PUNTO / System Architecture Diagram

01 / The generation contract

Current intent selects one of six scenarios; ambiguity asks for clarification.
The application loads skills, assembles facts, then generates three drafts.
Rules allow one repair. Provider failure or repeated violations return no approved draft set.

02 / Deployment and evidence

Docker Compose: nginx fronts one Python backend and a persistent SQLite volume.
The database container initializes or migrates, then exits; it is not a database server.
Runs and draft candidates are audit records, never messages already sent to a client.

03 / Publication and review

Static-only Studio shows a labelled saved example; it cannot generate a new answer.
Human scores are validated and imported before reports are re-exported. No ratings are assumed.
Optional batch generation uses the same service; AI judging is separate from live generation.

04 / Scope and limits

Live history is explicit input; seeded chats and automatic long-term memory are not added.
Engineering tests isolate protocol and rules. They do not establish real-model writing quality.
This map reflects current working files; source evidence is local-only. No credentials are included.