Why the model only ever sees the schema

Every conversation with an insurer or a bank gets to the same question within ten minutes. Not how fast is it, but what does it see. If the answer involves rows leaving the building, the meeting is over, and it should be.

So we made one rule early and built everything else around it. The only thing a language model ever receives is the schema: table names, column names, types, and how they join. Not a sample row. Not a value. Not a masked value either, because masking is a promise, and a promise is not a control.

What that buys you

• There is nothing to redact, because nothing sensitive is ever in the request.
• The audit question "what did the model see" has a one-word answer, and your security team can verify it rather than take our word for it.
• Air-gapped works the same way as hosted. The boundary does not move when the deployment does.

What it costs

It is not free, and we would rather say so here than on a call. A column called amt_2 tells a model less than a column called paid_amount. When the name and the type do not settle what a column means, the agent asks you, and you answer. That is a few minutes on the day you connect a source. It is the price of the rule, and we think it is the right price.

Masking is a promise. Schema-only is a control. Regulated buyers know the difference, and so do their auditors.

Everything else on the platform sits downstream of this. The pipelines the agents write run inside your perimeter, against your tables. The questions people ask in plain English are answered by queries that run there too. The model drafts the shape of the work. Your data never leaves to do it.