AlphaDraft
Draft night is an AI problem
How the engine thinks today — and the plan for what it becomes. The two are separated by a hard line below, because only one of them exists.
What ships today
AlphaDraft today is a deterministic engine — a calculator, and I mean that as a compliment. Built solo, used in a real draft, live in beta now. What it actually does:
- Before the draft, it prices every player. Not by projected points — by points above what the waiver wire would hand you for free. That one idea is why a running back projected for fewer points than a quarterback can be worth several times as much at auction.
- During the auction, it watches every sale, measures how hot the room is paying position by position, and adjusts what it expects each remaining player to cost.
- It tracks the whole room — every team's remaining budget, roster, and needs — and when the live feed isn't faithful enough to trust, it suppresses opponent-specific claims rather than guessing.
- When a player is on the block it advises a max bid or a pass — weighing what your roster needs, what you can afford, and what comparable players are still out there to pivot to.
- Every recommendation arrives with its reason in plain English, so if you know something the engine doesn't, you can overrule it in one glance.
- And when the data goes thin — a broken feed, a signal it can't measure — it says so. A labeled estimate, never a confident fiction.
Same inputs, same advice, every time. No learning loop, no black box. (The valuation math itself is getting a ground-up rebuild over the 2026 season — that story will get its own write-up.)
Everything below this line is a plan, not the product.
None of it ships in AlphaDraft today. It's written down here so you can see where the work is pointed — and hold us to it.
Why draft night is an AI problem
An auction draft is a real game, in the game-theory sense. Sequential decisions under uncertainty. Hidden information — nobody announces their max bid. Nine opponents with their own budgets, needs, and tempers. A hard budget, so every dollar spent on one player is a dollar that can't chase another. And a clock: an auctioneer counting down is a hostile environment for careful thinking, which is exactly when you want a machine that doesn't panic.
What makes it tractable rather than just hard: a draft is small. Ten teams, sixteen rounds, a few hundred players. A complete draft can be simulated in milliseconds — millions of practice drafts on ordinary hardware. Games with that shape have a well-known playbook now, and the rest of this page is that playbook, pointed at draft night.
Seeing futures instead of one forecast
Today's engine plans against a single expected future — this player will probably cost about that much. The first planned step is to replace the one forecast with thousands of simulated futures: price swings, tier runs, scarcity panics. Finish your roster in each one, then read the distribution. The recommendation would stop being "this price seems fair" and become something like "bidding here wins in 68 of 100 simulated futures; passing wins in 79" — illustrative numbers, from a system that isn't built yet.
And a draft only matters through the season it produces. So the plan judges a drafted roster by simulated seasons — injuries, byes, lineup calls, the waiver wire — not by how it looks on draft night. A roster that photographs well in August and a roster that survives into the playoffs are different things, and only a season simulator can tell them apart.
The AlphaGo-shaped part
Past that sits the recipe this project is named for: leagues of AI agents drafting against each other, millions of times, discovering strategy rather than being programmed with it — nomination bluffs, budget traps, tier timing. The moves no human league plays enough drafts to find.
Two honest notes on that recipe. First, self-play has a famous failure mode: agents that are brilliant against their own clones and strange against people. So the plan trains against a population of realistic drafter types — rankings-followers, aggressive spenders, bargain hunters — because a useful advisor has to be robust to how humans actually draft, not optimal against itself. Second, at decision time the pattern would be the AlphaZero one: a trained model supplies the intuition, and a bounded search over simulated futures checks it — thinking harder exactly when the decision is close.
The rule that never bends
No learned strategy gets near a real draft until it can explain itself in plain English. "Bid — 62% confidence" is not advice; it's a shrug with decimals. The discipline already exists in the product — every call ships with its reason — and it carries into the agent unchanged: if the model can't say why ("the tier is drying up; two opponents still need the position and can outbid you later"), it doesn't get to say anything.
The same goes for numbers. Anything the system leans on gets measured from real football history, never invented. If we haven't measured it, the engine doesn't pretend to know it.
The sober part
The whole plan hangs on one unglamorous truth: simulator realism and evaluation quality matter more than fancy algorithms. A powerful model in an unrealistic simulator learns the wrong thing — efficiently. So the boring parts lead: realistic simulation, strong baselines, and evaluation designed to catch us fooling ourselves. If the agent can't demonstrably out-draft simple, well-tuned heuristics across thousands of drafts, it isn't real yet — no matter how sophisticated the training run behind it.
So, plainly: today there's a calculator I trust — it's the product, and it's live. The agent is a plan, being built with the boring parts first. This page will change as the plan becomes product, and it won't claim anything before it does.