🐟 DraftFish

Predict auction draft prices by simulating the entire information ecosystem.

The idea

We built a simulated world. Inside it, 50 people exist — each with a childhood, a career, opinions, biases, a sports information diet, and memories from past drafts they've won or lost. A quant whose immigrant father lost $40K on gut-feel stock picks. A retired firefighter who's watched Kentucky basketball for 30 years. A 28-year-old marketing grad who joined the pool because her coworkers were doing it and she doesn't lose at anything.

They live inside a simulated sports media ecosystem. 20 media personalities post content — film analysts breaking down footwork, stats bloggers posting ETP calculations, influencers hyping freshmen with fire emojis, rumor accounts whispering about injuries. Each drafter scrolls through it all, but they each see different things based on who they follow. The ESPN casual never sees the KenPom thread. The analytics nerd never sees the Instagram hype reel.

Then they sit down at a 16-person auction table and bid using the beliefs they formed from everything they consumed. The prices that emerge aren't assigned — they're emergent.

Built on MiroFish (multi-agent social simulation) + Stanford's Generative Agents. We used it to draft in a real 15-team auction on March 18, 2026. Built the highest-projected-ETP roster in the pool.

 REAL DATA (stats, odds, bracket, injuries)
     |
     +------------------+--------------------+
     |                  |                    |
 ETP CALCULATOR    PERSONA GEN          ECOSYSTEM GEN
 (deterministic)   (Claude Sonnet)      (Claude Sonnet)
     |                  |                    |
 PPG x P(advance)  50 personas:         20 agents:
 per round         4-layer backstory    analysts (stats/eye)
 = expected pts    Latin hypercube      influencers, reporters
                   6 behavioral axes   rumor, mock draft
                        |                    |
                        +--------+-----------+
                                 |
                    SOCIAL SIMULATION (OASIS)
                                 |
                   ecosystem agents post content
                   (96 posts across 6 rounds)
                                 |
                   drafters consume SELECTIVELY
                   (ESPN fan != KenPom reader)
                                 |
                   memory streams accumulate
                   (Stanford: recency x importance x relevance)
                                 |
                         REFLECTION (Claude)
                                 |
                   memories -> 3-5 beliefs -> JSON params
                   {valuations, risk, biases, budget}
                                 |
                      AUCTION ENGINE (Python)
                                 |
                   16 agents x ascending price x noise
                   1,000+ runs -> price distributions
                                 |
                   +-------------+
                   |
              VALUE = ETP - Market$
              rankings + strategy + draft assistant

What the simulation predicted

The simulation identified three market dynamics, calibrated against the room's actual bidding aggression:

The sim identified two price tiers

Agents who consumed mainstream content (ESPN, Instagram, mock drafts) drove known players to 70-80% of ETP in auction price. Agents who consumed analytics content correctly valued depth — but they were outnumbered. The result: a two-tier market.

7 out of 8 price predictions landed within range. The simulation correctly identified which players would be expensive, which would be cheap, and the magnitude of the gap between them.

The Dybantsa trap — called before the draft

AJ Dybantsa: 25.3 PPG, highlight reels everywhere, "generational freshman" narrative. But 6-seed × 2.1 expected games = 47.9 ETP. Pryce Sandfort? 49.7 ETP. More expected points than Dybantsa. Cost: $4 vs $31. The sim's influencer agents hyped Dybantsa, casual agents consumed it, and the price inflated to 7x what a comparable player costs. The team that bought him ranked 9th in projected ETP.

Stars-and-scrubs: the math doesn't work

The sim ran this scenario hundreds of times. Spending $80+ on one player leaves $20 for 5 picks. Six players at 45-55 ETP total ~290. One star at 100 ETP plus five scraps total ~260. The balanced roster wins by 30 points every time. The Boozer owner proved it — $82 on the best player in the draft, ranked 10th in projected ETP.

How we used it in a real draft

On March 18, 2026, we used DraftFish to guide a 15-team auction draft ($100 budget, 6 players). The system ran in real-time alongside the draft — calculating ETP for every nominated player, recommending max bids calibrated to the room's aggression, flagging traps, and tracking budget.

The strategy going in

DraftFish's guidance was simple: let the room overpay for stars, target 1-2 seed depth players nobody's fighting over, diversify across teams, and spend the full budget. We watched Cameron Boozer go for $82 and let him walk. We watched Dybantsa go for $31 and let him walk. We targeted the players the sim said would be cheap — and they were.

Our roster: highest projected ETP out of 15 teams

Why it worked: the room confirmed every prediction

The Boozer team ranked 10th in projected ETP. They spent $82 on Boozer (103.6 ETP) — the best individual player by far. But it left $9 for 5 picks. Their remaining roster: 37.7 + 32.0 + 31.4 + 27.6 + 25.2 = 153.9 ETP. Total: 257.5. Our six balanced picks project 30 more points.

The Dybantsa team ranked 9th. They paid $31 for a 6-seed player with 47.9 ETP. We picked up Sandfort (49.7 ETP — more than Dybantsa) for $4. Same projected points, $27 cheaper. The hype tax is real.

The Haugh team ranked 14th. They paid $58 for Haugh (67.1 ETP) — the most expensive non-Boozer player. Spent their entire budget and filled the rest with minimum-price scraps.

The top 3 teams by projected ETP all used balanced depth strategies. No team in the top 3 spent more than $44 on a single player. The bottom 3 teams all had at least one $45+ overpay. The pattern DraftFish predicted — casuals overpay for names, depth goes cheap, balanced rosters win — played out exactly.

The Illinois stack ranked last in projected ETP. One team drafted 4 Illinois players (3-seed). One early upset and half their roster goes home together. DraftFish's diversification guidance — spread across 4+ teams — was validated. Our roster spans 5 different teams.

Sleepers — invisible to casuals, max spend ~30% of ETP

Overpay traps — name recognition inflates past ETP, max spend ~65% of ETP

Full rankings — scroll to explore all 56 players

The information ecosystem

20 ecosystem agents posted content across 6 rounds — analysts, influencers, beat reporters, rumor accounts. Each with a distinct voice and bias. Drafters consumed selectively based on their information diet.

The people in the room

50 drafter personas, each with a life story that produces behavior — not trait labels. 16 drawn per simulation.

What's next

Highest impact (building now)

• Draft replay calibration. Feed tonight's real draft transcript — 90 players, actual prices, actual sequence — back through all 50 personas. One Claude call per persona: "here's what happened, how would YOU have bid?" Compare their bids to reality. Personas whose bids match = realistic. Personas that diverge = identify why. This turns one real draft into a calibration dataset that makes every future simulation more accurate.
• Real content scraping. Seed ecosystem agents with actual ESPN articles, real tweets, real injury reports, real KenPom data — not Claude-generated approximations. Tonight we missed Caleb Foster being OUT with a foot fracture because we didn't ingest injury feeds. Real content in = real beliefs out.
• LLM-driven auction. The current rules engine caps agents' bidding behavior with formulas. Real humans bid emotionally — a casual will spend $82 on Boozer because they WANT him, not because they calculated his ETP. Each bid decision should be a Claude call where the agent feels the pressure of the room and decides in character.

High impact

• Expanded player database. 108 players wasn't enough for a 90-player draft. Scrape top 3-4 scorers from every tournament team = 250+ players. No more blind spots when someone nominates a mid-major's second scorer.
• Mid-auction adaptation. Agents should react to the draft as it unfolds — "Boozer just went for $82, prices are insane, I need to adjust" or "three stars went cheap, this room is passive." Currently beliefs are locked before the first nomination.
• More irrational casual personas. Claude-generated agents are too rational. Real first-timers don't calculate ETP — they bid on names they've heard. The persona generator needs to produce agents who genuinely plan to spend $80 on "their guy."

Future extensions

• Fantasy football (August). Same persona + auction architecture, different ETP formula (projected points × games × health factor), position scarcity, snake and auction formats.
• Full OASIS integration. Replace the simplified social sim with MiroFish's full platform simulation — follower graphs, reposting mechanics, algorithmic content ranking.
• Live draft assistant. Real-time UI that tracks the auction as it happens, updates remaining budget, recalculates targets, and suggests bids — all powered by the pre-computed simulation data.