Compliance framework
MagenDetect maps directly to the EU Digital Services Act and the IHRA Working Definition of Antisemitism — the two regulatory frameworks that define what platforms must detect and report.
DSA Article 34
Article 34 of the Digital Services Act requires very large online platforms (VLOPs) to identify, analyze, and mitigate systemic risks — including the dissemination of content that incites hatred or violence against protected groups. MagenDetect provides the detection layer.
Detection
Every piece of content is scored against 21 signal types across three tracks, with per-snippet evidence and IHRA references.
Audit trail
Every detection is persisted with confidence scores, the routing decision, character-level snippets, and the model's rationale. Queryable via API.
Transparency
The four-stage pipeline is fully documented. Signal taxonomies, attribution multipliers, and routing thresholds are published — not a black box.
IHRA Working Definition
The International Holocaust Remembrance Alliance's Working Definition of Antisemitism, adopted by 43 countries and the European Parliament, provides 11 illustrative examples of what constitutes antisemitism. MagenDetect's dog-whistle dictionary and three-track analysis are grounded in these examples.
| IHRA example | MagenDetect signal type | Track |
|---|---|---|
| IHRA-1: Denying the right of self-determination | delegitimization_of_israel | Anti-Israel bias |
| IHRA-3: Double standards | double_standards_against_israel | Anti-Israel bias |
| IHRA-4: Blood libel | blood_libel_modern | Anti-Israel bias |
| IHRA-5: Holocaust denial/distortion | holocaust_inversion | Antisemitism |
| IHRA-6: Dehumanization | dehumanization | Antisemitism |
| IHRA-7: Jewish conspiracy/control | classic_trope_conspiracy_control | Antisemitism |
| IHRA-8: Dual loyalty | dual_loyalty_insinuation | Antisemitism |
The full 11-example mapping is available in the API documentation.
Attribution awareness
The most common failure mode in content moderation is flagging journalism about antisemitism as antisemitic. MagenDetect's attribution-aware calibration solves this by distinguishing who said what:
"The ADL condemned speakers who said the Jewish lobby controls foreign policy"
Detected as rebuttal → confidence multiplied by 0.15× → routed to auto-accept
"It is well known that the Jewish lobby controls American foreign policy"
Detected as publisher voice → confidence at 1.0× → routed to auto-reject