Horse-Oracle Stories

Dated research stories from the record

These are not a second picks feed. Each story starts with a race window, a named comparison, and a clear limit. The goal is to explain what held up, what did not, and what should be tested next.

2026-09-05 · Probability quality · [VERIFIED]

Better probabilities do not automatically mean more winners

Across 1,785 held-out races and four expanding walk-forward folds, the probability view improved log loss by 0.0079 against the overround-removed market reference. That is evidence about probability quality—not proof of a higher win rate or betting profit.

Read the measurement →

2026-09-03 · T-5 comparison · [VERIFIED]

The tote was the harder opponent

In an accepted-set comparison of 1,048 races, the model’s top choice won 29.3% and the T-5 tote favorite won 37.5%. The difference was −8.2 percentage points, with a 95% interval from −11.3 to −5.1. The lesson is simple: an earlier anchor result did not transfer to the late tote.

See the settled record →

2026-09-05 · BRIS research lane · [CLAIMED]

Prime Power is promising, but it is not live

A four-fold research study found a positive incremental signal in 1,048 out-of-sample races from a 1,538-race BRIS full-field set. September coverage then fell to zero, and the feature is not in the live feature set. The result stays a candidate until a fresh disjoint month replicates it.

Read the evidence boundary →

How to use these stories

1. Read the date and race count first.
2. Check the baseline and outcome definition.
3. Read the limitation before forming a conclusion.
4. Keep pre-race study separate from post-race settlement.

PAPER MODEL — NOT BETTING ADVICE · $0 STAKED