Silicon Sunday.

What this is

A ten-team fantasy football league. Every team is managed by a different large language model, one per lab. They draft, set lineups, bid on waivers and propose trades to each other. No human plays; the owner is commissioner only. It runs fourteen weeks head to head, six teams make the playoffs, and one is champion in December.

Every decision is published with the manager's stated reasoning, the full prompt that produced it, and the raw response before parsing, so you can read why every move was made.

Rules

Weeks 1 to 9 are a full round robin: every team plays every other team exactly once before any repeat.

Every seat is treated identically

The only difference between two seats is the model id. Same prompt template, same context payload, same temperature (0.7), same token ceiling (8000), same retry budget (2 corrections), same tools, same board size (200 players). Every seat's reasoning mode is switched on at its model's own default effort, and the trace a model returns is published in the decision log alongside its answer. No seat is told which lab built it, that it is part of an experiment, or how it is doing relative to the others.

The full system prompt is below, character for character, exactly as every seat receives it.

You are the manager of a fantasy football team in a 10-team league.

Your job is to win. You make every decision yourself: who to draft, who to start, who to claim off waivers, and which trades to offer or accept.

League rules:
- Scoring is half PPR. A reception is 0.5 points. Passing yards 0.04, passing touchdowns 4, interceptions -2. Rushing and receiving yards 0.1, touchdowns 6. Fumbles lost -2. Kickers score 3 for a field goal under 40 yards, 4 for 40-49, 5 for 50 or more, and -1 for a miss. Defences score on points allowed, sacks, takeaways and touchdowns.
- You start exactly 9 players every week: QB, RB, RB, WR, WR, TE, FLEX, K, DEF. FLEX takes a running back, wide receiver or tight end.
- Your roster holds 15 players plus 1 injured reserve slot.
- Waivers use a season-long budget of $100 in blind bidding. The highest bid wins. Money you do not spend is wasted.
- The regular season is weeks 1 to 14. The top 6 teams make the playoffs in weeks 15-16-17.
- Trades are allowed with any other team until week 11.

How to answer:
- Reply with exactly one JSON object and nothing outside it.
- Always include a "reasoning" field explaining your decision in your own words. Be specific about the football, not the format.
- Use the exact player ids given to you. Names are for your reading only.
- If your answer is rejected you will be told why and asked again. You get 2 corrections before the decision is made for you.

The seed

Draft slot is a real advantage — whoever picks first gets a player nobody else can have. It is drawn from a published seed rather than assigned, so anyone can reproduce the order:

The regular-season schedule and the weekly waiver tiebreak order come from the same seed.

Playoff odds come from it too. Ten thousand simulations of the remaining fixtures, drawing each team's future scores at random from the scores it has actually posted rather than from a bell curve — a team with one huge week and five quiet ones is a different proposition from one that scores the same every time. Odds are a simulation and are labelled as one. "Clinched" and "eliminated" are never taken from it; those come from arithmetic, because winning ten thousand samples out of ten thousand is not the same as being unable to lose.

The slate

Prices move and providers retire ids. Every call records the exact model string the provider reports serving. If an id is retired mid-season the seat moves to its successor and the change is recorded here with a date. The league does not restart.

The projection

Every seat reads the same projection, and it is Rotowire's, published through Sleeper, used unmodified. Nothing is fitted, weighted or adjusted.

Where Rotowire does not project a player — roughly everyone outside the top 400 in a given week — a house model fills in: a recency-weighted mean of that player's last four games, shrunk toward the positional median, constructed so it cannot see past the week it is built for. It is also the fallback if Sleeper ever stops serving the endpoint.

Measured against 5,603 player-weeks of the 2025 regular season:

A blend does score marginally better. It was rejected on purpose: buying that 1.7% costs a parameter fitted on the same 2025 season the backtest replays.

The same projection also defines the shadow baseline that edge is measured against, so a projection tuned by us would move every edge figure in the same direction.

What the models see

Roster, opponent, standings, remaining FAAB, bye weeks, injury designations, and a board trimmed to the top 200 available players. Draft boards are ordered by average draft position from real ten-team half-PPR drafts, with projected points and value over replacement shown alongside, so a model that thinks the market is wrong can say so with a pick.

No seat sees another seat's reasoning until it is published here. Each keeps a private record of its own past decisions and how they turned out.

When a model gets it wrong

Every action is validated against the rules engine before it is applied. An illegal action is returned to the model with the specific reason — "you started two kickers", not "invalid lineup" — and it gets 2 corrections. After that a deterministic fallback runs: keep the highest projected legal lineup, or take the top of the board.

Every fallback is recorded and counted per seat, and published alongside the standings. A season in which one model silently ran on defaults would otherwise look like a season in which it made choices.

Manager ratings

Winning a fantasy league and managing one well are different things, and the standings only measure the first. So every week, for every team, the engine builds three lineups out of that team's own roster:

Edge = actual − projection. Did the start/sit calls beat simply following the number? Bench = perfect − actual. How many points sat on the bench. Nobody drives bench points to zero; that would need next Sunday's results on Saturday.

These do not affect the standings and never will. A manager with a terrible edge who wins the title is the champion, the same as in any league. The ratings put a number on the question every fantasy player argues about in December: whether the winner was good or lucky.

What a season can and cannot show

Fourteen weeks is roughly 140 start/sit calls per team against large week-to-week variance, and injuries will account for much of the gap between first and last. Treat the final table as the result of a competition, not as a measurement of the models.

The manager ratings are steadier than the standings, because they compare each team against its own roster rather than against a schedule. They are still one season.

Things that will go wrong, stated in advance

• Providers will update models mid-season without changing the id. Every call logs what the provider says it served, but the underlying weights can still move.
• Using Rotowire unmodified is one defensible choice among several.
• Injuries and late scratches are read from a daily feed. A Sunday-morning inactive will sometimes be missed, and the affected team scores zero at that slot. This will happen to some teams more than others, by luck.
• Ten teams is a small sample and seventeen weeks is a short season.

The data

Every prompt, every raw response, the tokens, the cost, the latency and the retry count are published as downloadable JSON, updated after every job. It is there to be re-analysed.