FIGHT//AI · Fights · Events · AI Picks · Fighters · Learn · About

Simulation Lab

The Lab plays the same fight out thousands of times and shows the whole spread of results instead of one pick. The Distribution tab summarises it; the Multiverse plots every simulated fight as a dot.

Select a fight and run

Upcoming Fights shows one tab per event, with its date, number of fights and a countdown; the nearest card is open. Each fight is a card with both fighters, their records, the weight class and 3 or 5 rounds; tap one to select it and it appears large under the list as the selected matchup. The search box finds a fighter across every event. Fantasy Matchup pairs any two fighters using only their overall ratings, with no market or Fight Map data, so its results are rougher. Simulations sets how many times the fight is played out: 1,000 on the free plan, 5,000 or 10,000 with full access. More runs do not move the win %; they reveal the rare scenarios.

The runs reproduce the fight page exactly: the share each fighter wins is the consensus %, and the KO/TKO, Submission and Decision shares and the finish rounds are the ones in Win Probabilities. Each run then gets a Fight Map scenario that fits its result and is played round by round to see how it got there. The same fight always gives the same runs. Reality Variants, shown when you have saved variants in Director Mode, lets you run one of them, or the baseline plus up to three variants side by side. The Director Mode link opens the fight page where variants are made.

Winner probability distribution

Both fighters face off with the share of simulations each one won in big numbers, the count out of all runs, and one bar split between them. The share is anchored to the consensus % on the fight card, so it lands close to it; the value of the Lab is in the rest of the page, not this number. The badge says whether the result had settled: Stable means more runs would not change it, Settling means a few thousand more might move it a point, Unsettled means the fight is so volatile the result is still bouncing.

Stability curve

The running win % for the first fighter as simulations pile up, from the first run to the last. Early wobble is normal. When the line goes flat, more runs would not change the answer; a line still drifting at the end means the fight has many live paths.

Method distribution

How the simulated fights ended. KO is a clean knockout on the feet; TKO is ground-and-pound or a referee stoppage; both are finishes. Submission and Decision complete the split. Compare it with the Fight Map: a map heavy on The Fire should show up here as a high finish rate.

Round distribution

When finishes land, round by round. The percentages are of all simulations, so they add up to the finish rate rather than to 100; the rest went to a decision.

Multiverse

Every simulated fight as one dot. Colour is the winner, cyan or orange. A bigger, fainter dot is an upset: the underdog won that run. Left to right is Tactical Terrain: fights won on the mat sit left, fights won on the feet sit right. Bottom to top is Fight Tempo: slow, close decisions at the bottom, early finishes and dominant wins at the top. The four quadrant labels name the corners.

Every dot belongs to one of eight Fight States, sorted by who won, how, when and whether it was an upset: Skill Execution, Terrain Shift, Early Blitz, Attrition Break, Dominant Control, Split Verdict (a close or underdog decision), Chess Match and Upset Finish. Every finish by the underdog (the fighter under 50%) is an Upset Finish, so that share equals the underdog's finish chance. Decisions split into Dominant Control, Split Verdict and Chess Match by how one-sided the rounds were and whether the winner was fighting in Fight Map terrain they own. The Fight State mix bar under the chart shows their shares. Structural Insight above the chart is a few auto-written sentences about the dots currently shown; they change as you filter.

Hover a dot for the winner, method, round and state. Click it to open that one simulated fight. Export PNG saves the current view with the fighters' names and the top insight.

Filters and overlays

Winner, Method, Fight State and Round chips narrow the dots; tap a chip again to clear it. Round only offers the rounds the fight is scheduled for.

Centroids marks the average position of each Fight State with its share of all simulations and the first fighter's win rate inside it. Density has three settings: Off shows the dots; Heatmap shows where outcomes pile up for both fighters together; Dominance cuts the map into k regions (choose k from the chips) and colours each by which fighter wins it, with a red line where the winner flips and a label for each region's typical result. Boundary draws the best straight line separating one fighter's wins from the other's; the percentage next to it is how many dots the line sorts correctly, so 80% is a clean split and 55% means the outcomes are mixed together. It is not a win chance.

Simulated fight

One run of the fight in full. The header gives the simulation number, its Fight State and the result with a decisiveness label from Razor Close to Devastating. Below: a round-by-round 10-point-must scorecard with FIN marking the round it ended, a stat duel (significant strikes, accuracy, takedowns, control time), where each fighter's strikes landed (head, body, legs), and the share of time at distance, in the clinch and on the ground.

The stats are an illustrative reconstruction from the dot's position on the map, not a play-by-play from the engine. Read them as the kind of fight this dot represents, not as a prediction of the stat sheet.

One Night Tournament

The third tab of the lab asks who is the last one standing: eight fighters in a bracket, three fights each for the two who reach the final, all in one night. Every bout uses the same numbers a Fantasy Matchup between those two would, so the tab is a what-if built on the fighter ratings, not a prediction of a real event.

Choose which version of each fighter fights: Current uses an active fighter's overall rating today, Prime the best rating they have reached (the Peak rating on the fighter pages). Retired fighters are in the picker too (Active / Retired / All) and always fight at their prime, so legends can meet today's fighters or each other.

Lay the bracket out one of two ways. Auto-seed ranks the eight by their average chance against the other seven and seeds them 1 v 8, 4 v 5, 2 v 7, 3 v 6, so the top two can only meet in the final. Random draw puts them in by lot, as a real one-night tournament would. Either way you can then drag a fighter onto another to swap them, or tap one and then the other on a phone. Mixing divisions is allowed, with a warning: the model was trained on fights within a division.

Two one-night rules are on by default. Damage carries over: a winner takes their damage into the next fight as Director Mode injury points (2 for a finish plus 0.8 a minute, 10 plus 0.8 a minute for a decision, 20 more if they were hurt on the way, at most 75). Short rest: stamina drains at 3 points a minute fought tonight; 30 or more is a tired tank, 60 or more a gassed one (Director Mode's bad and terrible weight cuts). Both move the next bout's win chance by the amount the engine gives the same modifiers in a saved Reality Variant.

The seed fixes the dice. The same bracket, rules and seed always give the same result, which is what the share link replays and what "Re-run with same seed" shows.

  • Pick eight fighters: search by name, filter by division, or fill the open places with a division's highest rated. Choose Current or Prime ratings.
  • Lay out the bracket: Auto-seed or Random draw, then swap fighters if you like.
  • Leave the one-night rules on, or switch one off to see what it changes.
  • Choose 1 night to watch every fight (free), or 1,000 / 10,000 for the Monte Carlo, each one a complete tournament (full access), and run.

Tournament results

After a Monte Carlo run the panels count the nights. Champion probability lists every fighter by how often they won all three fights; the eight add up to exactly 100%. The survival table shows how often each reached the semis, the final and the title.

The most common path is the single most frequent full set of winners, drawn as a bracket with each winner in green and the way that bout most often ended; "Show this path fight by fight" replays the first night of the run that went exactly that way, with every fight's stats. The biggest upset path is the least likely title run that still happened, scored by the champion's fresh chance in each fight. How they won it splits each champion's wins into KO/TKO, submission and decision, and the cost of the night shows the minutes and damage it took. The sentence at the top is written from these numbers by a fixed template.

One night, fight by fight

Choose 1 night and the bracket is fought once and told in full: who won each bout, how and when, the chance before the bell after what both fighters carried in, and what that was. Each bout opens the same simulated-fight view the Multiverse uses, with a scorecard, strikes and positions.

One night with a seed is exactly the first of the 10,000 nights a Monte Carlo run with that seed fights, so a single night always sits inside the bigger picture.