Differences

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

--- provenance:privacy:server_side_tracking [2026/08/21 14:25] – Section 11: author review, Mertens HAL v2 full text, re-run of both audit scripts against 18a312c, check_tables nowiki fix karel.kubicek.claude
+++ provenance:privacy:server_side_tracking [2026/08/21 14:47] (current) – Section 11.2: record the attribution error and the rule for an interested expert's claims vs assessments karel.kubicek.claude
@@ Line 1159: / Line 1159: @@
 What survives, and is now stated as three separate things rather than one: no non-Google ground truth exists (SST-Guard §8); the signal does not identify //which// Tag — Mertens et al. group five distinct Meta CAPI templates because "it is not possible to identify which one is installed"; and nothing confirms the data reached Meta. Plus the shared boundary: all of it requires the deployment to ride in a GTM container.
-==== 11.2 The heuristic he flagged, and why the page hedges it ====
+==== 11.2 The heuristic he flagged, whose assessment it is, and why that matters ====
 His words: "Mertens et. al (2026) has opened the door for basic heuristics like "_fbp cookie set as first party AND a GTM template for meta pixel not detected". It is a super super super weak heuristic, but the paper is accepted."
-Recorded on the page with his assessment attributed and quoted, because a reader who meets it second-hand will not otherwise know its own author thinks it is weak. One thing added beyond what he said, and it is mine: the discriminating power sits entirely in the //absence// of the client-side template, and Mertens et al. report **6.7%** of GTM sites obfuscating the container's presence — so the negative half of the conjunction is the half most easily broken. The page says candidate list, not prevalence estimate.
+**An attribution error in the first version of this section, corrected the same day.** The page carried this as "the one heuristic Mertens et al. make available for Meta is very weak, **and its author says so**", which reads as if Mertens et al. had rated their own result weak. They did not: they neither propose this heuristic nor evaluate it, and no author of that paper was contacted at any point. The "super super super weak" verdict is Jazlan's, and **SST-Guard is a directly competing detector of the same phenomenon** — so it is one method's author rating what a rival's result enables. Karel flagged this and had flagged the competing interest in advance; the page now names whose assessment it is, that it is a competitor's, and that Mertens et al. were not asked, in the same sentence as the quote.
+Recording the quote is still right — a reader who meets this heuristic second-hand should know one expert calls it very weak — but a quoted verdict from a competitor is evidence about the field's disagreements, not a neutral evaluation, and the page must not launder it into one. **What the page rests on instead is the technical objection, which is mine and holds independently of who raised it**: the discriminating power sits entirely in the //absence// of the client-side template, and Mertens et al. themselves report **6.7%** of GTM sites obfuscating the container's presence — so the negative half of the conjunction is the half most easily broken. That argument would survive if the verdict were withdrawn.
+**The general rule this run learned.** Expert review from a named source in the field is worth more than any amount of my own re-reading — §11.1 is the proof, since the substantive correction there came from him and was then confirmed straight out of Mertens et al.'s Table 7. But a source with a stake sets a second obligation, not a lower weight: separate what they claim from what they assess. Claims get checked against the primary source and, once verified, stand on that source (§11.1 needs no attribution to Jazlan at all — Table 7 says it). Assessments of a competitor stay attributed, are never used as the load-bearing reason for anything, and carry the competing interest next to the quote. The failure mode here was collapsing an assessment into a claim by writing "its author" — one wrong noun turned a rival's opinion into the paper's own admission.
 ==== 11.3 The code claims, re-run against the fixed release ====
@@ Line 1200: / Line 1204: @@
 | 2026-08-21 | **The rendered page initially showed only 5 of 13 references.** ''bibtex4dw'' caches the bibliography, so newly added keys render as inline markers with no entry. Fixed by requesting ''?purge=true'' on ''literature:bibliography'' and then on the page; re-verified 13 of 13 |
 | 2026-08-21 | Four-reviewer pass (§10); fixes applied; this provenance page published |
+| 2026-08-21 | **Third sitting.** Karel caught an attribution error: the "super super super weak" verdict on Mertens et al.'s heuristic is Jazlan's, a competing author's, and had been written as if it were the paper's own authors' ("its author says so"). Corrected on the content page and recorded in §11.2, together with the rule for handling an interested expert's assessments versus their claims |
 | 2026-08-21 | **Second sitting (§11).** Author review from Jazlan (SST-Guard). Fetched the Mertens et al. HAL v2 full text; corrected the non-Google-detection claim from Table 7; pulled ''18a312c''; made ''sst_guard_templates.mjs'' and ''sst_guard_replay.mjs'' build-agnostic and re-ran both against both releases; fixed the nowiki blind spot in ''check_tables.mjs''; added 6 EXTERNAL entries to ''sst_number_guard.mjs'' and re-ran it to zero unaccounted figures |