Add knowledge
"Teach my AI" means add knowledge — it does not mean retrain the model.
Those two are constantly collapsed into one another and they should not be. Retraining changes the model's weights, is expensive, is hard to undo, and makes it impossible to say afterwards which source produced which sentence. Adding knowledge leaves the model alone and gives it something to read, with a citation attached — which is why a source can be withdrawn later and actually be gone.
Indexing
Reversible. Attributable. A source can be removed and the system genuinely no longer knows it. may_train_model is not consulted, because nothing is being trained.
Training
Not reversible in any practical sense. A source that goes into weights cannot be taken out of them, so may_train_model is a separate declaration and defaults to refusal.
Try it — nothing is uploaded
Drop a file. It is measured in your browser.
The file is read with FileReader, hashed with crypto.subtle, and classified by the same formatOf() and laneOf() the build uses. There is no request. Watch your network tab.
Drop a file here, or click to choose one.
CSV, JSON, Markdown, text, Word, code — anything.
What the ingestion engine will not decide for you
Whether it may leave
Measurable facts are measured. may_leave_device is not measurable and is never assumed — it stays refused until somebody writes it down in data/declarations.json.
How authoritative it is
A file's format says nothing about its authority. A statute and a rough note can both be markdown. Authority is declared with a reason recorded beside it.
When it was last true
A modification time is when the file was written. It is not when anybody confirmed the contents were still correct, and treating one as the other is how a system starts quoting last year's rule with this year's confidence.