One element at a time
New mechanics arrive one by one: a pressure, a lever, a pathway. Difficulty comes from how simple elements combine, never from information you were not given.
Oversight · A game about governing AI
The hardest job in AI is keeping it good.
A story-driven campaign where governance is the gameplay.
You run the desk at an office that just adopted an AI system. The AI is genuinely useful, and error can spread through it: from outputs to operators, into the records, and back out again. You have a limited budget, a catalog of governance levers, and pressure that never quite lets up. Your job is to assemble a stack that holds.
Oversight is fiction. The systems in it are invented, and its effects are authored for play, not cited to sources. But it is built on the same engine as the PAN Lab, and it teaches the same lesson the evidence teaches: safety is a property of the whole deployed system, not of the model alone.
The curriculum
Simple elements. Real combinations.
Containment is half the job
The design principle: an AI that is safe but helping no one is not a success. Winning means the system is both contained and genuinely doing the work.
Controls rot
Checks that get rubber-stamped, alarms that get tuned out, training that becomes theater. The game makes the slow decay of oversight, the thing that actually kills governance regimes, playable.
No guessing games
Every level is machine-verified winnable at every difficulty, with multiple distinct winning approaches. If a level feels impossible, that is a design bug we fix, not a trick we played.
Under the hood
A game you can audit.
Before any level ships, an automated judge enumerates its entire legal move space, every combination of levers, placements, and strengths a player could buy, and verifies it is winnable, measures how scarce the solutions are, and scores its difficulty against explicit per-act bands. The judge itself is frozen and version-fingerprinted: if its code changes, every shipped level is automatically flagged for re-verification.
Story text is checked against the engine too. If the copy says a lever is not enough on its own, an audit replays that claim through the judge and fails the build if it is false. We built a game this way for the same reason we build everything this way: teaching people to govern AI systems is a trust business.
The engine and the disciplineOversight is how the public meets the work. The PAN Lab is where the work meets the evidence. The practice is where the work meets your organization.