9 September 2026. This lane reuses the accepted AI4TRUST D2.1 original in ../gdi-horizon-role-wave-2026-09-09/horizon-d21.pdf, SHA256 1222963ce874a0455e10109cafff95ad201e1a43c1d6be6cfdd1848f9a2f2dfe. No new independent source or download. Current classification-run root text scope: PDF4-5 contents,53-59 and81-82. Earlier Horizon root text scope was2,3,11,12,34,35,55,65,68,69 and visual2,11,55. These are selected pages of90, not a complete annex audit.
PDF53-54 makes climate, health and migration initial development/test topics, with reusable selection criteria including cross-country relevance, historicity, factual assessability and consortium feasibility. It acknowledges that misinformation need not be a coordinated campaign. Its feasibility discussion is a dated2023 design account, not a current exclusion decision. It points to D2.2 for active topic-mining elaboration.
PDF55 combines research/media/factchecker inputs with GDI's existing internal lists and some ChatGPT suggestions. Broad initial keyword lists are to be refined by removing irrelevant terms and adding co-occurring terms. There is no per-term provenance: individual entries cannot all be attributed to GDI. These are collection seeds, not a website-risk codebook or an automatic sanction list.
PDF56-58 includes neutral subjects such as weather, renewable energy, clinical trials, adverse events, immigration law and migrant rights alongside charged narratives, public figures and conspiracy language. Inclusion does not label a named person or proposition false. Duplicates and mixed capitalization occur; this document does not prove runtime configuration. The migration notes permit national or origin terms to vary by country and period. No bulk targeting list is reproduced here.
PDF58-59 explicitly declares an intersectional methodological framing attentive to harm, discrimination and social identity. That is an observable normative choice. Its effect on annotation labels or thresholds still needs a rule-level record; the framing alone does not prove suppression. PDF81-82 shows German annex terms are locally adapted, rather than a literal English copy. Selected pages do not identify translator approval or update authority. They do not show German GDI authored every German term. French tail/Greek opening on these pages were contextual reading, not complete language-list review.
The revealing distinction is collection perimeter versus treatment rule. Broad search terms can shape a sample, but a bias finding requires inspecting resulting coverage, annotation decisions and validation. The final report separately describes aggregate model scoring and human validation. Sampling, annotation, inference, human judgment and downstream action stay separate in this case. Actual codebook, aggregate thresholds, override rules and agreements remain unacquired. No person/site score audit or model execution occurred.