No, Silas is not a trained model. Silas is methodology: guardrails, checks and bounds. It uses AI models to read and draft, and it is not trained on customer or government data.
The distinction changes what there is to ask about. What we built is the layer around the reading and the drafting: eligibility checked against the set-aside actually named in a notice, a compliance matrix assembled against your real profile, and a durable record of why each judgement was made. Silas also refuses controlled material, deliberately, which is a design decision rather than a limit we ran into.
