Architecture

Parse (deterministic keyword routing to a governed view) → guard (PII-free view enforced) → plan (measure × dimension → parameterized SQL) → execute (governed aggregates) → narrate (one-line summary). Only the narrate phrasing is the metered model step; routing, planning and aggregation are deterministic, so the answer is reproducible and the guardrail is not at the model's discretion.

Four governed metric views are the entire read surface: bookings, reviews, guest segments (k-anonymised, minimum cell size ≥ 50), and partner scorecards (no contract terms or settlement amounts). There is no row-level path. BI Chat performs no canonical writes; the only thing it persists is a cost-ledger row per ask.

Each ask emits cost.logged on the shared event bus and writes a row into the shared C8 cost ledger that #27 FinOps reads. The metered stage runs dual-mode: Cloud `claude-haiku-4-5` (cost-capped, fail-closed) or OSS recorded on local hardware at $0. Honesty rule: an un-recorded OSS run fails closed to "not captured" — the demo never fabricates a model output it did not record.

The dataset is synthetic and the warehouse is in-memory; the "metered" model round-trip is simulated over recorded behavior. No real BI warehouse, no real guest data, no production prompts or proprietary metrics appear anywhere. In Stage-2 the same method surface maps to real governed repositories with no change to the chat UI.