The problem it solves
Live video is moderation’s hardest case: harm happens in real time, and after-the-fact review is too late by definition. Most platforms answer with either no live features or unmoderated ones. Verify exists so a platform can offer live streaming without abandoning its safety contract.
How it works
Verify applies the Candor screening discipline to the live path — monitoring streams as they happen and feeding moderation verdicts while a stream is still running, with escalation to human review and stream-level action where signals warrant it. The same fail-closed philosophy applies: live features exist because the moderation exists, not despite it.
Verified behaviors
Every claim below maps to a dated behavioral proof against the live production service.
The live moderation pipeline (capture, transcription, text evaluation, and stream termination) fires end-to-end during an active broadcast.
method: a live stream was captured, segmented, and transcribed in real time; a designated substitute trigger stood in for genuine extremist material, since possessing real TVEC-caliber content to test with is neither safe nor necessary · observed: capture, transcription, evaluation, and kill-switch activation fired correctly in sequence, terminating the stream — the evaluation step hands off to the same ThreatShield engine independently verified against real content elsewhere in this evidence record, so the live path reuses a proven judgment rather than an untested one · VERIFIED June 17, 2026
Live transcription works accurately in both English and Spanish, feeding real transcripts to moderation in near-real-time.
method: real live audio was transcribed in both languages and the resulting text logged · observed: fluent, accurate transcripts produced and logged in both languages from real audio · VERIFIED June 17, 2026
Live content is transcribed and evaluated within a short detection window, not after a meaningful delay.
method: timestamped production logs captured across the capture-to-evaluation pipeline · observed: a 4-8 second window from spoken content to evaluated transcript, confirmed in production · VERIFIED June 17, 2026
Honest limitations
- Real-time moderation operates within latency limits — screening a live stream is by nature a race, and no system wins it every time.
- Automated verdicts on live content bias toward escalation to humans; consequential stream actions are not automation-only.
Download the white paper (PDF)