Our Fantasy Baseball League
Blake's Body Shop

Seventy-six hours of metadata

Nine drafts. Eight and a half hours each. Somewhere north of seventy-six cumulative hours of watching eleven men make decisions, drink beer and have the time of our lives. Or academically, nine years of field work in behavioral economics that isn’t assigned, nor graded. You can't place all those observations into AI to predict outcomes.

An auction room is a behavioral economics experiment running live. The first aggressive bid anchors an afternoon of prices. Loss aversion and sunk cost do their best work in the keeper rules: pay heavy for a player one March and his price follows him into the next, so that the carry-over decision quietly stops being "is he worth it" and becomes "I can't let that money have been for nothing." Rosters in this league hold monuments to auctions past.

Here is the point that matters for the framework: none of this was ever written down. How do you give an AI the notes that you don’t write over nine years? And what’s more, how do you describe the relationships that exist outside of the body shop?

There are no notes, no transcript, no spreadsheet column for "gets loose after the third hour." Seventy-six hours of behavioral data exists in my own "neural network." They cannot be pasted, parsed, or uploaded. To borrow from AI parlance, Description, the second D, has a hard boundary, and this is it. What can be described can be delegated. What can't be described can't be — and that, not any philosophy about human dignity, is why the judgment calls stay in the room. Those 76 hours of observations in the body shop, and the relationships that have existed since 1989 are the frameworks through which judgment calls get made.

Four competencies, one season

Baseball and behavioral economics have been teammates for twenty years. Moneyball was never really about baseball. It was about a market mispriced by human bias, and Michael Lewis spent his next book, The Undoing Project, tracing that idea back to Kahneman and Tversky. Every fantasy league is a miniature of the same experiment: a closed market where twelve people price the same assets, and the biases do the mispricing.

While I don’t use AI in our draft room, I’ve used Claude this year to see where I might be able to make a trade and catch up on a marginal lead. As I was taking a Claude class, built on the AI Fluency framework developed by Professors Rick Dakan and Joseph Feller with Anthropic, I thought it would be cool to show a fun AI use case, and where AI is going to fall short.

Stop describing your team. Hand over the box score.

Description is usually taught as prompting. I described my league in sentences and got back plausible, reputation-based analysis — the machine remembered who players were supposed to be, what the Internet was writing about past performance. Behavioral economists have a name for that: availability bias, judging by what comes easily to mind. It turns out a machine trained on the internet’s opinions has the internet’s biases. It "knew" Kyle Tucker was a star the way a casual fan does, regardless of what Kyle Tucker was hitting (Let’s not talk about how bad Kyle Tucker was hitting). The cure for its bias was the same as the cure for mine: force the actual numbers into the conversation.

Description also means constraints. The analysis only got sharp once the machine knew the rules that shape everything: weekly roster locks (bench players produce nothing, so bench pieces are free trade currency), one move per month (no streaming, every pickup is a commitment), and a three-year keeper clock (some stars are rentals, some are franchises — and the difference is invisible in the stats).

Well-described data buys computed strategy. This is the weekly report that good description produces:

One home run passes Team Minnow. Nine stolen bases buys almost nothing. No amount of clever prompting produces that sentence — only real data, handled with real rigor.

Showing that it works

Diligence is the simplest D: show that it’s working. The not so humble brag shows me in first place in the standings. The chart below is the season plotted at the three points I checked in with Claude: 8th in May, 3rd in early July, tied for 1st three days later. Three check-ins make a line to show the progress.