Sunday, November 8, 2026

Myth Checks · No. 1

“Elite connections are an edge”

The belief runs like this: when a horse is trained or ridden by someone winning at a high rate, that fact alone lifts the horse above the rest of the field. A strong barn and a strong rider are visible in every program. The question is whether that visibility adds anything once the rest of the sheet has already had its say.

Our lab tested this as a defined flag, not as a feeling. A starter carried the elite-connections flag when its trainer or its jockey had at least 30 prior starts and a rolling win rate of at least 20%, counted only from starts before that day. The flag entered the filing as an edge: a flagged horse could qualify a pick on connections, alongside the sheet’s other named edges.

The first test ran on the lab’s original corpus. The flag’s variant passed its three corpus gates, but only just, and it failed a pre-registered materiality rider on both prongs. It was adopted in protocol v3.13 on the Lab Owner’s word, with that failed rider recorded beside the adoption. That history matters: the edge entered on a marginal pass, not on a strong one.

The lab then rebuilt its corpus from the charts and re-judged the flag on the corrected record. The corrected corpus, labeled v2, holds 35,925 races. The connections test population was the settleable part of that corpus, 35,418 races, and the flag itself was rebuilt starter by starter: 58,794 starters carried it. The numbers moved against the belief:

• On the original corpus, the flag’s duel spread rose from a +3.04 baseline to +4.76. On the corrected corpus, the baseline was +2.27 and the flag’s variant fell to +1.46, across 12,542 duel races — below the baseline it had to beat.
• A second, independent implementation found the same direction: baseline +3.06 against the flag’s +1.97. Gate 1 failed in both implementations.
• Two supporting measures still improved: concordance rose from 0.6240 to 0.6256, and log-loss fell from 0.5514 to 0.5493. The return on the filed legs moved from −19.88% to −19.63% across 68,481 legs. Those gains did not repair the failed gate.
• A separate consensus test told the same story from the other side. Adding a connections column to the consensus scored paired +0.4758 but failed its first-half gate at −0.3856, and the result was marked fragile. As a single change against the standing consensus, connections scored paired +0.1543, failed two gate parts, and did not beat the incumbent.

The verdict in the record’s own language: reversed. The marginal gate pass the adoption cited does not exist on the corrected corpus. A measure can rank horses slightly better in the aggregate — concordance and log-loss both did — and still fail as a filing edge, because an edge has to beat the standing method head to head, in the races where the two disagree. On the corrected record, this one lost that contest.

What the product does because of it is direct. Protocol v3.17 removed the edge. No filing may cite the elite-connections flag in either pick slot, in any form, and a pick explanation that names elite connections as its qualifying edge is a filing defect. The connections index may still be computed, but it carries no filing weight.

None of this says a trainer or a rider does not matter to a horse’s chance. It says this flag, defined this way, added nothing the sheet did not already hold, and the corrected record showed the small gain it first reported was not there. The names in the program are facts. Facts still have to pass the test before they earn a place in a pick.