-ology) in fact satisfied the rule. Every shortlisted paper now has a verdict.
+  ['NDSS/2015/bloom-cookies-web-search-personalization-without-user-tracking', 'a privacy-preserving personalisation design evaluated on search logs; no arms'],
+  ['CCS/2018/peeling-the-onions-user-experience-layer-examining-naturalistic-use-of-the-tor-b', 'a 19-participant UX study of Tor Browser; outcome is user experience'],
+  ['USENIX/2023/strategies-and-vulnerabilities-of-participants-in-venezuelan-influence-operation', 'interviews with 19 influence-operation participants; the sockpuppets are theirs, not the authors'],
+  ['CCS/2011/policy-auditing-over-incomplete-logs-theory-implementation-and-applications', 'compliance checking over audit logs; shares the word "audit" and nothing else'],
+  ['WWW/2013/your-browsing-behavior-for-a-big-mac-economics-of-personal-information-online', '168 recruited participants valuing their own PII in an auction; a user study'],
+  ['IEEE-SP/2015/effective-real-time-android-application-auditing', 'program analysis of Android apps; "auditing" means taint tracking'],
+  ['USENIX/2017/exploring-user-perceptions-of-discrimination-in-online-targeted-advertising', 'a survey of user perceptions; the randomised factors are survey vignettes, not platform arms'],
+  ['IMC/2018/403-forbidden-a-global-view-of-cdn-geoblocking', 'vantage points in 177 countries, but the outcome is REACHABILITY. Excluded by the tightened (O) clause; belongs to design:blocking_and_geodifference'],
+  ['CCS/2019/the-art-and-craft-of-fraudulent-app-promotion-in-google-play', 'a study of app-store fraud workers who operate sockpuppets; observational'],
+  ['WWW/2022/characterizing-detecting-and-predicting-online-ban-evasion', "Wikipedia's own labelled sockpuppet groups; observational, no arms"],
+  ['WWW/2022/fairness-audit-of-machine-learning-models-with-confidential-computing', 'ML fairness auditing inside a TEE; no live platform'],
+  ['IEEE-SP/2022/towards-automated-auditing-for-account-and-session-management-flaws-in-single-si', 'SSO implementation flaws; outcome is a security bug, not a discriminating response'],
+  ['NDSS/2023/tactics-threats-targets-modeling-disinformation-and-its-mitigation', 'interviews with fact-checkers and analysts; sockpuppets are the subject, not the instrument'],
+  ['WWW/2023/scoping-fairness-objectives-and-identifying-fairness-metrics-for-recommender-sys', "practitioner interviews about fairness metrics; no measurement of a platform"],
+  ['IEEE-SP/2023/when-and-why-do-people-want-ad-targeting-explanations-evidence-from-a-four-week', 'a four-week field study of what people want from ad explanations; outcome is attitudes'],
+  ['USENIX/2023/auditing-framework-apis-via-inferred-app-side-security-specifications', 'Android framework API access control; "auditing" is static analysis'],
+  ['CCS/2024/curator-attack-when-blackbox-differential-privacy-auditing-loses-its-power', 'differential-privacy auditing; a different object with the same name'],
+  ['USENIX/2024/efficient-privacy-auditing-in-federated-learning', 'membership inference against an FL model; not a platform'],
+  ['USENIX/2024/what-do-you-want-from-theory-alone-experimenting-with-tight-auditing-of-differen', 'DP synthetic-data auditing; not a platform'],
+  ['USENIX/2024/fledging-will-continue-until-privacy-improves-empirical-analysis-of-googles-priv', 'security analysis of the FLEDGE/Protected Audience API; the outcome is API behaviour and attack feasibility, not a served-ad contrast across arms'],
+  ['NDSS/2025/exploring-user-perceptions-of-security-auditing-in-the-web3-ecosystem', 'user perceptions of smart-contract audits'],
+  ['PETS/2025/privacy-perceptions-and-behaviors-towards-targeted-advertising-on-social-media-a', 'an n=412 cross-country survey; outcome is attitudes'],
+  ['PETS/2026/audagent-automated-auditing-of-privacy-policy-compliance-in-ai-agents', 'policy-compliance checking of AI agents; no arms'],
+  ['PETS/2026/privacy-in-theory-bugs-in-practice-grey-box-auditing-of-differential-privacy-lib', 'DP library auditing; not a platform'],
+  ['WWW/2026/question-the-questions-auditing-representation-in-online-deliberative-processes', 'algorithm design for question selection in deliberative polls; not a measurement of a deployed platform'],
 ];
+// Does the paper measure outcome variation under the SAME treatment? That is the
+// step the page argues is skipped, and a term probe cannot answer it: the eight
+// phrasings in _aa_noise.mjs let in a "noise floor" that describes someone else's
+// paper, and miss Cart-ology's four identically-measured baseline profiles, which
+// match none of them. So it is hand-adjudicated here, like the population, with
+// the sentence that settled it. Three routes count:
+//   (a) A/A  -- two or more arms that differ in nothing;
+//   (b) repeats -- the same condition run several times, with the spread reported;
+//   (c) generated -- a null distribution built from the paper's own observations
+//       (permutation over arm labels, or a randomised baseline).
+// Anything not listed here was read and found to state a difference without ever
+// measuring what a non-difference looks like.
+export const NULLS = {
+  'IMC/2010/challenges-in-measuring-online-advertising-systems':
+    ['A/A', '"Even queries launched simultaneously from two identically configured clients on the same subnet can produce wildly different ads"; and "In this paper we perform all analysis relative to a control experiment"'],
+  'CCS/2015/sunlight-fine-grained-targeting-detection-at-scale-with-statistical-confidence':
+    ['generated', 'exact/random permutation test against an explicit H0, on a held-out split of the profiles, with Benjamini-Yekutieli or Holm correction'],
+  'IMC/2015/location-location-location-the-impact-of-geolocation-on-web-search-personalizati':
+    ['A/A', '"The red line compares two treatments at the baseline location (i.e., the experimental control), and thus shows the noise floor."'],
+  'PETS/2015/automated-experiments-on-ad-privacy-settings':
+    ['generated', 'permutation test over the arm labels -- "The permutation test randomly permutes ... the control and experimental treatments" -- run over "blocks of nearly identical agents"'],
+  'WWW/2019/measuring-political-personalization-of-google-news-search':
+    ['A/A', 'edit distances tested "among four identical browser profiles" as the reference for the trained-profile comparison'],
+  'CCS/2022/cart-ology-intercepting-targeted-advertising-via-ad-network-identity-entanglemen':
+    ['A/A', 'four unused baseline profiles measured identically; one diverged and the authors read that as the noise: "Such non-determinism is expected, validating our strategy of deploying numerous baseline profiles and the use of normalization."'],
+  'WWW/2024/tiktok-and-the-art-of-personalization-investigating-exploration-and-exploitation':
+    ['generated', '"a randomized baseline" against which the personalisation features are scored'],
+  'PETS/2025/more-and-scammier-ads-the-perils-of-youtubes-ad-privacy-settings':
+    ['repeats', 'each condition run six times at least a week apart ("we repeat each experiment six times"), with outliers cut at standard deviations of the country-specific mean'],
+  'IMC/2025/from-voice-to-ads-auditing-commercial-smart-speakers-for-targeted-advertising-ba':
+    ['A/A + generated', 'paired puppets differing in nothing, scheduled in parallel; plus a permutation test against a random dataset'],
+  'PETS/2026/ad-personalization-and-transparency-in-mobile-ecosystems-a-comparative-analysis':
+    ['A/A + generated', '"our baseline persona uses the same neutral persona type for control and treatment. For the baseline persona, the majority of measurement pairs produce insignificant tests" -- an A/A that behaved like one -- plus a 9,999-permutation test'],
+};
 </file>
@@ Line 325: / Line 434: @@
 const byKey = new Map(papers.map((p) => [`${p.venue}/${p.year}/${p.slug}`, p]));
-import { AUDITS, REJECTED } from './algorithm_audits_set.mjs';
+import { AUDITS, REJECTED, NULLS } from './algorithm_audits_set.mjs';
 // ---------------------------------------------------------------------------
@@ Line 352: / Line 461: @@
 // Pin them here so an accidental edit to the set fails the run instead of
 // silently moving every figure on the page.
-if (audit.length !== 32) die(`audit population is ${audit.length}, the page says 32 -- update the page and this contract together`);
+if (audit.length !== 34) die(`audit population is ${audit.length}, the page says 34 -- update the page and this contract together`);
-if (REJECTED.length !== 12) die(`rejected set is ${REJECTED.length}, the page says 12 -- update the page and this contract together`);
+if (REJECTED.length !== 36) die(`rejected set is ${REJECTED.length}, the page says 36 -- update the page and this contract together`);
 if (AUDITS.some(([k]) => REJECTED.some(([r]) => r === k))) die('a key is in both AUDITS and REJECTED');
 for (const [key] of REJECTED) if (!byKey.get(key)) die(`rejected paper not in the extraction: ${key}`);
@@ Line 373: / Line 482: @@
 // --- by year -----------------------------------------------------------------
-console.log('## Audit papers by year (denominator: the 32-paper audit set)');
+console.log('## Audit papers by year (denominator: the audit set)');
 const years = [...new Set(papers.map((p) => p.year))].sort();
 let run = 0;
@@ Line 387: / Line 496: @@
 console.log(`  years with zero audit papers: ${gaps.join(', ')}`);
 console.log('');
+let wsum = 0;
 for (const [lo, hi] of [[2010, 2015], [2016, 2019], [2020, 2023], [2024, 2026]]) {
   const n = audit.filter((p) => p.year >= lo && p.year <= hi).length;
-  console.log(`  ${lo}-${hi}: ${n} of ${N} (${pct(n, N)})`);
+  wsum += n;
+  const span = hi - lo + 1;
+  console.log(`  ${lo}-${hi} (${span} yr): ${n} of ${N} (${pct(n, N)}), ${(n / span).toFixed(2)} papers/year`);
 }
+if (wsum !== N) die(`window buckets sum to ${wsum}, not ${N}`);
 console.log('');
@@ Line 432: / Line 545: @@
   ['states a vantage location', audit, vantageStated, papers.filter((p) => p.vantage.length > 0), vantageStated, 'measuredFrom (3,908)'],
 ];
-console.log(`  ${'indicator'.padEnd(34)} ${'audit (32)'.padStart(14)}   baseline`);
+console.log(`  ${'indicator'.padEnd(34)} ${`audit (${N})`.padStart(14)}   baseline`);
 for (const [label, aSet, aFn, bSet, bFn, bName] of rows) {
   const a = aSet.filter(aFn).length;
@@ Line 440: / Line 553: @@
 console.log('');
 console.log('## The crawl-config rows again, on a SHARED denominator');
-console.log("   The rows above compare 32 audits (8 of which have no crawl config) against");
+console.log(`   The rows above compare ${N} audits (${N - audit.filter((p) => p.crawlConfig !== null).length} of which have no crawl config) against`);
 console.log('   1,120 crawled papers (40 of which have none). A field that can only be stated');
 console.log('   on a paper that HAS a config must be counted over papers that have one, or the');
@@ Line 526: / Line 639: @@
 for (const [k, [c, a]] of Object.entries(vocabCount))
   console.log(`  ${k.padEnd(32)} ${String(c).padStart(7)}  ${String(a).padStart(5)} of ${N}`);
+console.log('');
+// --- the null, hand-adjudicated ---
+console.log('## Establishing a null, HAND-ADJUDICATED (not a term count)');
+console.log('   Does the paper measure outcome variation under the SAME treatment?');
+console.log('   Routes: A/A (arms differing in nothing), repeats (same condition run');
+console.log("   several times, spread reported), generated (null built from the paper's");
+console.log('   own data). Evidence sentence per paper in algorithm_audits_set.mjs.');
+for (const k of Object.keys(NULLS)) if (!AUDITS.some(([a]) => a === k)) die(`NULLS key not in AUDITS: ${k}`);
+const byRoute = {};
+for (const [k, [route]] of Object.entries(NULLS)) (byRoute[route] ||= []).push(k);
+for (const [route, ks] of Object.entries(byRoute)) console.log(`  ${route.padEnd(16)} ${ks.length}`);
+const nNull = Object.keys(NULLS).length;
+console.log(`  ${'TOTAL'.padEnd(16)} ${nNull} of ${N} (${pct(nNull, N)})`);
+console.log(`  no null established: ${N - nNull} of ${N} (${pct(N - nNull, N)})`);
+for (const [k, [route, why]] of Object.entries(NULLS)) {
+  const p = byKey.get(k);
+  console.log(`    ${p.year} ${p.venue.padEnd(8)} [${route}] ${p.title}`);
+  console.log(`        ${why}`);
+}
 console.log('');
@@ Line 611: / Line 744: @@
   crawled                                     1120
   inferential (any non-descriptive statistic)  1762
-  AUDIT (hand-adjudicated, rule above)        32
+  AUDIT (hand-adjudicated, rule above)        34
-  adjudicated and rejected                    12
+  adjudicated and rejected                    36
-## Audit papers by year (denominator: the 32-paper audit set)
+## Audit papers by year (denominator: the audit set)
    1  #
    0
@@ Line 627: / Line 760: @@
    2  ##
    0
-   5  #####
+   6  ######
    4  ####
-   2  ##
+   3  ###
    5  ##### *provisional
    3  ### *provisional
-  total 32
+  total 34
   years with zero audit papers: 2011, 2012, 2013, 2017, 2021
--2015: 6 of 32 (18.8%)
+-2015 (6 yr): 6 of 34 (17.6%), 1.00 papers/year
--2019: 5 of 32 (15.6%)
+-2019 (4 yr): 5 of 34 (14.7%), 1.25 papers/year
--2023: 11 of 32 (34.4%)
+-2023 (4 yr): 12 of 34 (35.3%), 3.00 papers/year
--2026: 10 of 32 (31.3%)
+-2026 (3 yr): 11 of 34 (32.4%), 3.67 papers/year
 ## Audit papers by venue, against that venue's own output
-  CCS        3 of   990    0.3%
+  CCS        4 of   990    0.4%
   IEEE-SP    1 of   767    0.1%
-  IMC        7 of   638    1.1%
+  IMC        8 of   638    1.3%
   NDSS       1 of   701    0.1%
   PETS      10 of   510    2.0%
@@ Line 650: / Line 783: @@
 ## Platform measured (multi-valued; does not sum to N)
-  web                    25 of 32  78.1%
+  web                    27 of 34  79.4%
-  other-online-service   18 of 32  56.3%
+  other-online-service   19 of 34  55.9%
-  mobile                  4 of 32  12.5%
+  mobile                  4 of 34  11.8%
-  iot                     3 of 32  9.4%
+  iot                     3 of 34  8.8%
-  offline                 0 of 32  0.0%
+  offline                 0 of 34  0.0%
 ## What audit papers report, vs the comparable corpus baseline
    Each row states both populations. The baseline is the population named,
    not "all papers", so the two cells are comparable.
-  indicator                              audit (32)   baseline
+  indicator                              audit (34)   baseline
-  runs a non-descriptive statistic    21/32 (65.6%)   1637/5118 (32.0%) of empirical (5,118)
+  runs a non-descriptive statistic    21/34 (61.8%)   1637/5118 (32.0%) of empirical (5,118)
-  states an ethics-review outcome     20/32 (62.5%)   1728/5118 (33.8%) of empirical (5,118)
+  states an ethics-review outcome     20/34 (58.8%)   1728/5118 (33.8%) of empirical (5,118)
-  states artifact availability        22/32 (68.8%)   2890/5118 (56.5%) of empirical (5,118)
+  states artifact availability        22/34 (64.7%)   2890/5118 (56.5%) of empirical (5,118)
-  has a crawlConfig at all            24/32 (75.0%)   1080/1120 (96.4%) of crawled (1,120)
+  has a crawlConfig at all            26/34 (76.5%)   1080/1120 (96.4%) of crawled (1,120)
-  states crawl statefulness           22/32 (68.8%)   219/1120 (19.6%) of crawled (1,120)
+  states crawl statefulness           24/34 (70.6%)   219/1120 (19.6%) of crawled (1,120)
-  states interaction depth            21/32 (65.6%)   841/1120 (75.1%) of crawled (1,120)
+  states interaction depth            22/34 (64.7%)   841/1120 (75.1%) of crawled (1,120)
-  states a vantage location           14/32 (43.8%)   1228/3908 (31.4%) of measuredFrom (3,908)
+  states a vantage location           15/34 (44.1%)   1228/3908 (31.4%) of measuredFrom (3,908)
 ## The crawl-config rows again, on a SHARED denominator
-   The rows above compare 32 audits (8 of which have no crawl config) against
+   The rows above compare 34 audits (8 of which have no crawl config) against
 ,120 crawled papers (40 of which have none). A field that can only be stated
    on a paper that HAS a config must be counted over papers that have one, or the
    two cells are not comparable. Both framings are printed; the page uses this one.
-  statefulness       audit 22/24 ( 91.7%)   crawled 219/1080 (20.3%)
+  statefulness       audit 24/26 ( 92.3%)   crawled 219/1080 (20.3%)
-  interactionDepth   audit 21/24 ( 87.5%)   crawled 841/1080 (77.9%)
+  interactionDepth   audit 22/26 ( 84.6%)   crawled 841/1080 (77.9%)
-  consentAction      audit 11/24 ( 45.8%)   crawled 349/1080 (32.3%)
+  consentAction      audit 12/26 ( 46.2%)   crawled 349/1080 (32.3%)
-  headless           audit 5/24 ( 20.8%)   crawled 140/1080 (13.0%)
+  headless           audit 5/26 ( 19.2%)   crawled 140/1080 (13.0%)
   -- the 8 audit papers with no crawl config at all:
 IMC      Peeking Beneath the Hood of Uber.
@@ Line 704: / Line 837: @@
   Counts and percentages of advertisements
   -- correction families, folded by hand (paper-counted; the skeleton fold does not merge these):
-       Holm-Bonferroni (any spelling)   5 of 32
+       Holm-Bonferroni (any spelling)   5 of 34
-       Bonferroni without Holm          3 of 32
+       Bonferroni without Holm          3 of 34
-       any Bonferroni-family            7 of 32
+       any Bonferroni-family            7 of 34
-       Benjamini-Hochberg/Yekutieli     1 of 32
+       Benjamini-Hochberg/Yekutieli     1 of 34
-  -- 11 of 32 audit papers report no non-descriptive statistic at all:
+  -- 13 of 34 audit papers report no non-descriptive statistic at all:
 IMC Challenges in measuring online advertising systems.
 CCS Your Online Interests: Pwned! A Pollution Attack Against Targeted Advertising.
@@ Line 715: / Line 848: @@
 WWW AdBudgetKiller: Online Advertising Budget Draining Attack.
 PETS Investigating sources of PII used in Facebook’s targeted advertising
+CCS Cart-ology: Intercepting Targeted Advertising via Ad Network Identity Entanglement.
 NDSS HARPO: Learning to Subvert Online Behavioral Advertising
 CCS Marketing to Children Through Online Targeted Advertising: Targeting Mechanisms and Legal Aspects.
 PETS A Utility-Preserving Obfuscation Approach for YouTube Recommendations
+IMC Poster: Identifying Filter Bubble Based on Feed-Level Embedding Similarity Analysis.
 PETS Echoes of Privacy: Uncovering the Profiling Practices of Voice Assistants
 IEEE-SP Setting the Course, but Forgetting to Steer: Analyzing Compliance with GDPR's Right of Access to Data by Instagram, TikTok, and Youtube.
@@ Line 724: / Line 859: @@
   full text present for 5855 of 5859 papers
   term                              corpus  in audit set
-  sock puppet                           26      5 of 32
+  sock puppet                           26      5 of 34
-  persona                              190     14 of 32
+  persona                              190     14 of 34
-  control profile/account/persona       33     11 of 32
+  control profile/account/persona       33     11 of 34
-  treatment group/profile               66      5 of 32
+  treatment group/profile               66      5 of 34
-  trained profile                        4      1 of 32
+  trained profile                        4      1 of 34
-  A/A test                               2      0 of 32
+  A/A test                               2      0 of 34
-  noise floor                           35      2 of 32
+  noise floor                           35      2 of 34
-  price discrimination/steering         33      9 of 32
+  price discrimination/steering         33      9 of 34
+## Establishing a null, HAND-ADJUDICATED (not a term count)
+   Does the paper measure outcome variation under the SAME treatment?
+   Routes: A/A (arms differing in nothing), repeats (same condition run
+   several times, spread reported), generated (null built from the paper's
+   own data). Evidence sentence per paper in algorithm_audits_set.mjs.
+  A/A              4
+  generated        3
+  repeats          1
+  A/A + generated  2
+  TOTAL            10 of 34 (29.4%)
+  no null established: 24 of 34 (70.6%)
+IMC      [A/A] Challenges in measuring online advertising systems.
+        "Even queries launched simultaneously from two identically configured clients on the same subnet can produce wildly different ads"; and "In this paper we perform all analysis relative to a control experiment"
+CCS      [generated] Sunlight: Fine-grained Targeting Detection at Scale with Statistical Confidence.
+        exact/random permutation test against an explicit H0, on a held-out split of the profiles, with Benjamini-Yekutieli or Holm correction
+IMC      [A/A] Location, Location, Location: The Impact of Geolocation on Web Search Personalization.
+        "The red line compares two treatments at the baseline location (i.e., the experimental control), and thus shows the noise floor."
+PETS     [generated] Automated Experiments on Ad Privacy Settings
+        permutation test over the arm labels -- "The permutation test randomly permutes ... the control and experimental treatments" -- run over "blocks of nearly identical agents"
+WWW      [A/A] Measuring Political Personalization of Google News Search.
+        edit distances tested "among four identical browser profiles" as the reference for the trained-profile comparison
+CCS      [A/A] Cart-ology: Intercepting Targeted Advertising via Ad Network Identity Entanglement.
+        four unused baseline profiles measured identically; one diverged and the authors read that as the noise: "Such non-determinism is expected, validating our strategy of deploying numerous baseline profiles and the use of normalization."
+WWW      [generated] TikTok and the Art of Personalization: Investigating Exploration and Exploitation on Social Media Feeds.
+        "a randomized baseline" against which the personalisation features are scored
+PETS     [repeats] More and Scammier Ads: The Perils of YouTube's Ad Privacy Settings
+        each condition run six times at least a week apart ("we repeat each experiment six times"), with outliers cut at standard deviations of the country-specific mean
+IMC      [A/A + generated] From Voice to Ads: Auditing Commercial Smart Speakers for Targeted Advertising based on Voice Characteristics.
+        paired puppets differing in nothing, scheduled in parallel; plus a permutation test against a random dataset
+PETS     [A/A + generated] Ad Personalization and Transparency in Mobile Ecosystems: A Comparative Analysis of Google's and Apple's EU App Stores
+        "our baseline persona uses the same neutral persona type for control and treatment. For the baseline persona, the majority of measurement pairs produce insignificant tests" -- an A/A that behaved like one -- plus a 9,999-permutation test
 ## Screening loss: audit-topical papers in the index but not in the extraction
@@ Line 768: / Line 935: @@
   crawled                                     1120
   inferential (any non-descriptive statistic)  1762
-  AUDIT (hand-adjudicated, rule above)        32
+  AUDIT (hand-adjudicated, rule above)        34
-  adjudicated and rejected                    12
+  adjudicated and rejected                    36
-## Audit papers by year (denominator: the 32-paper audit set)
+## Audit papers by year (denominator: the audit set)
    1  #
    0
@@ Line 784: / Line 951: @@
    2  ##
    0
-   5  #####
+   6  ######
    4  ####
-   2  ##
+   3  ###
    5  ##### *provisional
    3  ### *provisional
-  total 32
+  total 34
   years with zero audit papers: 2011, 2012, 2013, 2017, 2021
--2015: 6 of 32 (18.8%)
+-2015: 6 of 34 (17.6%)
--2019: 5 of 32 (15.6%)
+-2019: 5 of 34 (14.7%)
--2023: 11 of 32 (34.4%)
+-2023: 12 of 34 (35.3%)
--2026: 10 of 32 (31.3%)
+-2026: 11 of 34 (32.4%)
 ## Audit papers by venue, against that venue's own output
-  CCS        3 of   990    0.3%
+  CCS        4 of   990    0.4%
   IEEE-SP    1 of   767    0.1%
-  IMC        7 of   638    1.1%
+  IMC        8 of   638    1.3%
   NDSS       1 of   701    0.1%
   PETS      10 of   510    2.0%
@@ Line 807: / Line 974: @@
 ## Platform measured (multi-valued; does not sum to N)
-  web                    25 of 32  78.1%
+  web                    27 of 34  79.4%
-  other-online-service   18 of 32  56.3%
+  other-online-service   19 of 34  55.9%
-  mobile                  4 of 32  12.5%
+  mobile                  4 of 34  11.8%
-  iot                     3 of 32  9.4%
+  iot                     3 of 34  8.8%
-  offline                 0 of 32  0.0%
+  offline                 0 of 34  0.0%
 ## What audit papers report, vs the comparable corpus baseline
    Each row states both populations. The baseline is the population named,
    not "all papers", so the two cells are comparable.
-  indicator                              audit (32)   baseline
+  indicator                              audit (34)   baseline
-  runs a non-descriptive statistic    21/32 (65.6%)   1637/5118 (32.0%) of empirical (5,118)
+  runs a non-descriptive statistic    21/34 (61.8%)   1637/5118 (32.0%) of empirical (5,118)
-  states an ethics-review outcome     20/32 (62.5%)   1728/5118 (33.8%) of empirical (5,118)
+  states an ethics-review outcome     20/34 (58.8%)   1728/5118 (33.8%) of empirical (5,118)
-  states artifact availability        22/32 (68.8%)   2890/5118 (56.5%) of empirical (5,118)
+  states artifact availability        22/34 (64.7%)   2890/5118 (56.5%) of empirical (5,118)
-  has a crawlConfig at all            24/32 (75.0%)   1080/1120 (96.4%) of crawled (1,120)
+  has a crawlConfig at all            26/34 (76.5%)   1080/1120 (96.4%) of crawled (1,120)
-  states crawl statefulness           22/32 (68.8%)   219/1120 (19.6%) of crawled (1,120)
+  states crawl statefulness           24/34 (70.6%)   219/1120 (19.6%) of crawled (1,120)
-  states interaction depth            21/32 (65.6%)   841/1120 (75.1%) of crawled (1,120)
+  states interaction depth            22/34 (64.7%)   841/1120 (75.1%) of crawled (1,120)
-  states a vantage location           14/32 (43.8%)   1228/3908 (31.4%) of measuredFrom (3,908)
+  states a vantage location           15/34 (44.1%)   1228/3908 (31.4%) of measuredFrom (3,908)
 ## The crawl-config rows again, on a SHARED denominator
-   The rows above compare 32 audits (8 of which have no crawl config) against
+   The rows above compare 34 audits (8 of which have no crawl config) against
 ,120 crawled papers (40 of which have none). A field that can only be stated
    on a paper that HAS a config must be counted over papers that have one, or the
    two cells are not comparable. Both framings are printed; the page uses this one.
-  statefulness       audit 22/24 ( 91.7%)   crawled 219/1080 (20.3%)
+  statefulness       audit 24/26 ( 92.3%)   crawled 219/1080 (20.3%)
-  interactionDepth   audit 21/24 ( 87.5%)   crawled 841/1080 (77.9%)
+  interactionDepth   audit 22/26 ( 84.6%)   crawled 841/1080 (77.9%)
-  consentAction      audit 11/24 ( 45.8%)   crawled 349/1080 (32.3%)
+  consentAction      audit 12/26 ( 46.2%)   crawled 349/1080 (32.3%)
-  headless           audit 5/24 ( 20.8%)   crawled 140/1080 (13.0%)
+  headless           audit 5/26 ( 19.2%)   crawled 140/1080 (13.0%)
   -- the 8 audit papers with no crawl config at all:
 IMC      Peeking Beneath the Hood of Uber.
@@ Line 861: / Line 1028: @@
   Counts and percentages of advertisements
   -- correction families, folded by hand (paper-counted; the skeleton fold does not merge these):
-       Holm-Bonferroni (any spelling)   5 of 32
+       Holm-Bonferroni (any spelling)   5 of 34
-       Bonferroni without Holm          3 of 32
+       Bonferroni without Holm          3 of 34
-       any Bonferroni-family            7 of 32
+       any Bonferroni-family            7 of 34
-       Benjamini-Hochberg/Yekutieli     1 of 32
+       Benjamini-Hochberg/Yekutieli     1 of 34
-  -- 11 of 32 audit papers report no non-descriptive statistic at all:
+  -- 13 of 34 audit papers report no non-descriptive statistic at all:
 IMC Challenges in measuring online advertising systems.
 CCS Your Online Interests: Pwned! A Pollution Attack Against Targeted Advertising.
@@ Line 872: / Line 1039: @@
 WWW AdBudgetKiller: Online Advertising Budget Draining Attack.
 PETS Investigating sources of PII used in Facebook’s targeted advertising
+CCS Cart-ology: Intercepting Targeted Advertising via Ad Network Identity Entanglement.
 NDSS HARPO: Learning to Subvert Online Behavioral Advertising
 CCS Marketing to Children Through Online Targeted Advertising: Targeting Mechanisms and Legal Aspects.
 PETS A Utility-Preserving Obfuscation Approach for YouTube Recommendations
+IMC Poster: Identifying Filter Bubble Based on Feed-Level Embedding Similarity Analysis.
 PETS Echoes of Privacy: Uncovering the Profiling Practices of Voice Assistants
 IEEE-SP Setting the Course, but Forgetting to Steer: Analyzing Compliance with GDPR's Right of Access to Data by Instagram, TikTok, and Youtube.
@@ Line 881: / Line 1050: @@
   full text present for 5855 of 5859 papers
   term                              corpus  in audit set
-  sock puppet                           26      5 of 32
+  sock puppet                           26      5 of 34
-  persona                              190     14 of 32
+  persona                              190     14 of 34
-  control profile/account/persona       33     11 of 32
+  control profile/account/persona       33     11 of 34
-  treatment group/profile               66      5 of 32
+  treatment group/profile               66      5 of 34
-  trained profile                        4      1 of 32
+  trained profile                        4      1 of 34
-  A/A test                               2      0 of 32
+  A/A test                               2      0 of 34
-  noise floor                           35      2 of 32
+  noise floor                           35      2 of 34
-  price discrimination/steering         33      9 of 32
+  price discrimination/steering         33      9 of 34
+## Establishing a null, HAND-ADJUDICATED (not a term count)
+   Does the paper measure outcome variation under the SAME treatment?
+   Routes: A/A (arms differing in nothing), repeats (same condition run
+   several times, spread reported), generated (null built from the paper's
+   own data). Evidence sentence per paper in algorithm_audits_set.mjs.
+  A/A              4
+  generated        3
+  repeats          1
+  A/A + generated  2
+  TOTAL            10 of 34 (29.4%)
+  no null established: 24 of 34 (70.6%)
+IMC      [A/A] Challenges in measuring online advertising systems.
+        "Even queries launched simultaneously from two identically configured clients on the same subnet can produce wildly different ads"; and "In this paper we perform all analysis relative to a control experiment"
+CCS      [generated] Sunlight: Fine-grained Targeting Detection at Scale with Statistical Confidence.
+        exact/random permutation test against an explicit H0, on a held-out split of the profiles, with Benjamini-Yekutieli or Holm correction
+IMC      [A/A] Location, Location, Location: The Impact of Geolocation on Web Search Personalization.
+        "The red line compares two treatments at the baseline location (i.e., the experimental control), and thus shows the noise floor."
+PETS     [generated] Automated Experiments on Ad Privacy Settings
+        permutation test over the arm labels -- "The permutation test randomly permutes ... the control and experimental treatments" -- run over "blocks of nearly identical agents"
+WWW      [A/A] Measuring Political Personalization of Google News Search.
+        edit distances tested "among four identical browser profiles" as the reference for the trained-profile comparison
+CCS      [A/A] Cart-ology: Intercepting Targeted Advertising via Ad Network Identity Entanglement.
+        four unused baseline profiles measured identically; one diverged and the authors read that as the noise: "Such non-determinism is expected, validating our strategy of deploying numerous baseline profiles and the use of normalization."
+WWW      [generated] TikTok and the Art of Personalization: Investigating Exploration and Exploitation on Social Media Feeds.
+        "a randomized baseline" against which the personalisation features are scored
+PETS     [repeats] More and Scammier Ads: The Perils of YouTube's Ad Privacy Settings
+        each condition run six times at least a week apart ("we repeat each experiment six times"), with outliers cut at standard deviations of the country-specific mean
+IMC      [A/A + generated] From Voice to Ads: Auditing Commercial Smart Speakers for Targeted Advertising based on Voice Characteristics.
+        paired puppets differing in nothing, scheduled in parallel; plus a permutation test against a random dataset
+PETS     [A/A + generated] Ad Personalization and Transparency in Mobile Ecosystems: A Comparative Analysis of Google's and Apple's EU App Stores
+        "our baseline persona uses the same neutral persona type for control and treatment. For the baseline persona, the majority of measurement pairs produce insignificant tests" -- an A/A that behaved like one -- plus a 9,999-permutation test
 ## Screening loss: audit-topical papers in the index but not in the extraction
@@ Line 947: / Line 1148: @@
       WWW/2020/stop-tracking-me-bro-differential-tracking-of-user-demographics-on-hyper-partisa
       IN: "We create 9 carefully crafted personas representing different genders and age groups"
+CCS      Cart-ology: Intercepting Targeted Advertising via Ad Network Identity Entanglement.
+      CCS/2022/cart-ology-intercepting-targeted-advertising-via-ad-network-identity-entanglemen
+      IN: five profiles -- two blank baselines, two account baselines, one entangled attack profile: "All profiles are created and mechanistically measured in the same way, with separation between attacker, victim, and baselines on different machines with different IPs"
 IMC      Measurement and analysis of implied identity in ad delivery optimization.
       IMC/2022/measurement-and-analysis-of-implied-identity-in-ad-delivery-optimization
@@ Line 977: / Line 1181: @@
       PETS/2024/opted-out-yet-tracked-are-regulations-enough-to-protect-your-privacy
       IN: "We then visit the filtered websites with user interest personas" -- 16 interest personas, two regions
+IMC      Poster: Identifying Filter Bubble Based on Feed-Level Embedding Similarity Analysis.
+      IMC/2024/poster-identifying-filter-bubble-based-on-feed-level-embedding-similarity-analys
+      IN: four bot arms differing only in the video-selection strategy: "The bot selected a video having (a) the minimum JCC, (b) JCC larger than the minimum (random selection), (c) JCC larger than the mean, and (d) JCC larger than WCC"; outcome is the recommended feed
 WWW      TikTok and the Art of Personalization: Investigating Exploration and Exploitation on Social Media Feeds.
       WWW/2024/tiktok-and-the-art-of-personalization-investigating-exploration-and-exploitation
@@ Line 1018: / Line 1225: @@
 USENIX   Problematic Advertising and its Disparate Exposure on Facebook
       OUT: donated ad exposure from real users; disparity is observational
-CCS      Cart-ology: Intercepting Targeted Advertising via Ad Network Identity Entanglement.
-      OUT: attack on ad-network identity linking; no persona arms
 CCS      Privacy Limitations of Interest-based Advertising on The Web: A Post-mortem Empirical Analysis of Google's FLoC.
       OUT: randomised control is a data permutation over a browsing panel, not a live-platform arm
@@ Line 1030: / Line 1235: @@
 IEEE-SP  Deployment of Source Address Validation by Network Operators: A Randomized Control Trial.
       OUT: an RCT, but the treatment is a notification to operators; outcome is not platform output
+NDSS     Bloom Cookies: Web Search Personalization without User Tracking
+      OUT: a privacy-preserving personalisation design evaluated on search logs; no arms
+CCS      Peeling the Onion's User Experience Layer: Examining Naturalistic Use of the Tor Browser.
+      OUT: a 19-participant UX study of Tor Browser; outcome is user experience
+USENIX   Strategies and Vulnerabilities of Participants in Venezuelan Influence Operations
+      OUT: interviews with 19 influence-operation participants; the sockpuppets are theirs, not the authors
+CCS      Policy auditing over incomplete logs: theory, implementation and applications.
+      OUT: compliance checking over audit logs; shares the word "audit" and nothing else
+WWW      Your browsing behavior for a big mac: economics of personal information online.
+      OUT: 168 recruited participants valuing their own PII in an auction; a user study
+IEEE-SP  Effective Real-Time Android Application Auditing.
+      OUT: program analysis of Android apps; "auditing" means taint tracking
+USENIX   Exploring User Perceptions of Discrimination in Online Targeted Advertising
+      OUT: a survey of user perceptions; the randomised factors are survey vignettes, not platform arms
+IMC      403 Forbidden: A Global View of CDN Geoblocking.
+      OUT: vantage points in 177 countries, but the outcome is REACHABILITY. Excluded by the tightened (O) clause; belongs to design:blocking_and_geodifference
+CCS      The Art and Craft of Fraudulent App Promotion in Google Play.
+      OUT: a study of app-store fraud workers who operate sockpuppets; observational
+WWW      Characterizing, Detecting, and Predicting Online Ban Evasion.
+      OUT: Wikipedia's own labelled sockpuppet groups; observational, no arms
+WWW      Fairness Audit of Machine Learning Models with Confidential Computing.
+      OUT: ML fairness auditing inside a TEE; no live platform
+IEEE-SP  Towards Automated Auditing for Account and Session Management Flaws in Single Sign-On Deployments.
+      OUT: SSO implementation flaws; outcome is a security bug, not a discriminating response
+NDSS     Tactics, Threats & Targets: Modeling Disinformation and its Mitigation
+      OUT: interviews with fact-checkers and analysts; sockpuppets are the subject, not the instrument
+WWW      Scoping Fairness Objectives and Identifying Fairness Metrics for Recommender Systems: The Practitioners' Perspective.
+      OUT: practitioner interviews about fairness metrics; no measurement of a platform
+IEEE-SP  When and Why Do People Want Ad Targeting Explanations? Evidence from a Four-Week, Mixed-Methods Field Study.
+      OUT: a four-week field study of what people want from ad explanations; outcome is attitudes
+USENIX   Auditing Framework APIs via Inferred App-side Security Specifications
+      OUT: Android framework API access control; "auditing" is static analysis
+CCS      Curator Attack: When Blackbox Differential Privacy Auditing Loses Its Power.
+      OUT: differential-privacy auditing; a different object with the same name
+USENIX   Efficient Privacy Auditing in Federated Learning
+      OUT: membership inference against an FL model; not a platform
+USENIX   "What do you want from theory alone?" Experimenting with Tight Auditing of Differentially Private Synthetic Data Generation
+      OUT: DP synthetic-data auditing; not a platform
+USENIX   Fledging Will Continue Until Privacy Improves: Empirical Analysis of Google's Privacy-Preserving Targeted Advertising
+      OUT: security analysis of the FLEDGE/Protected Audience API; the outcome is API behaviour and attack feasibility, not a served-ad contrast across arms
+NDSS     Exploring User Perceptions of Security Auditing in the Web3 Ecosystem
+      OUT: user perceptions of smart-contract audits
+PETS     Privacy Perceptions and Behaviors Towards Targeted Advertising on Social Media: A Cross-Country Study on the Effect of Culture and Religion
+      OUT: an n=412 cross-country survey; outcome is attitudes
+PETS     AudAgent: Automated Auditing of Privacy Policy Compliance in AI Agents
+      OUT: policy-compliance checking of AI agents; no arms
+PETS     Privacy in Theory, Bugs in Practice: Grey-Box Auditing of Differential Privacy Libraries
+      OUT: DP library auditing; not a platform
+WWW      Question the Questions: Auditing Representation in Online Deliberative Processes.
+      OUT: algorithm design for question selection in deliberative polls; not a measurement of a deployed platform
 ## Screening loss in full
@@ Line 1398: / Line 1653: @@
 <file text _aa_noise-output.txt>
-AUDIT keys parsed from report script: 32
+AUDIT keys parsed from report script: 34
 term                                    corpus  audit
 A/A test                                     2      0
@@ Line 1409: / Line 1664: @@
 two arms with the same treatment            13      1
-audit papers with >=1 noise-baseline term: 11 of 32
+audit papers with >=1 noise-baseline term: 11 of 34
   CCS/2015/sunlight-fine-grained-targeting-detection-at-scale-with-statistical-confidence
      permutation / randomisation test
@@ Line 1441: / Line 1696: @@
   PETS/2020/inferring-tracker-advertiser-relationships-in-the-online-advertising-ecosystem-u
   WWW/2020/stop-tracking-me-bro-differential-tracking-of-user-demographics-on-hyper-partisa
+  CCS/2022/cart-ology-intercepting-targeted-advertising-via-ad-network-identity-entanglemen
   IMC/2022/measurement-and-analysis-of-implied-identity-in-ad-delivery-optimization
   IMC/2022/what-factors-affect-targeting-and-bids-in-online-advertising-a-field-measurement
@@ Line 1450: / Line 1706: @@
   WWW/2023/a-method-to-assess-and-explain-disparate-impact-in-online-retailing
   PETS/2024/opted-out-yet-tracked-are-regulations-enough-to-protect-your-privacy
+  IMC/2024/poster-identifying-filter-bubble-based-on-feed-level-embedding-similarity-analys
   PETS/2025/echoes-of-privacy-uncovering-the-profiling-practices-of-voice-assistants
   PETS/2025/sheeps-clothing-wolfish-intent-automated-detection-and-evaluation-of-problematic
@@ Line 1490: / Line 1747: @@
 <file text _aa_residue-output.txt>
-loose=190 adjudicated=32 residue=158 residue_with_apparatus_and_outcome=34
+loose=190 adjudicated=33 residue=157 residue_with_apparatus_and_outcome=33
   2021 CCS      The Effect of Google Search on Software Security: Unobtrusive Security Interventions via Content Re-ranking.
   2020 WWW      Finding a Choice in a Haystack: Automatic Extraction of Opt-Out Statements from Privacy Policy Text.
   2025 IEEE-SP  Restricting the Link: Effects of Focused Attention and Time Delay on Phishing Warning Effectiveness.
-  2023 USENIX   Strategies and Vulnerabilities of Participants in Venezuelan Influence Operations
   2022 PETS     Increasing Adoption of Tor Browser Using Informational and Planning Nudges
   2019 WWW      How Intention Informed Recommendations Modulate Choices: A Field Study of Spoken Word Content.