Quake II: The Vault

Running a 1997 C engine with 17 years of somebody else's history in a browser is the smaller problem. The larger one: how do you build such a system in weeks rather than months with AI sessions without the claim "it works" standing in for a measurement? And how far can development run unattended without the machine taking over direction? An outside AI assistant read four screenshots, declared the project "fully autonomous", proposed a test and read a dashboard counter as the test passing. The test had never been run.

A gate set instead of trust: 126 Node, 173 browser and 45 C test files under sanitizers, per-file coverage floors with a ratchet that only raises from green runs, 70 audits over tree and game data, and a rule that "it looked right" is not a measurement. On top of it a loop of four lanes in their own git worktrees, fired by a timer every 300 seconds, with four guards that clear themselves: a lock, a state check, a 75-minute ceiling, STOP files. The machine invents no features: priorities 000 to 499 belong to the person, a lane's proposal becomes OPEN only by a human, and done means both CI checks green at the landed commit. The case study reads every number from git, the tree, the audits, the watchdog logs and every session transcript, and puts it into the text by script.

Live at quake2vault.app: 16 engine binaries, 8 renderers, 32 mounts, PWA, multiplayer. 392 sessions since 18 August 2026, 81.9 million output tokens and 28.4 billion cache-read tokens, $19,327 at API list price, paid for with three flat-rate seats. 98.5% of it with a person at the keyboard; the unattended loop spent 1.5% and has run for 32 hours. The case study shows: "built in three weeks with AI" holds, "fully autonomous" does not, and a counter that moves is evidence that something happened, not of what.