AI4TRUST reports a final warning system, a textual component, a human-validation dashboard and user evaluation. Each finding belongs to its stated component and test.
Choose a subject or a numbered connection.
Solid lines: documented in a cited record. Dashed: reported or inferred. Hollow arrows: proposed powers. Choose a connection for its evidence.
A textual weighted F1 near 56% is not aggregate system accuracy. Reported users and annotations do not establish production deployment, automatic sanctions or independent replication.
Report period1March2024-28February2026; updated3August2026, not operational launch date.
Reported by the cited source
Final-period report describes an ensemble of LightGBM classifiers with isotonic regression calibration, combining six AI tools and providing triage categories/feature attribution. Exact inputs, weights, category thresholds and approval rules are not published in this read record.
The metric applies to the textual signals tool detecting39signals across6tactics/8languages. It is not aggregate warning-system performance or probability a particular claim is false. No benchmark was reproduced.
Final report describes the named dashboard and iterative user evaluation. It supplies no override, score replacement, reasons, publisher appeal or correction-propagation rule.
Population includes media professionals, fact-checkers, policy experts and students. These reported rounds are not an independent Commission audit, actual downstream recipient list or measured advertising effect.