<?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/6610b17cec874164ba78db4926d02838&quot; frameborder=&quot;0&quot; width=&quot;1920&quot; height=&quot;1440&quot; webkitallowfullscreen mozallowfullscreen allowfullscreen&gt;&lt;/iframe&gt;</html><height>1440</height><width>1920</width><provider_name>Loom</provider_name><provider_url>https://www.loom.com</provider_url><thumbnail_height>1440</thumbnail_height><thumbnail_width>1920</thumbnail_width><thumbnail_url>https://cdn.loom.com/sessions/thumbnails/6610b17cec874164ba78db4926d02838-8ad3ae186aa5ac69.gif</thumbnail_url><duration>293.518</duration><title>Hearth —AI Product Launch | TPM Program Walkthrough | Zana Yocum</title><description>This Loom explains how I designed and gated a kitchen planning guide to prevent unsafe advice. It emphasizes five hard rules and an explicit safety constraint that is outside any cutability ranking, with costs and ranges made visible in the interface. An eval harness with 28 cases including indirect attacks found the guide passed quality at 93% but failed the safety gate at 80%, including attempts that used credential claims or escalation. Launch gating relied on an onboarding experiment, a reality check cost of 5.5 points of completion, and a suppression countermetric that detected a routing change dropping safety referrals for 4 days; the team halted, rolled back, added a check, and shipped safely to six of seven targets.</description></oembed>