Architecture

The classify pipeline is four stages: Intake & normalise (deterministic), Layer 1 baseline rules (deterministic, immutable), Layer 2 owner overlay (deterministic), and Classify — the one metered stage. The quality and toxicity scoring is deterministic. The owner's final approve / hold / reject is a separate, owner-gated action that publishes on approve.

The app owns the review-with-moderation entity in its demo_eco_c3_review table and reads the shared host #01 property catalog. On approve it performs the campsite-canon plain row write into the core demo_eco_review store (one published-review store — not a schema change). The data invariant holds throughout: owner writes are canonical (credential_id NULL); a viewer's writes are credential-scoped, ephemeral, and reset-clearing, and a DO-UPDATE scope guard means a viewer edit can never clobber a canonical row.

Moderating a review emits review.moderated on the in-process event bus (persisted to the durable event log), carrying the id, decision, and property. The upstream trigger is a guest review submission. Downstream, the property detail page and the trust/analytics surface (#13) consume the published outcome through the shared core store.

The one metered stage (classifyReview) runs dual-mode: Cloud claude-haiku-4-5, cost-capped at $0.05 per session and fail-closed — over the cap it falls back to recorded OSS (qwen2.5-7b-instruct) at $0, which is recorded on local hardware and never fabricated. Each run writes a cost_ledger row and a demo_cache replay key (slug · stage · prompt-version · input-hash). Out of scope and clearly labelled: the classifier is a deterministic simulation in the prototype; no real PII, no real imagery, no production prompts appear anywhere. Honesty is inspector-only.