ax1om could tell that a dataset arrived as a CSV. It could not tell whether the file held your org’s real records or a copy of the sample data we ship, and model selection now depends on that answer: once real data exists, a model trained on sample data drops out of the recommendation.
The console’s CSV import wizard asks the question once, at upload. The MCP
upload tools take the same data_origin field. The answer is stored as a
first-class provenance input on the data source, beside the provider, and
every readout that names a model’s provenance reads it from there.
We also fingerprint the sample datasets we ship. An uploaded copy of our demo data is recognised whatever it is named, so a model trained on it is labelled sample-backed even when someone declared otherwise.