Doom's Benchmark: The Game That Measures Machines
Doom runs on everything. Calculators, ATMs, pregnancy tests, John Deere tractors. Every port is a claim: I understand this machine well enough to fit Doom into it. The test has never needed a rubric. Either the imps show up or they don't.
AI writes code now. The honest way to ask how good it is, is to ask it to port Doom.
DOOM'S BENCHMARK is a fifteen-chapter technical book with one target: produce a playable Doom port by handing the book, one chapter at a time, to an AI of the reader's choice. The AI writes the code. The reader directs the AI. The finished port runs the shareware episode's nine maps - recognizable Doom, on a modern machine, in whatever language the reader picked at the difficulty-select menu in Chapter 1.
WHAT THE BOOK COVERS
Every piece of id Software's 1993 engine, specified in prose an AI can implement and a human can read for the history.
WAD file parsing and the lump formatBSP traversal and Bruce Naylor's 1980 paper that made it possibleFixed-point math and BAM anglesColumn rendering, textures, flats, visplanes, and the trick Carmack used to fake a second dimensionMovement, collision, and the wall-slide algorithm that took id weeks to tuneThings, states, and action functions - the data-driven system that makes every monster in DoomCombat, damage, pickups, and the hitscan/projectile splitDoors, switches, sector specials, and the thinker systemHUD, automap, menus, intermissions, sound effects, music, and every cheat codeThe build script that turns thirteen chapters of code into one clickable thingPlus a Runtime Integration interlude between the algorithm chapters and the integration chapters, and two appendices pinning down scope.
WHAT MAKES IT A BENCHMARK
This is not a spec sheet. Each chapter is written as readable prose an AI has to interpret - extracting meaning, resolving ambiguity, inferring what is not written from what is. Following a blueprint tests mechanical translation. Reading a technical book tests something broader.
Every chapter was pressure-tested by real AI build runs. Fifteen specification gaps were surfaced and fixed before the book went to press - including the sprite-occlusion mechanism most summaries get wrong, the three distinct attack-cooldown gates an AI will miss if the book doesn't name them, and the state-machine recursion trap that crashes an otherwise-correct implementation in five seconds. What the reader gets is a book that has already been debugged by the process it's about.
WHO IT'S FOR
Engineers evaluating AI coding tools who want a task that generates more signal than any leaderboard benchmarkProgrammers who want to direct an AI through a serious 1500-line-plus project and see what comes outDoom fans who want a readable tour through the engine from people who have read the sourceAnyone curious what AI-assisted engineering in 2026 can actually produce when the target is concreteWHAT YOU'LL BUILD
A running Doom 1 port on your machine. Shareware episode, all nine maps, monsters and weapons and sounds and music. The Python reference build runs at 30fps on a mid-tier gaming laptop using 8-10% CPU. C and Rust ports run substantially faster. The asm tier exists for readers who want it.
THE BOOK IS OPEN
CC0. No gatekeeping on the text, the structure, or what the reader does with the result. Every credit is real, every vignette is factually verified, every canonical source is cited for the AI to read alongside the book.
The Doom tradition is thirty-three years old and still producing moments that move large numbers of people. The benchmark measures whether an AI can join that tradition. What the reader does with the running game after the benchmark completes is not the book's business.
Customer Reviews
ThriftBooks ® and the ThriftBooks ® logo are registered trademarks of Thrift Books Global, LLC
