Pattern radar
Combines static analysis hints with policy keywords to flag modules that deserve human review — not a court verdict.
EU AI Act deadline: Aug 2, 2026 · compliance window narrowing
Compliance console
Article 5 · Safety scan
Catch prohibited patterns before they ship
This workspace models how Agent Mai highlights potential Article 5 conflicts — social scoring cues, workplace emotion analytics, and biometric surveillance patterns — so product and legal can intervene while code is still cheap to change.
Live heuristics
Pattern library
Traceable
Module-level pointers
Escalation
Legal + eng
Audit link
Integrity report
How the safety scan complements your audit
Use it alongside the Quick Audit: classification tells you documentation gaps; prohibited-practice scanning tells you whether the use-case itself may be off-limits.
Article 5 lens
Mapped to the latest consolidated text — social scoring, manipulation, biometric surveillance, and emotion inference in sensitive contexts.
Evidence hooks
Each hit surfaces file paths, log excerpts, and recommended mitigations so engineering can patch or scope down features.
Governance bridge
Export bundles align with your integrity report and Annex IV narrative — one story for regulators and boards.
Legal handoff
Escalation prompts include jurisdiction notes and when to involve outside counsel or DPA consultation.
Continuous use
Re-run on every release candidate; prohibited signals can appear when models or integrations change.
Safety scan flow
DetectInspectEscalateResolve
Illustrative scan
Review required on sample workspace
Demo timestampSynthetic · CET
Modules reviewed
14
Flags for review
04
Heuristic confidence
High
Article 5 prohibited practices
Illustrative findings — always validate with your counsel.
Social scoring
Critical riskParameters that evaluate trustworthiness from social behavior or personality proxies — often incompatible with fundamental rights safeguards.
Emotion recognition in workplaces
High riskSystems inferring emotional states to monitor employees — Article 5(1)(f) triggers fast for productivity surveillance patterns.
Remote biometric ID
Prohibited caseLive biometric identification in publicly accessible spaces for law enforcement — narrow exceptions, heavy documentation.
Subliminal manipulation
ProhibitedTechniques beyond human perception meant to distort behavior — requires immediate product and legal escalation.
Technical deep-dive (sample)
Detection logic
[SYSTEM_LOG]: Initializing tensor analysis on Module 'HR_EVAL_PRO_V3'…
[WARNING]: FacialFeatureAnalyzer API reachable from worker runtime.
[RESULT]: Pattern consistent with workplace productivity surveillance.
[COMPLIANCE_NOTE]: Assess Article 5(1)(f) — emotion recognition in work/education contexts.
[WARNING]: FacialFeatureAnalyzer API reachable from worker runtime.
[RESULT]: Pattern consistent with workplace productivity surveillance.
[COMPLIANCE_NOTE]: Assess Article 5(1)(f) — emotion recognition in work/education contexts.
Codebase pointer
/src/analytics/worker_sentiment/facial_engine.pyL: 243–289
Suggested mitigation
Suspend facial metadata in production; route oversight workflows through human review before re-enabling any affective inference.
EU AI Act compliance engine (illustrative)Encrypted audit trail in production
Pair this scan with a full audit pass
Upload documentation in the Quick Audit, then cross-check prohibited heuristics — your board gets one coherent narrative.
Open quick auditOperationalize remediation with owners
Connect flagged modules to your integrity report, assign engineering and legal leads, and track fixes through release.