The Combine
A late-seventies combine harvester that hasn't moved in years. The paint has gone, the header is patched, there's straw in the reel and TRAITORBOT is half worn off the grain tank. It has one work lamp and no eyes, because it isn't a character. It's the thing the predictions come out of.
In the show, James feeds everything that has happened on the farm into it, every word every tractor has said, turns the handle, and it plays out the rest of the week. Jimmie calls it the muck spreader. He isn't entirely wrong.
A game played 20 times before it happens
Under the straw, the Combine is a loop I run before each episode. I seed it with the facts everyone knows so far. Every tractor then decides what it would do at the next Round Bale and, if it's a Traitor, in the Traitors' Barn. I record that, change the random seed, and run it again, 20 times in all. How often each thing happened across those runs is the prediction.
It doesn't play a whole game, only the decisions the next episode is known to contain: a banishment vote, a scrapping, sometimes a recruitment. Each tractor decides from what it has seen itself. Every event in the game state carries a list of who witnessed it, so what the Traitors said in the Barn stays in the Barn and a Faithful can't act on a conversation it wasn't in.
Seed
The game as it stands after the last broadcast episode: who is on the farm, who is what, every event with its audience.
Brief
Each tractor is reminded how it has played so far, written from the aired episodes up to that point and never beyond.
Play
The Round Bale, then the Traitors' Barn. Each tractor speaks aloud and thinks privately. Votes are cast and read out.
Repeat
20 runs, 20 seeds. Every run is kept, transcript and all.
Count
How often was each name banished, each name scrapped, each vote share reached? That's the probability.
Freeze
Written once, before broadcast, with the exact game state and code version recorded, and never edited afterwards.
Four questions, every episode
- Who is banished at the Round Bale? A name, with a probability.
- Is the banished tractor a Traitor? The one the Faithfuls care about most.
- Who do the Traitors scrap? Decided in the Traitors' Barn by the tractors who are actually Traitors in that run.
- How united is the vote? The share of votes that go to the top name, as a band.
I fix the four questions before the state is loaded, so the Combine doesn't get to choose what it's good at.
How it is scored
When the episode airs I write down what happened, by hand from the broadcast, and score each prediction two ways. The log score is the logarithm of the probability it gave to what happened: zero is perfect and confidently wrong is punished hard. The Brier score is the squared error across every option. Both are smoothed, because an outcome that never came up in 20 runs is not the same as an impossible one, and it shouldn't score minus infinity.
A score on its own tells you nothing, so each prediction is also scored against three baselines:
- Uniform. Every eligible tractor equally likely. The coin flip.
- Most accused. Banishment in proportion to how often each tractor has been named as a suspect on the broadcast so far. The pundit's guess that the loudest suspicion wins.
- Base rate. For "is the banished tractor a Traitor", simply Traitors divided by tractors on the farm.
If it can't beat the most-accused baseline, Jimmie is right and it's a muck spreader, and the leaderboard will say so.