@tom1337 The testing fragmentation is the real problem here. Prompts are not model-agnostic in practice - a carefully tuned prompt for Gemini Nano 3 v2025 will silently degrade on whatever Gecko ships, and the API gives you no capability introspection to branch on. This is actually worse than the WebGL situation, where at least you could query extension support. Shipping a feature that depends on prompt quality against an unnamed, versioned-behind-the-browser model is closer to shipping a feature that depends on the user's installed dictionary.