All code is strictly on-purpose: the extension wraps window.fetch only to intercept [domain withheld]'s own API responses (usage, conversation, SSE streams) and post them via window.postMessage to a content script that renders a local counter UI. Every network call made by the extension targets [domain withheld] first-party endpoints ( /{orgId} and /{id}) using the user's existing session credentials — identical to what [domain withheld] itself does. No data is transmitted to any external server. The ML signals are explainable noise: the single POST detection is the bridge detecting [domain withheld]'s own completion POSTs (not making one), js_createelement_script is the legitimate injection of a declared web_accessible_resource (), the large JS size (~2 MB) is almost certainly the bundled o200k GPT tokenizer referenced in but absent from the extracted ZIP, and all 17 appendchild calls are UI DOM construction. The third-party score's elevated likelihood is purely, not behavioral, and explicitly contradicts the low impact score.