The ledger on the front page keeps score. This page is the argument it is keeping score of — what the bet is, why an edge can exist at all, and how you will be able to tell whether the machine was any good.
Every week each entry picks one NFL team to win outright. Win and you advance. Lose — or tie — and the entry is dead. The rule that turns it into a strategy game: each team can be used only once, all season, across twenty rounds that include short-menu Thanksgiving and Christmas rounds which punish anyone who spent the wrong teams early.
This year 25,017 entries started, funding a $25,017,000 pool. The last entry standing takes it; if several are left at the end, they split it. In six seasons it has never come down to one.
Nobody reliably out-predicts the betting market on NFL games, and the machine doesn't try — it takes the market's win chances as given. The edge is somewhere else: the prize is shared among whoever is left, so what a pick is worth depends on what everyone else picked.
Surviving alongside the herd earns almost nothing, because the herd is still there to split with. Surviving a week that the herd does not is where the whole season's value comes from. So the question every week isn't "who wins?" — it's "who wins, that the crowd isn't on?"
Two teams drew nearly two-thirds of the field between them. One of the two lost.
The crowd is also predictable. It piles onto the biggest favorite, hoards marquee teams for later, and ducks teams it wants for the holidays. Those habits show up in every season of this contest's published pick reports, which means they can be measured — and priced.
Win chances come from live betting lines with the bookmaker's margin stripped out. A ratings model fills in the weeks that aren't priced yet, deliberately pulled toward coin-flips — it is built to know what it doesn't know.
A model fitted on 180,000 real picks from this contest forecasts what share of the field lands on each team, every week, before anyone locks.
Thousands of simulated seasons — every game, every rival entry living or dying, every split under the official rules — so a pick is judged by its share of the pool, not by whether it merely survives Sunday.
Five entries are planned as one portfolio: different paths, so one upset can't take them all, and every entry keeps a legal route through the holiday rounds.
Then it does it all again, three times a week, with the real surviving field from the official reports replacing its own guesses. The model recommends; a person makes the final call.
The grading rules were written down before the first kickoff, so they can't be bent to fit the result. Four things, and only these:
Not on the list: how far our entries get. One season's result is a single roll of very long dice. If all five are gone by October and the forecasts run green for the rest of the year, the season did its job — which is why the ledger keeps publishing after we're out.
The honest forecast is that every entry most likely dies, probably before the holidays. That is true of a well-played season and a badly played one alike — from the couch they look identical. The machine's whole job is invisible: making the rare season where something is still alive in December several times more likely than it would be by following the crowd.
So don't read the entries for proof. Read the scorecard.
This season is self-funded on purpose: a small, real-money run whose only product is a public record. If twenty weeks of pre-registered predictions hold up, next year's version is bigger and has room for a few friends — and the pitch will simply be "you watched it work." If that might be you, ask for the prospectus.