This note describes an internal analytics pattern without naming the product, customers, connector configuration, or report contents.
The system ingests official-source platform data, processes it through background workers, computes scores and anomalies deterministically, then generates structured reports.
The language model does not calculate the score.
Why the boundary matters
Scoring rules must be repeatable. The same validated input should produce the same score, anomaly flags, and ranking before any prose is written.
The model receives sanitised, validated report JSON and converts it into a readable narrative. It can explain a result but cannot silently change the result.
This separation improves reviewability, makes regression testing possible, and prevents prompt changes from altering business logic.
Operational controls
The private build also separates credentialed ingestion, database access, scoring, report generation, and narration. Credentials remain encrypted, OAuth state is expiry-bound, and browser writes sit behind identity and role checks.
The public lesson is simple: language is a presentation layer when the underlying decision must remain deterministic.