Campsite Content Generation

Browse the synthetic campsite catalog. Each property's description carries a provenance chip - "AI-drafted · editor-approved" - and a "How this was written" link that opens the inspector with the full two-stage trace. A language switcher swaps the displayed copy across the supported languages (a labelled mock translation, $0).

Operators open the Console → Content studio: a priority-sorted editorial queue with QC and claim-flag badges, a Generate picker (campsite × content type × voice × language × Cloud/OSS), and a per-draft panel that shows the copy beside its two-stage QC result - each check passed or failed, with claim flags called out. From there: Approve, inline Edit, Regenerate, Reject or Publish. Publishing writes the approved text to the live website.

Open the inspector to see the two-stage generation trace on the uniform trace-row contract: attributes → validate and clean the brief → cache check → template select → WRITE (model A) → QC CHECK (model B) → capped retry → final draft, with provenance, model, tokens and cost per stage. A Cloud/OSS toggle and cost meter cover the two metered stages, which fail closed to the $0 OSS path at the budget cap, and a cache hit is shown honestly as a $0 stored-draft return. Recorded OSS-vs-Cloud divergences appear here too, including one honest "not captured" case. Everything is labelled synthetic and simulated.

Owners tune the brand voices, the quality rules (required region and facility, claim strictness, banned words), the per-voice-by-language prompt templates, and the generation settings (retry cap, QC-pass threshold, voice strength). Every change is audited and re-keys the generation cache, so fresh output reflects the new rules immediately. The project-page walkthrough video narrates the full tour.