Model weights are not retrained. Approved context is retrieved at generation time from what you’ve written, edited, and approved — the same way retrieval works for research sources.
Learning Controls
Each source of personalization can be turned off independently:
Turning a control off stops that signal from being used in new generations; it does not delete history already recorded.
Active Signals
The signals panel lists what evidence is currently available to the personalization layer — each with an evidence count and whether it’s active or paused. A signal with zero evidence has nothing to draw on yet regardless of whether its control is on.Pinned Rules
Pinned rules are durable, freeform instructions applied to every generation for your account — for example, a standing correction like never describing the organization as an industry leader without evidence. Add or remove them at any time; a change applies to generations from that point forward.Learned Voice Examples
Voice exemplars are passages recovered from substantial human rewrites, not raw model output. A generation that was accepted unedited does not contribute a voice example; a paragraph you materially rewrote does. This is what “learn from substantial edits” actually populates.Verified Claims
Manage the approved facts adaptive intelligence can retrieve as evidence.
Brand Voice
Style profiles define voice at the account level; adaptive intelligence refines it from your edits.