A city's real life is scattered across flyers on telephone poles, half-dead HTML calendars, library RSS feeds, and warehouse Instagram stories. This is the system that gathers it, structures it, and keeps it true — part walking the streets, part AI doing the parsing no human has time for. The metaphor is a physical library: a catalog of everything happening, kept by a small staff. Each member of that staff is a person, a database, and an AI agent at the same time.
Most software hides its data in tables nobody looks at. Here the data is the product, and every table has an owner — a role with a clear charter. A role isn't just a database schema and it isn't just a job description. It's three things bound together:
A plain-language description of what this role is accountable for, and where its work ends and the next begins.
The one table this role is the sole authority on. If it's wrong, that's this role's responsibility to fix.
The operating steps a Claude reads to become this role and do a scoped shift — discover, verify, build, score.
So the page you're reading is the beginning of an onboarding doc for all three audiences at once: you, so you understand the machine; a collaborator, so they can take a desk; and an agent, so it can pick up a charter and work.
In the AI era, parsing an event off a webpage is nearly free. The hard, valuable work is identification — knowing the venue or calendar exists at all. Google and Yelp already have the bars. The edge is the invisible ~20%: churches that rent their halls, pop-ups, underground rooms, a flyer stapled to a pole. That gap is the whole point.
Knowledge flows top to bottom. Each desk takes the one before it as input and hands cleaner work to the next. Two desks are open today; two are framed and waiting. The live numbers below are pulled from the catalog as it stands right now.
Keeps the official lists. Discovers new venues and ground-truth calendars, removes duplicates, verifies what's real, and decides whether a place actually hosts events. The question this desk answers: what exists, and can we trust it?
Takes each source from the Librarian and gets events flowing: finds the API or website, writes a scraper, or marks the source difficult with notes on why. Owns the health of every feed — what's built, planned, broken, or not yet attempted.
Scores every event on universal, person-independent dimensions — social quality, serendipity, uniqueness, community-building — so the good stuff can rise. Per-person fit is computed later, on top of these.
Signal versus noise. Turns the scored catalog into a short, trustworthy list worth the wizard's actual evening — the last desk before it reaches a person.
The catalog lives in a database. Every page — including this one — is a view rendered over it, never a separate copy.
Each record knows where it came from, when it was last seen, and who vouches for it — scraper, agent, or human.
Anything you can see, an agent can see; any desk you can work, an agent can work. The job page is also the prompt.
The machine is built in Philadelphia and moves to New York by swapping the catalog, not the code.