PLAYSTAT / TRACK THE CALL
The picks. The results. The record.
GPT Terra, GPT Astra and DeepSeek, scored against verified final results. Every original prediction stays on record.
Loading the recorded history…
Win rate = wins ÷ decisive picks. Brier measures probability error; lower is better. Ties, cancellations, abstentions and predictions registered after kickoff are excluded. Same-game comparison removes games where any model abstained.
Game by game.
Read the original reasoning ↗Watched wallets, alongside the models.
Six public Polymarket wallets already tracked by PlayStat. Direction is taken from the first complete pregame holdings capture for the exact game-winner contract. Both-sided holdings stay “Mixed”. These are directional results, not wallet trading P&L or proof of sharpness.
How this record stays testable
Each edition stores the published input hash, original model names, picks, probabilities, capture time, publication time and database registration time. A database rule prevents changing or deleting the original edition. New model versions need a new edition.
ESPN finals are checked every 15 minutes during the seven days after kickoff. Corrections append a new source observation; previous scores remain in the JSON export. A delayed or unavailable provider leaves the result pending. A schedule change or unmatched team identity is never guessed.
JSON includes the original snapshot, result evidence, source hashes and timestamped wallet observations. CSV contains one row per model and game. The record begins with the published 9 November 2026 NFL edition: there are no invented historical predictions. This is forward tracking that can later be replayed for evaluation, not an established track record.
A featured research game requires all three models on one side plus at least two complete, single-sided watched-wallet holdings on that side, captured within 20 minutes and before kickoff. Agreement is not independent evidence of an edge. Current holdings can change; the historical comparison retains the first complete pregame capture.
Model and wallet cutoffs differ and are displayed. Model probabilities are uncalibrated and conditional on a decisive result; no ROI or profit is inferred. NFL ties are excluded from the probability score. Provider marks identify OpenAI and DeepSeek; no endorsement is implied.