How It Works Now

The platform answers a question every sports fan has asked, every history reader has wondered about, and every barroom debate has chased: who would win.

The answer is knowable the way most interesting questions are knowable. With rigor, with documented sources, and with honesty about the limits of what a model can say.

Any contest where the record supports scoring can become a category. What exists today is a working sample rather than the roster: fighters, commanders, machines, explorers, athletes.

The catalogue grows by multiplication rather than addition. A competitor built once competes against everyone already there, so each addition creates as many new contests as there are opponents to meet. Every venue, century, ruleset and season multiplies it again, which is why a handful of competitors already produces thousands of distinct contests.

What a condition means changes with the category. Venue in boxing is a ring and a crowd. Venue in an expedition is two thousand miles of country.

You choose the competitors and the conditions. Venue, distance, weather, era, ruleset, whatever the category offers. Then you run it.

What you see is one simulation. Its exact times, its winner, its finishing margin or wound count. The account narrates that run and no other. Run it again and the engine produces a fresh result.

The published odds come from a separate exercise: the same engine run thousands of times to characterize the full range of outcomes. They tell you what to expect across many contests. A single run tells you what happened in this one.

A competitor favored at 67% loses a third of the time. The run you watch may be one of those losses. That is the model working.

The odds are frequencies. A competitor shown at 67% won 67% of many thousands of simulated contests under exactly the conditions you selected.

They are written the way sports odds are written. A negative number is a favorite and tells you what you would stake to win 100. A positive number is an underdog and tells you what 100 would return. So minus 250 means the outcome happens often, and plus 400 means it does not.

Change any condition and the odds change with it, because they were calculated for that configuration rather than for the matchup in general. The same two competitors carry different prices at a different venue, in a different century, or in different weather.

A price only appears where the sample behind it supports one. Where an outcome is too rare for thousands of runs to measure honestly, that option is withheld rather than shown at a number the model cannot defend.

Every number traces to a documented source.

For modern athletes, parameters come from peer-reviewed biomechanical analysis, published race data, coaching memoirs and interviews. Bolt's Berlin 2009 world record was analyzed by the IAAF with sport scientists at the German Sport University Cologne and published in Graubner and Nixdorf. Lewis's Tokyo 1991 record serves the same function.

For historical figures, parameters come from primary sources and academic scholarship. Achilles draws on the Iliad, classical scholarship on Mycenaean warfare, and the warrior tradition his myth was built on. Spartacus draws on Plutarch, Appian, archaeological evidence from the gladiatorial schools at Capua, and modern scholarship on Roman military history.

Physical models apply real physics, and what they model depends on the contest. Wind correction and altitude-coupled drag for a sprint. Surface grip and fatigue onset for a race. Disease incidence, provisioning and navigation error for an expedition.

The engine uses documented attribute ratings, dimension weights, stochastic draws within those ratings, and deterministic physics. It is a computational model built from sources, and every output traces back to them.

This is deliberate. It keeps every parameter attached to its evidence, and it handles the asymmetry between competitors honestly, since Bolt has biomechanical data and Achilles has Homer.

Each run draws random values from within documented distributions. A sprinter runs within a range shaped by physiology, training and the day. A fighter has good exchanges and bad ones within a documented baseline. The engine captures that range rather than reducing a competitor to a single number.

Achilles never fought in a Louisiana swamp. Amundsen never crossed the Darién. Columbus never sailed with a chronometer.

It runs the other way too. A Formula One car has never raced on the dirt of the Circus Maximus, and its aerodynamics are useless without tarmac to press them into. Lewis Hamilton has never driven anything that could be spooked. A modern sprinter has never run on a track laid before rubber compounds existed. Every one of them is out of their element too.

Every contest here puts somebody somewhere they were not, and that is not a flaw in the premise. It is the premise.

Why it has to be this way. A matchup that only ever ran under conditions both competitors actually met would be a very short list, and it would answer nothing. The interesting question is never who was better on the day they happened to compete. It is what each of them was made of, and the only way to find that out is to move them and watch what holds.

Put two fighters on open ground and the taller one with reach wins more often. Put the same two in dense woods and reach becomes a liability. Neither result is the true one. Both are true, and the pair of them says more about those two men than either could alone.

The same holds for machines. On a modern circuit a Formula One car beats a chariot by a margin too large to be interesting. Move the contest to a packed dirt oval with tight turns and no run-off, and the car loses the downforce it was designed around while the chariot is on the only surface it ever knew. The answer changes, and the reason it changes is the point.

That is what the platform is for.

What travels with a competitor, and what does not. The model separates three things, and keeping them apart is what makes the answer defensible.

Skill travels. A man who can read weather can read weather anywhere. A navigator who can hold a course by dead reckoning does not lose that crossing an ocean he has never seen. A driver's car control is his own, whatever he is sitting in. Judgment, nerve, physical capacity, the ability to hold a party together under strain: these belong to the competitor and they go where the competitor goes.

Equipment does not travel. It belongs to the era, not to the competitor. Choose an earlier century and the same explorer sails a smaller ship, carries cruder instruments, and has nothing that protects him from a disease nobody had yet learned to treat. He brings what he knew. The century decides what he has.

Machines carry this more sharply than people do, because a machine is equipment. A car built around aerodynamic downforce has it on tarmac and loses it on dirt, and no amount of driver skill puts it back. What the vehicle can do is a property of the vehicle and the surface together, and the model treats it that way.

Familiarity does not travel either, and this is the part that does the work. A polar explorer has no jungle knowledge, and the model does not pretend otherwise. What he has instead is a documented capacity to function in country he does not know, and that capacity is scored from his record the same way his endurance is.

How adaptation is modeled. Two competitors facing an unfamiliar environment do not fail identically, and the difference between them is real and documented. Some people improvise, reassess and find another way. Others hold to a plan that has stopped working.

So adaptability is an attribute like any other, sourced from what each competitor actually did when circumstances turned against them. And it is applied where it belongs: not as a flat bonus for being clever, but as a change in how a competitor responds once something goes wrong. A passage where nothing goes wrong is unaffected by it. A passage that turns is where the difference shows, and it compounds, so the third setback is harder than the first.

And some things cannot be adapted to. Altitude sickness at four and a half thousand metres is imposed on everybody, because nobody acclimatises by being resourceful. A ship with a keel too deep cannot enter a shallow river however good her master is. An untreated infection in 1550 does not respond to competence. A tyre compound built for tarmac finds no grip in dust, whoever is steering.

The model is careful about this line, because a simulation that lets skill solve everything has stopped being a simulation. Where the record says a thing was survivable by judgment, judgment helps. Where the record says it was not, it does not.

What this buys you. The same matchup gives different answers under different conditions, and the answers are legible. A sea captain beaten across a mountain range, a mountaineer beaten across an ocean, and a Formula One car beaten on dirt are all the model saying the same thing: the ground decides as much as the competitor does.

That is the thrill of it. Not who was better, but under what conditions, and why.

The parameters are the platform. What makes them trustworthy is the discipline around them, and that discipline is the same in every category.

Sourced beats fitted, and fitted is recorded as fitted. Where a value can be anchored to evidence outside the model it is, and where it was tuned to make the model behave, that is written down along with what it was tuned against. The ratio of sourced to fitted parameters is tracked rather than assumed.

A value that cannot be sourced is marked and kept away from anything load-bearing. Where the record is silent, the model says so rather than inventing a figure that looks like the others.

Where sources conflict, the most rigorous is used and the conflict is recorded. Two defensible numbers are a fact about the evidence, not a problem to be averaged away.

Every calibrated figure is re-measured automatically. A number that stops reproducing is reported at its new value rather than carried forward on the strength of having once been right.

A contest is measured against its own noise. Where the difference between two competitors is smaller than the run-to-run variation of either one, that is reported as a contest without a favorite rather than presented as a close one.

Structural checks run continuously. Populations are conserved. Distances agree with geography. Declared paces agree with declared durations. Every value the model carries is read by the code that runs on it.

None of this makes the model right. It makes the model checkable, and it means an error is found by the platform rather than by a reader.

Some contests resolve in seconds. Others take months.

The expedition category sends parties over the same passage and measures each against the environment. Instead of two competitors, a field of eight sets out. Instead of a winner and a margin, the result is elapsed time, survivors, and how a party ended. Some arrive. Some turn back. Some come apart.

The engine tracks each party day by day across every leg: what it eats, who it loses, whether it finds the ship, which side of an island it passes, whether fever caught in a jungle lands three weeks later at sea. The leaderboard ranks by ground covered, so the board and the map always agree.

Hazards are anchored to the places they belong to, and what a hazard costs is sourced the same way an attribute is. A snake bite is priced against what an untreated bite did to a man with no antivenom. Starvation over two weeks weakens a party rather than burying it, because that is what two weeks without food does. Where the model was wrong about this it was corrected against evidence, and the correction is recorded.

These contests run for minutes rather than seconds, and they produce an account rather than a result.

Any contest can be shared. Create a session, send the link, and everyone who joins makes their call and watches the same run at the same moment.

Calls are locked before the contest begins, so nobody sees another player's pick until the result is in.

IM Bills are play money. They exist to let you back your judgment and see how it holds up.

They are granted freely, they live inside the platform, and they are used for prediction alone. Everything on the platform is free.

Markets are offered only where they can be priced honestly. Where an outcome is too rare for the sample behind it to give a real number, that option is withheld rather than shown at a price the model cannot support. Where a market has nothing left to decide at the chosen conditions, it closes rather than offering a bet on a settled question.

Sprint events carry a two-tier wind rule that mirrors how athletics governs records.

A race run within roughly ten miles an hour of wind either way is race-legal, and the result binds for that race: winner, margin and times. A race is record-legal only within the World Athletics threshold of a 2.0 metre per second tailwind, and only record-legal results feed cross-race comparison, calibration and leaderboards. Outside the race-legal range, the race is voided.

You can run a race in extreme wind and watch the engine resolve it honestly, while the platform's published content stays clean.

The method is described here, and on the competitor profiles, which document the sources behind each rating. How a contest is modeled, what a parameter has to satisfy, and where the model declines to guess are all stated rather than left to be inferred.

The methodology is meant to be defensible to sport scientists, historians and biomechanists who care about precision in their fields. Corrections, citations and challenges are welcome. Taking the question seriously means accepting scrutiny on the answer.

Counterfactual simulation quantifies uncertainty rather than eliminating it. The model produces distributions.

Historical data carries its own measurement limits. Electronic timing was standardized late, hand-timed performances carry real error, and wind readings before the 1980s are inconsistent. The engine accounts for this where it can, and the profile flags it where it cannot.

Some competitors are documented far better than others. Bolt's Berlin race was analyzed at a depth no other sprint has matched. Achilles survives in epic poetry written centuries after his time. Each competitor is handled with the best record available, and the gaps are stated.

Calibration involves judgment even when grounded in sources. Deciding how much weight psychological stability carries against physical dominance is a defensible call rather than a derivable one, and it changes what the model says. Those judgments are documented rather than buried.

The model is a tool for thinking about the question. Its output is one informed contribution to a debate that has always been worth having.

What exists today is an engine that produces results, and a rough prototype that displays them. Those are two different things at two very different stages. The simulation is the work of months and it runs properly. What you are looking at is a working body built to hold it while the real one gets made, closer to cardboard than to coachwork.

The direction is broadcast: contests generated as watchable events rather than read as results, with the same engine underneath. That depends on technology which is improving quickly and is not there yet, so the platform is built to receive it rather than waiting for it.

What is being established now is the part that does not change. The methodology, the sourcing, and the discipline that makes an answer defensible. The rest arrives when it can.

Bolt against Lewis. Jordan against LeBron. Achilles against Spartacus. A Roman charioteer against a Formula One car. The matchups history never gave us are the ones every fan has imagined.

Impossible Matchups answers with the rigor of methodology rather than the volume of opinion. It documents its sources, surfaces its assumptions, and honors the uncertainty of competition.

That is the project. The matchups history never gave us, simulated as honestly as the documented record allows.