<?xml version="1.0" encoding="UTF-8"?><oembed><type>video</type><version>1.0</version><html>&lt;iframe src=&quot;https://www.loom.com/embed/b159c9172b6840f3b504f8cf18ef8f11&quot; frameborder=&quot;0&quot; width=&quot;1718&quot; height=&quot;1288&quot; webkitallowfullscreen mozallowfullscreen allowfullscreen&gt;&lt;/iframe&gt;</html><height>1288</height><width>1718</width><provider_name>Loom</provider_name><provider_url>https://www.loom.com</provider_url><thumbnail_height>1288</thumbnail_height><thumbnail_width>1718</thumbnail_width><thumbnail_url>https://cdn.loom.com/sessions/thumbnails/b159c9172b6840f3b504f8cf18ef8f11-960717fce97619b8.gif</thumbnail_url><duration>106.938</duration><title>Fixed Vision Verifies Repairs With Evidence</title><description>This Loom explains a tool called Fixed Vision that detects whether a repair path is truly fixed or only appears fixed. It describes a half-blocked pump case where the pump still hums and pumps slowly, meaning a repair can look fixed without being properly verified. The team built the system during a Meta hackathon, using committed data and listing third-party tools, and notes it uses Meta models including Segment Anything 1.3, SAM 3.1, and Mouse Voice Transcription 1.0 for Hindi and English voice, plus video, transcript, retrieval, and approved guidance. It also states that a specific engine and defined rules and code decide verification instead of the underlying model.</description></oembed>