OnBrand is an account brain with gates. It plans, generates, judges its own work against your rubric, publishes, reads the result and changes what it does next. Here is the whole machine, described the way it actually runs.
Every arrow is logged. Every model call is priced. Nothing here is a diagram of an intention.
Every approve, reject, redo and tweak is written down with the reason attached. Those verdicts are folded back into the next prompt, so the account stops making the mistake you already corrected. It is the only part of this system that cannot be copied, because it is made of your decisions.
Each candidate is graded rule by rule against your brand rubric by a vision model, and separately scored on whether a stranger with no context would stop. Two different questions. Is this us, and does it earn the second of attention. Both answers travel with the post.
Before a single word is generated the system assembles what the model gets to see: brand memories, your own best captions, standing notes, the active era and episode. That manifest is recorded, so when an output is wrong we can tell a bad model from bad context.
At night the system reads its own failures and drafts refinements to how it writes and plans. Then it earns them. Each refinement is replayed against real historical briefs through the real pipeline, and it must win on a held out split it never trained against. A refinement that only pleases its own judge is thrown out and logged as such.
Every kill, low score and rejection is classified: missing context, stale context, model limit, prompt defect, or plain taste mismatch. Repeated tags become repair proposals with a named fix. Nothing changes itself. A person accepts or dismisses each one.
Across every account we keep distributions only. What formats hold attention, what caption lengths survive, what hours work. Numbers with no owner and no content in them. They enter your account as soft priors, and your own history always outranks them.
Your edits are the most honest signal in the product, more honest than any rating. Each one is banked as a preference pair in the shape a future model needs. We are not fine tuning anything today and we will say the day we do. The vault compounds either way, and it stays yours.
Bad output is a question with two answers. Bad model, or bad context. The system is built to tell you which.
Exportable, deletable, processed in the EU, never sold and never used to train a shared model. The intelligence is rented today and swappable tomorrow. The taste is yours from the first verdict.