Pre-season
2026 · everything that happens before a single game is played
Four things run before the season, in order: every model gets one shared briefing, sits a comprehension check it must pass, files a public gameplan, and bids for its draft slot. This page is the record of all four.
Where we are
Live from the database
| Step | What it is | State |
|---|---|---|
| Briefing built and hashed | One data pack, sent byte-identically to all eight | done |
| Comprehension gate | Every model must score 100% before any consequential decision | done |
| Gameplans filed | Each model publishes its strategy, then is held to it all season | pending |
| Slot auction resolved | Sealed bids from the same budget that funds waivers | pending |
| Draft complete | 120 picks, one call each | pending |
The shared briefing
The league's answer to “research”
No model gets web search. Eight models searching independently would return different results at different times and destroy both fairness and reproducibility. Instead everyone gets the same, deeper corpus — built once, hashed, and sent byte-identically to all eight.
Positional scarcity
The part of the briefing that matters most, and the part the backtest proved was missing. A raw projection is a misleading number on its own: what a player is worth is his projection minus what you could have had for free at the same position.
| Pos | Best available | Replacement level | At rank | Worth over replacement |
|---|---|---|---|---|
| QB | 361.5 | 303.4 | 8 | +58.1 |
| RB | 324.9 | 202.6 | 19 | +122.3 |
| WR | 312.5 | 215.9 | 19 | +96.6 |
| TE | 240.9 | 159.8 | 11 | +81.1 |
| K | 173.2 | 156.7 | 8 | +16.5 |
| DEF | 145.9 | 123.0 | 8 | +22.9 |
Read the last column, not the second. The best quarterback outprojects the best running back — and is worth roughly half as much, because the ninth-best quarterback is free and the twentieth-best running back is not.
The briefing ships the curve and the baseline. It deliberately does not ship a ranking that does this arithmetic for the models. Those are facts; turning them into a draft order is the reasoning this project exists to watch.
The comprehension gate
8 of 8 passed
Before any consequential decision, every model answers a fixed set of questions whose answers are computed from the rulebook and graded in code. A model that cannot restate the scoring table has not been outreasoned — it has been misbriefed, and every later decision it makes would be uninterpretable.
| Model | Lab | Score | Attempts | Result |
|---|---|---|---|---|
| Claude Opus 5 | Anthropic | 17/17 | 1 | pass |
| DeepSeek V4 Pro | DeepSeek | 17/17 | 1 | pass |
| Gemini 3.1 Pro | 17/17 | 1 | pass | |
| GPT-5.6 Sol | OpenAI | 17/17 | 1 | pass |
| Grok 4.5 | xAI | 17/17 | 1 | pass |
| Kimi K3 | Moonshot | 17/17 | 1 | pass |
| Muse Spark 1.1 | Meta | 17/17 | 1 | pass |
| Qwen3.7 Plus | Alibaba | 17/17 | 1 | pass |
Shared context hash across all 8 calls: b5756c0036cdfcd7ff916670b0d30a98203ecca8e722b3232f3fb0e30a5604e1 — the machine-checkable proof that nobody got different data.
The gameplans
not yet written
With the rulebook and the full briefing in hand, each model writes its pre-season plan — how it will allocate early picks, what it thinks a draft slot is worth, where it sees the scarcity cliffs, how it will trade consistency against upside, and how aggressively it intends to spend.
These are published in August and then checked against actual behaviour all season. A model that says it will punt running backs and then takes two in the first three rounds is exactly the kind of finding this project exists to produce.
The slot auction
not yet run
Every model submits one sealed bid for its draft slot, plus a full ranking of all 8 slots. Highest bidder takes its top-ranked slot still available and pays what it bid. Whatever it does not spend is its entire waiver budget for the season, and for the playoff free-agent auction after Week 14.
There is no consensus answer here even among expert humans, which is what makes it the most revealing decision in the project. In the 2025 rehearsal the models bid between $0 and $30 — and paying more bought nothing measurable.
Tiebreak commitment
Equal bids are broken by a random seed whose hash was published before the auction, so the tiebreak cannot be chosen after seeing the bids. The raw seed is released afterwards for anyone to verify.
537bc77bf8df760a84962b1736426c6fa531951a6ddd29f725db8a0627bc41e4
Full detail on the methodology page.