Building an All-Shot Expected-Score Distribution Model from Real-Match Curling Boards: Shot-Wise Accuracy and Plausibility Analysis

Curling is a strategic sport in which shot decisions involve both expected rewards and inherent risks; expected-score distributions (ESDs)—probability distributions over possible final scores—capture this uncertainty as a risk-aware strategic indicator. Although unified frameworks predicting ESDs across all shots have been proposed, their internal behavior and learned output characteristics have not been systematically examined. We construct an all-shot ESD prediction framework grounded entirely in real-match board configurations drawn from World Curling Federation championship events and conduct a shot-wise analysis along two complementary axes: accuracy, the model’s reproduction of its training targets, and plausibility, the validity of those targets at aggregate and in-match scales. Accuracy degrades monotonically with shot number; only shot 16 admits comparison with an independent reference, while intermediate-shot labels are bootstrap rollouts of the next-shot model. An independent ground-truth probe at the first backward step (shot 15,500 boards × 95 contexts) bounds the consequence of this bootstrap structure: the resulting deviation is a small selection-induced hammer-underestimation bias common to both FiLM and concatenation chains, whose magnitude is more than an order of magnitude larger than the FiLM chain versus concatenation chain stage difference, and full multi-step verification beyond the first step is structurally out of reach. Within this scope the aggregate ESD shares the gross shape of the empirical end-score distribution and the qualitative hammer/non-hammer ordering, with a hammer-favorable offset attributed—after disentangling intent from execution on the shot-percentage 100% subset—to a label execution-noise envelope that is tighter than the realized play of top-tier matches. Within real matches the chain responds smoothly to each delivered stone and, at the directly validated terminal shot, assigns mean probability ≈0.71–0.72 to the realized end score under intended execution (against 0.28 for a marginal-frequency baseline), transferring from senior to junior populations. The framework is best read as an internally coherent chain anchored to a directly validated final-shot calibration, statistically plausible under intended execution with a clear boundary at execution failure. Two structural limitations remain: rare high-magnitude outcomes are scarce in real data and produce a heavy upper tail of accuracy errors, and a single fixed execution-noise envelope that is tighter than top-tier realized play accounts for the aggregate hammer-side offset and motivates recalibration against real-match execution statistics.