Over the last seven days of tracked results, the model has settled into a fairly consistent rhythm: a 25.1% first-pick win rate and a 45.6% first-pick place rate across 1,083 races, with any of the top three picks winning 64.2% of the time and placing 88.7% of the time. Those any-top-3 numbers are the more telling ones - they show the model is identifying the right shape of a race far more often than it's nailing the exact winner, which is normal and expected given how much racing luck plays a part.
Day to day, performance has bounced around a touch. The 17th and 18th of July were the strongest days by volume, with 223 and 232 races respectively and first-pick win rates of 26.9% and 26.3%, both comfortably above the weekly average. The 14th and 15th were softer, with first-pick win rates dipping to 21.4% and 20.3% on smaller race counts, which is worth noting since thinner fields and fewer races per day tend to add more noise to any single day's numbers.
Nothing in this week's data points to a systematic blind spot - the fluctuation looks more like ordinary racing variance than a model weakness, particularly given the any-top-3 rate has stayed strong throughout, hovering between 63% and 76% on every single day in the sample. That consistency in identifying the right contenders, even on the days the exact winner slipped through, is the number worth watching over the win rate alone.