A guide to the recommendations
Every race has a box of four horses. Use it to focus your attention, then read how confident the model is.
The morning card is the cold read. Refresh after each earlier race is final: the remaining selections can change.
An experiment in public. Selections stay on the record.
Randwick · R7
A past result to explain the card. Not a current recommendation.
How the race day works
The model rules stay fixed for the meeting. New evidence from earlier races can change the calls for races still to come.
The opening assessment uses the horses’ form and the known race conditions. It has not yet seen how this meeting is unfolding. Treat the morning card as a starting point.
Once an earlier race is final, the day layer reassesses the remaining races using that result. Emerging jockey, stable and barrier patterns can change the percentages, order or membership of the four.
Check again after each earlier result and before your race jumps. A new assessment can also leave the picks unchanged. Interim results do not count.
That race’s last issued selections stay on the record. Its eventual result can inform later races, but it cannot rewrite its own prediction.
Look for the issue time and the latest earlier result used on the card. If the preceding result is still pending, the card has not absorbed it yet.
What changed in the recent record?
Across the same 45 scored races, the base four held 29 winners. The day-adjusted four held 32.
Annie’s Rose, Wonka and Marffiano were winners added to the four by the day layer. No winner was removed from the four in this window. Winning leans stayed at 11 of 45. This is evidence of better recent box coverage, not proof that every update improves the model. The comparison uses each race’s saved base and adjusted forecasts, rather than an unchanged morning card.
See the changes and the limits of the evidence →01 / Read the card
Read each selection’s role
alongside its estimated chance.
Four horses, identified by their saddlecloth numbers. In this example: 2, 8, 6 and 9. Start with all four horses. Read their form, watch how the race might unfold and decide what deserves a closer look.
The lean is the model’s first-ranked selection. It has won 18 of the 62 scored races in this window. The other three matter.
The fourth-ranked horse occupies the upset slot. It remains part of the box, even when the model gives it a small chance of winning.
The large number is the saddlecloth.
Use No. 8 to identify the horse. “Box rank 2 of 4” is the model’s preference and is shown underneath in smaller text.
The model’s estimated win chance.
It is not a price or a promise. Calibration is still being tested.
02 / Read the confidence
Race confidence describes the model’s view of the whole field. It is separate from any one horse’s percentage.
The model sees more separation between the runners. It can still be wrong.
There is a reasonable view, with meaningful uncertainty. Read the whole box.
The model has little conviction about the order. Treat the four as a shortlist to explore.
Dated example, 3–17 October 2026: 36 races. These are small samples, not promised success rates. Bars are ordinal labels, not calibrated probabilities.
Earlier cards pair Strong, Fair and Weak with Back, Small and Pass. These are the original published tier names, not instructions to stake money. The day layer can lower the tier but cannot raise it under the current rule.
High, Medium, Low and Speculative describe individual selections. They are model labels, not independently calibrated guarantees. Read them alongside the horse’s model percentage.
03 / Put it together
Check the date, conditions and issue time. Refresh after each earlier final result; check which earlier race the card has absorbed.
A Weak race deserves a different reading from a Strong one. Don’t turn a lean into a certainty.
Open the base comparison to see what the day’s evidence changed. Once the race jumps, its last issued picks stay beside the result.
winners in the box
leans won
9 meetings, 26 September–31 October 2026. R1 excluded. A small, exploratory sample across model versions. This is coverage, not evidence of beating the market.