Sourcine treats the model as an untrusted component. Facts come from your approved records; the model may improve wording, never author a fact, number, citation, or contract term. Here is the pipeline every governed output passes through — and why each step exists.
The order matters: facts are fixed before the model is ever invoked, and nothing unsupported leaves the pipeline by default.
Before anything runs, the facts the output needs are retrieved from your approved records — systems of record, answer libraries, policy documents, contract repositories. The model has no say in what is true.
Critical facts, identifiers, and required fields are placed into the output by software, directly from those sources. This is where names, numbers, dates, control statements, and contract terms come from — assembled, not generated.
Only after the facts are fixed does an LLM optionally smooth the wording. It runs behind a fail-closed gate, may change phrasing, and is barred from introducing new facts, entities, or figures. In many workflows this step is suppressed entirely.
The candidate output is checked back against its source. Protected fields — names, dates, IDs, numbers, citations, control statements, contract terms — must match exactly. Any drift is a failure, not a rounding error.
If validation fails, the output is blocked and routed for review — it is not published. The default state is “do not ship,” not “ship and hope.”
The accepted output is recorded with the source it used, the version of that source, and the validation result — a tamper-evident record you can produce on request.
Retrieval-augmented generation still lets the model form the final answer — so a wrong or altered fact remains possible. Deterministic-first inverts the flow: the answer is composed from approved sources before any model runs, and the model is reduced to an optional, gated style pass. The difference is where authority lives.
We scope a single high-value workflow against an approved source set — you see governed output, protected-field validation, and an audit trail before committing to more.