APEX Horse Racing is in its research phase. Before a single race is forecast, the published science on race outcomes is being assembled, screened and rated under PRISMA 2020 with two independent readers — the same standard behind every other APEX sport.
Picks are issued only when the engines are fit, tested and graded. Until then this page reports the research count.
PRISMA is the standard used in medicine for systematic reviews of the published evidence. Every study found is logged, screened against written criteria, read in full where the full text exists, and rated by two independent readers whose scores are compared. Nothing is cherry-picked and nothing is left out quietly: the count of what was found, what was kept and what was rated is published, as it is here.
In the 1980s Bill Benter built a computer model for Hong Kong racing that weighed dozens of signals about every horse against the public odds, and over two decades it is reported to have won close to a billion dollars. He did it with a small team and the computing power of the time.
APEX is building the same idea the way it would be built now: a team of autonomous agents reading the whole literature, every signal mapped once and tested once on point-in-time data, every result replicated by a second agent with its own code, and the engines graded every morning in public. More science, more compute, the same aim — an engine that wins more than the market says it should.
On November 5 and 6 the review ran as a team of eight headless agents on the APEX machine. Every direct-racing full text was read twice, by one Antigravity reader and one Kimi reader working blind to each other, each score tied to a quoted passage that a script checked against the paper (1,016 quotes, 90% verified), and every split of two points or more was ruled on in writing by the managing model. A further 1,070 records from a fresh SciSpace search were screened against written criteria and 286 kept for reading. Highest rated: William Benter's Hong Kong model, horse fundamentals combined with the public odds and beating the odds alone on later seasons after takeout; then the quantile-mixture running-time model from Hong Kong, HodgeRank rank aggregation in Britain, and the Southampton work on how bettors mis-read the path of the odds before the off.
The 23 newly screened papers that test a signal against the odds out of sample are read first, then the 123 with free full texts. Late money and the odds path (United States, Australia, Japan), draw bias by course, and sectional times are the signals queued for point-in-time tests. The 18 FanDuel tracks were fact-checked twice from official pages, with free results, entries, tote odds and timing found at all 18, and ranked by data and evidence: Sha Tin first, then Vincennes and the Australian city tracks. The first engine test is Hong Kong under the Benter template. Nothing is forecast until it is tested, locked and graded in public.