A speedrun is a peculiar kind of performance. The player is not there to experience a game so much as to solve it, in public, against a clock. The practice has acquired its own vocabulary, its own referees and its own annual festivals, and nearly all of that was invented by people with no authority whatsoever to invent anything, working out among themselves what ought to count.
Doom taught a community to record itself
The technical seed was unglamorous. Doom, released by id Software in 1993, could record a demo file, and crucially that file stored the player's inputs rather than the resulting picture. Recordings were therefore tiny, small enough to post to a bulletin board in an era when sharing video was out of the question, and they were verifiable, because anybody could replay the inputs through their own copy and watch the identical run unfold on their own machine.
Communities gathered around archives of those files, competing level by level and publishing tables of the best times. Collaborative projects assembled full-game runs out of separately optimised segments, distributed as demos and sometimes recut into films with commentary. Quake, with its own recording format and a fiercely inventive movement scene, carried the habit forward. An early lesson arrived alongside it: a demo replays correctly only on the engine build that produced it, so a record was always attached to one specific piece of software. The hobby has been arguing about versions and platforms ever since.
An older verification culture existed too, built around arcade high scores and the organisations that kept them, but the demo file changed the terms of proof. Evidence stopped being a photograph of a screen and became something a sceptical stranger could run again.
Inventing the rulebook
Left alone, a race to the end of a game collapses into ambiguity. Does it count if you skip half the map by walking through a wall? What about starting from a cleared save with everything unlocked? Does the timer run during loading? Communities answered by writing rulesets, and that taxonomy is arguably the discipline's real achievement.
Any per cent means reaching the ending by whatever means the rules permit. Full-completion categories fix a definition of everything and demand it. Glitchless categories forbid the exploits that make the other categories strange. Low per cent restricts what may be collected. Attempts are timed from a defined start in real time, or by the game's internal clock, or with loading removed so that different hardware can be set side by side. Single-segment runs must be performed in one sitting, while segmented runs are assembled piece by piece. A summary of how these conventions spread is kept in the general overview of the practice.
None of this is fragmentation. Each ruleset is a different sport played on identical equipment, and the rules exist so that two performances recorded years apart can be compared meaningfully. A category is a social contract, argued into existence on forums by people who then have to live inside it.
The machines that play perfectly
Tool-assisted runs sit deliberately outside the human competition. Using an emulator's frame advance, save states and re-recording, an author constructs an input sequence one frame at a time, retrying a single sixtieth of a second hundreds of times over until it comes out optimal. The product is not a video but an input file, replayed to produce the run, which is the Doom demo principle taken to its logical end. TASVideos has archived and adjudicated this work since the middle of the 2000s.
The tool-assisted world functions as a laboratory rather than a rival. It maps the theoretical floor of a game's completion time, demonstrates which routes are possible at all, and regularly produces techniques that human runners then spend months learning to approximate at a survivable rate of failure. It has also caused honest confusion, since at least one early tool-assisted performance circulated widely enough that viewers took it for live play, which is roughly when the community concluded that clear labelling could never be optional.
When a bug becomes a technique
Outsiders are often startled to learn that the fastest routes consist largely of mistakes, and that the mistakes belong to the developers. Runners clip through geometry, trigger sequence breaks that load a later area early, use wrong warps to leap over the intended order of events, and in the extreme cases manipulate a game's memory through ordinary controller inputs until the machine executes instructions nobody ever wrote.
The community's line on all this is consistent and pragmatic. If it can be produced on unmodified software running as shipped, using inputs the hardware can genuinely deliver, it belongs to the game. Modified copies, cheat devices and edited save files do not. That single distinction settles the overwhelming majority of disputes, and where it leaves real disagreement about whether the result is still recognisably the same game, the response has been to add another category rather than to legislate taste.
Proof, and the people who check it
Leaderboards are run by volunteers, and the standards they have converged on resemble a small, functioning bureaucracy. Submissions carry video, usually with audio, a visible timer, sometimes a display of the inputs themselves, and a declaration of hardware or emulator. Moderators retime from frame counts rather than trusting anybody's stopwatch. Emulated attempts are ranked separately or excluded where their timing differs measurably from original hardware.
That rigour was learned expensively. A long-standing record from the early home-console era was struck from the books in 2018 after analysis suggested the time was not achievable on the hardware in question. A heavily publicised dispute in 2020 turned on a statistical argument about random drop rates during a submitted run; the attempt was removed from the board, and the runner later stated that a modification had been active without his knowledge. Neither episode involved a court or a governing body. Both were settled by volunteers publishing their working and inviting anyone to check it, which is a sturdier institution than it sounds.
The marathon as institution
The last piece is social. Games Done Quick began in 2010 as a modest gathering of runners playing games back to back on a stream for charity. It grew into twice-yearly events raising millions of dollars per marathon for causes including cancer prevention and international medical relief, with a cumulative total across the series now running into tens of millions. Similar marathons operate in Europe and elsewhere on much the same model.
The format achieved something no leaderboard could. It put the sofa back into the room, with a runner performing while friends explain the tricks to an audience who have never touched the controller, and it turned an activity built on solitary repetition into something worth watching for hours at a stretch. That, more than any individual record, is what carried the pursuit out of the archives and into general culture. Underneath all of it sits a simple and slightly obsessive proposition: study something closely enough, for long enough, and you stop being its audience and start being its scholar.
