Seven days, 836 races. The model's first-pick win rate across the week was 23.6% — solid and consistent, sitting roughly 7 percentage points above what a random selection from a six-runner field would produce. That gap is the baseline test: the model should outperform chance, and over a large sample it does.
The top-3 coverage is more striking. In 507 of 836 races, at least one of the three model selections won — 60.6%. In practical terms, the winner came from the model's top three in six out of every ten races across the week. Place coverage is higher still at 86.8%, meaning in nearly nine out of ten races one of the three highlighted dogs finished in the first two.
The best individual day was June 21, when 39 first-pick wins came from 120 races at a 32.5% strike rate. The toughest spell was June 19 and 20, when the two largest cards of the week — 219 and 238 races respectively — returned 19.2% and 26.5%. Large-card days across many tracks and grade types tend to introduce more variation, and June 19 in particular shows up as a below-average day in the 7-day picture.
What is notable across the full week is the absence of a dramatic accuracy cliff between smaller and larger cards. The model held up across the range. Whether the sample runs at 12 races or 238, the first-pick win rate stayed within a band that reflects consistent underlying signal rather than luck in either direction. That stability is the most encouraging number in the dataset.