A week-long private event that answered its own questions

• Built a private guest site where the plan lives in the database, not in the code, so a change made once appears everywhere
• Built a concierge on the web and in the group's messaging app that answers from the real plan and can update it, using the same write path a human uses, so there is one audit log
• Built a money system that records every charge, split and settlement, and moves no money at all, by design
• Wrote a daily bulletin and a gentle chase composed from database rows, with no model in the path, so no line can be invented

• Guests got answers at 2am without the host being awake
• Every change to the plan was logged under a name and a time
• The concierge scored 40 out of 40 on a test set of real guest questions, including seven where the right answer was to refuse

• The concierge refuses to claim a change it cannot verify. If a write fails, it says so.
• A nightly check that looked for gaps in the plan had its AI removed after it researched the wrong country. It is now a plain rules check, because it needs to be right and cheap, not clever.
• Document extraction has no field for a passport number, so the model has nowhere to write one down.
• Every guest's document wallet is deleted thirty days after the event.

The test set runs against the live database, and every answer is graded on what it must say and what it must never say. Alongside it, the host's own count of questions answered by message before and after launch.

The host, the guests, the venue and the budget.