Two searchers work from two different indexes. Two models design your course without seeing each other's plan. Two more write every lesson independently. A separate model then compares the drafts and checks the claims against the sources. That is not a slogan — it is the pipeline, step by step, below.
No person reads your course.
The crew below are roles in an automated pipeline, not staff. Each one is a language model doing one job. The models run through OpenRouter (Anthropic, DeepSeek, Google, OpenAI, Thinking Machines, xAI) · Perplexity. We name them so you can see what actually happened to your material — not to suggest anybody signed off on it.
Two of them, sent down different roads. Sonar works its own search index. Ranger searches the live web in parallel, on a different index, and is noticeably better at Russian-language material. Where they agree, a fact rests on two independent findings; where they differ, the difference travels downstream instead of being quietly picked.
Currently standing down. During the pilot we do not download the source pages. Sources are found and cited, and the knowledge base is built from what the search step returned — so nothing here, or anywhere in the product, claims a page was read when it was not.
Turns what came back into an ordered knowledge base. Everything written later has to ground its facts in this, which is what makes the last step possible at all: you cannot check a claim without something to check it against.
Two models design the course independently: which lessons are the Pareto core you cannot skip, which are optional, and — stated out loud — what the course deliberately leaves out. That last list is on the course page. A course that claims to cover everything is a course that has not decided anything.
Reads both plans and builds one out of them, taking the better of each rather than tossing a coin. If one plan is unusable, the other still stands — one failure never costs you the course.
Two models write each lesson independently, to a fixed shape: an explanation, a worked example, something to try in five minutes, a way to check yourself, and what to take away. They draw the diagrams too — there is no separate illustrator. Figures are drawn as code and as plain vector schematics, so every label in them is text somebody can check.
The Arbiter again, in its second job. It compares the two drafts, then takes every risky claim — numbers, dates, names, causal statements — and checks it against the Librarian's base. What contradicts the sources is removed. What the sources are merely silent about loses its false precision.
Diagrams are claims too. Every label, number and arrow in a figure is checked the same way as a sentence. A diagram whose sources do not support it is corrected or removed — a picture is exactly as easy to be wrong in as a paragraph, and much easier to believe.
Not part of writing. A separate check run over the finished course that refuses it if a lesson is missing its structure, if a figure never reached the page, or if anything was left unchecked. Its whole value is in what it turns down.