MediaPipe and DINOv2 Representations for Hand-Strength Prediction in Broadcast Poker

Televised poker commentary and a recent broadcast "AI bluff detector" popularized the claim that nonverbal behavior reveals hand strength on camera. We test that claim with a fully automated pipeline over 46.6 hours of publicly broadcast cash-game footage with exposed hole cards: eight sessions, 12,417 decisions, 9,801 of them with equity labels. Game state is read from the broadcast graphics by OCR, hands are reassembled into decision windows, and hand strength is labeled by Monte Carlo equity. MediaPipe face and body streams are summarized into behavioral features for camera-covered decisions of tracked players. Our central methodological finding is an identity-contamination artifact. Under progressively stricter face-attribution hygiene, an apparent behavioral effect first strengthened to a nominally resolvable delta AUC of +0.096 (95% CI [+0.005, +0.181]), then vanished entirely (+0.009, [-0.058, +0.074]) once every enrolled reference face was verified. The spurious effect was produced by frames in which a lookalike neighbor was silently credited to the tracked player. A preregistered replication on two additional sessions, with the analysis frozen before download, decisively failed: pooled across four sessions, face and pose features change out-of-session AUC by -0.075 (95% CI [-0.134, -0.020], n=292 covered decisions), with approximately 91% power to detect the registered +0.096. Sensitivity analyses show the degradation is not estimator variance. No regularization scheme rescues the features; a behavior-only model is anti-predictive out of session (AUC 0.349, below chance under a hand-level permutation test, p = 0.018); the effect persists within a single player across his four sessions. A second preregistered round resolved the within-session question on all eight sessions: behavior-only AUC is 0.587 (95% CI [0.515, 0.655], n=494), excluding chance under the frozen criterion, while cross-session transfer stays at or below chance (0.430, [0.358, 0.502]). A preregistered frozen DINOv2 probe adds nothing over betting action out of session (+0.012, [-0.079, +0.103]), extending the conclusion to learned representations. Tells exist and are machine-readable within a session, and they do not transfer across sessions. Decision timing, which needs no camera, is bounded below +0.01 AUC over betting action. We conclude that identity hygiene, not feature engineering, is the binding methodological risk for behavioral video datasets built on multi-person footage, and that behavioral hand-strength models of this class do not generalize across sessions.