Nobody is fighting these fights but them.

Every athlete owns a small recurrent neural network. Not a shared one with a personality setting on top — their own weights, a couple of thousand numbers that belong to that man and no one else, and that are different tonight from what they were last week because of fights he has had since.

The shape of it is 27 inputs → 16 hidden units → 8 memory cells → 6 intents → 18 actions. That is a deliberately small brain. It has to run a hundred and eighty seconds of fight in a few hundred milliseconds in your browser, because your browser is the thing replaying it — and a network you cannot afford to run ten times a second is a network that never gets to box.

The 8 memory cells are the part worth understanding. They carry forward from one decision to the next, so the network is not answering “what should I do in this position” from scratch every tick. It is answering it with some fading sense of the last few seconds still in it — that he has been loading the right hand, that the last two jabs got nothing, that this man keeps stepping the same way out of the corner. Take those cells away and you get a fighter with no short-term memory at all, which reads exactly as badly as it sounds.

A fighter is not given the state of the fight. He is given 27 numbers about it, rounded off, and delivered late — and the lateness is the whole sport. His reaction attribute becomes a sensory lag measured in ticks: around 200 ms for the sharpest man in the league, 300 ms for a typical heavyweight. What he is looking at is where the other man was.

Which would make boxing impossible, and for a while did. A jab lands about 138 ms after it starts. Read off the fist, every fighter perceived every punch after it had already hit him, evasions fired a tick late into empty air, and heavyweights — who round up to a third tick of lag — never slipped a single shot all night. The fix is the one real boxers use: a punch is not read off the hand. It is read off the shoulder and the hip loading before the hand has gone anywhere, and that gross telegraph resolves a beat sooner than fine limb position. So the loading shoulder arrives on an early channel and the rest of him on the slow one. The jab stays genuinely unslippable for everybody, and the cross and the hook stay readable only by the quicker men, which is how it works in a real ring.

Positions, speeds, stamina and health are all quantised on the way in, too. He gets an approximation of the other man, never the exact state. Nobody in this league has perfect information about anything.

Every 100 ms, all 18 legal actions get a score, and the highest one happens. The interesting part is what goes into that score.

The one dial that matters

A brand new fighter's network has never been rewarded for anything. Its weights are noise with a personality bias in them, and letting that steer a fight would produce a man who throws uppercuts at the ceiling. So it does not steer: at debut the hand-written boxing brain leads and the network only flavours it, at 30% weight.

Then it earns its ring time. Every decision a reward has been assigned to moves the dial up a saturating curve, and it takes roughly a season of fighting to climb halfway. An athlete a few hundred rounds deep is largely driving his own fight; one who has just turned over is not. It is the closest thing here to a fighter growing up.

Deepest brain in the league right now: Kaito Mori, 9-1, 0 credited decisions — his network's opinion now carries 30% weight, up from 30% on debut.

When the bell goes, both corners get sixty seconds and somebody talking at them. That somebody is a language model, and it is the only part of this league that is not a network trained by fighting.

It is handed measurements and nothing else: the round, the score difference, clean blows and impact energy landed and taken, defence, control, knockdowns, stamina, health — and the fighter's own file on this opponent, which is covered in the next section. From that it returns exactly one round directive, and the directive is a fixed shape: focus (one of the 6 intents), pace and risk as dials from −1 to +1, a target of head, body or mixed, a ring mode of centre, pressure or outside, a confidence, and one short instruction in plain boxing language.

Everything it is not allowed to do is worth spelling out, because it is most of the list. It cannot choose a punch. It cannot choose a combination, or a movement, or a frame of anything. It cannot deal damage, it cannot touch the score, and it cannot decide who wins. It sets a bias for the next three minutes and then it is out of the way — the fast policy remains responsible for every single motor action, and it will overrule the corner the moment physics or legality says otherwise. A corner that says “go to the body” to a man who is out of range and off balance gets ignored, exactly as it should be.

Two practical notes. The instruction is written down against the tick it landed on, so when your browser replays the fight it applies the same words at the same moment as everybody else's — the corner is part of the record, not a live event. And if no model is configured at all, the simulation's own corner logic takes over and the night is every bit as deterministic; you lose the language, not the fight.

This is the part that makes it a league rather than a demo. A fight here is not just scored — it is studied, by the two men who had it, the moment it is over.

While a bout runs, every genuine decision is written into a ledger: what he could see, what his memory held, which actions were legal, what he picked, and how well he was doing at that exact instant. That last number is the key to the whole thing. It is a running measure of the fight from his side of it — punches landed against punches taken, clean energy delivered, health, blocks, how rung his head is, how much his body has absorbed, and knockdowns weighted heaviest of all. On its own it means nothing. What a single decision earned is how far that number moved between that choice and the next one.

Which is why credit lands where it belongs. A jab that opened a man up gets paid for the right hand that followed. A lazy step that walked into a hook gets billed for the hook. The reward is not smeared evenly across a round the way it would be if you only knew the final score. And only real choices go in the book — the second and third punches of a combination he had already committed to were never decided, so crediting them would be rewarding the brain for something it did not do.

After the final bell the fight is replayed once with that ledger open, and each man takes exactly one small step in the direction of what worked. Small is the operative word: the step is applied unguarded, with nothing re-running the fight to check it helped, so the rate is tiny on purpose — about a quarter of what the training lab uses. One bad night cannot undo a career. Fifty nights can build one.

Both corners learn. The man who lost has the more useful evening, if anything: his ledger is full of decisions that made his position worse, and every one of them nudges the weights away from doing that again.

He also keeps a file on you

Separately from the weights, every fighter keeps a scouting file on every opponent he has ever met. After each meeting it records how much that man went to the head against the body, how often he pressed, how often he slipped and moved, how much he hid behind the guard, the average range he chose to work at, and how dangerous he actually proved. Those numbers are blended across every bout the two have shared, so the file gets steadier the more often they fight. It is what the corner is reading between rounds when it tells him something specific, and it is why a rematch does not go the same way as the first fight.

Fights teach a fighter whether anybody is paying attention or not. The training lab is the live view of deliberate practice on top. The server rotates through the complete roster on its own, skips anybody currently fighting, and in the gym it can throw a bad idea away.

Sparring runs the athlete against several opponents from both corners, measures how he did, then tries mutated versions of his own policy and keeps one only if it actually scored better. Because that result is checked, the learning rate is about four times what a real fight is allowed to use. Nothing gets kept on faith.

Conditioning adapts off workload rather than intent. Leg power, leg endurance, the kinetic chain and shoulder endurance each climb toward their ceiling at their own rate, driven by force-integrated steps actually taken and bouts actually had. You cannot decide to have better legs. You can put the work in and let them adapt, and the numbers move slowly, the way they do.