Small games about big infrastructure. Real incidents, deliberately outrageous numbers, playable in minutes.
You are the on-call engineer. Orders arrive under a countdown, every instrument is a real production knob, and the waves replay a real five-month incident — the one where five reasonable decisions starved an LLM summarizer. When the ordered knob is hardcoded, slam the big red bar. Survive to the end and run the real fix (fallbacks −98%).
Checkpoint bureaucracy at the Ministry of Ingestion. You are the LLM content classifier: stamp documents against the filters, survive the fail-open jump scare, run the two-key dry-run → apply sequence on the stored corpus, and face the coworker whose charity fun run you soft-deleted. Citations quote the real spec.
The opposite problem: what a company's memory shares back out. Documents ride a belt through gates you place. Flip a drop cartridge to SHARE, learn the hard way that an enforcing drop filter always wins — and that a disarmed one protects nothing — then calibrate the scored gate: share above, private below, ask a human in between. Teach it until it stops asking.
Two cofounders, two calls, one wall. Every red string is a grant. You explain who can see what, your cofounder takes issue with that, and the wall fills with documents, rule versions, and Temporal receipts until nobody can read it. Then you pick up the scissors: cut string for tables until the whole model is three lines. The dialogue is real.
The red-string wall became a real grant ledger. Now you have to wire it. Connect documents to their readers, keep both owners, turn an Inbox decision into a grant, and cut only the cables a filter owns. Six small puzzles about a week of Content Rules work. Every cable needs a reason.