Every horse on the Daily Pace Card carries one speed figure. The figure feeds the handicap weight beside the name and the speed column in the consensus. Readers see the output every day; this post shows the build, then walks through the examination we put it through this week — grading 23 Breeders’ Cup Classic winners on one scale, pricing all 14 races at the 2026 Breeders’ Cup, and then trying to prove the figure’s own adjustment wrong.
Every past-performance line starts as a raw figure. We measure the horse’s final time, plus beaten lengths, against a constant for that track, distance, and surface — Charles Carroll’s track-constant method, the foundation the Books post describes. The constants come from our own line store. A track-distance-surface cell prices a race only when at least 30 lines support it; a thinner cell falls back to the track-and-surface rung, then the track rung, then a global value.
A raw figure treats every day as equal, and days are not equal. Tracks run fast and slow. The projection variant measures the day from the horses who ran on it. The build walks the population in date order. Each horse carries a level: the median of its adjusted figures from earlier days only. In each race, every starter with a level contributes the gap between today’s raw figure and its own level. The median gap is the race’s sample. The day’s variant for a surface is the median of that surface’s race samples, capped at six points in either direction. A surface with too few sampled races takes the whole day’s median instead. Turf lines take no variant at all.
Each horse acts as its own control. A strong field cannot inflate the variant, because every starter is measured against itself — never against the class of the race it happens to sit in.
A horse’s speed figure is the median of its last three adjusted lines. The scale is native: 0 means typical for the track, distance, and surface. Positive is faster than typical; negative is slower. That is the whole build — a constant, a projection, a median.
The variant is the step a skeptic should attack, so we attacked it first. We pre-registered a test on Corpus v3: 13,745 races, the same starters ranked two ways — by the raw figure, and by the adjusted figure. The adjusted ranking won. Its top pick won 25.13% of races against 23.08% for raw. In the 6,396 races where the two rankings disagreed at the top, the adjusted pick won 1,452 head-to-head duels and lost 1,171 — a net of +281. The adjusted ranking also placed the winner higher (mean rank 3.24 against 3.40) and found the winner in the top three more often (61.6% against 58.0%).
We report the limit with the result: the money edge is small. The paired return difference is +0.42 points, and its interval crosses zero; neither ranking profits as a standalone betting book. The verdict rests on the ranking evidence, which runs one direction. Under the frozen decision rule, the adjusted figure wins, and the variant stays in the build.
We graded the 23 Classic-winning performances listed on our comparison table — 2004 through 2026 — on one population: our line store merged with all 23 official chart cards, 121,816 lines. The variant, not the raw running, does most of the sorting.
Arrogate (2016) ran a raw of only +0.20 — an ordinary number against the constant alone. His card’s variant was −3.33: the day ran slow against the levels of the horses on it, so the adjustment lifts him to +3.53, first of the 23. Authentic (2020) ran the best raw in the set, +3.45, and his day ran fast: a +2.98 variant cuts him to +0.47, fifteenth. Ghostzapper (2004) needed no help — no variant resolved on his card, so his figure is his raw, +2.29, tied for second. Sorted by raw alone, the order runs Authentic, Sovereignty, Ghostzapper, American Pharoah, Flightline. The adjusted table reads differently, because championship days are rarely typical days.
We then priced the 2026 Breeders’ Cup at Keeneland — all 14 races, every starter — and compared each horse’s actual figure against the projection its last three lines had earned.
Sovereignty won the Classic at +0.81. His raw was +2.85, the second-best raw in the 23-Classic table; Saturday’s dirt variant, +2.04, cut it to +0.81. He also won below his own form: his projection stood at +3.21. T O Elvis ran the highest figure by any winner all weekend, +4.72 in the Sprint. Both dirt cards drew positive variants — +2.77 on Friday and +2.04 on Saturday — meaning the chain read both cards as fast against the fields’ own levels, and cut dirt figures across the board.
Sovereignty’s +0.81 raised the obvious question. A Classic winner with the second-best raw in a 23-race table landed twelfth on our scale because one number — the pooled day variant — charged him the full +2.04. We went looking for the error.
The dissection found the pressure point. Saturday’s dirt variant pooled eight race samples: five sprints and three routes. Its median landed on two sprint races, +2.02 and +2.07. The Classic’s own sample was +1.34, the lowest on the card — the Classic field, measured against its own levels, reported a slower day than the sprints did. Routes alone gave a variant of +1.70, which prices Sovereignty at +1.16. The pooling is real, and it costs him about a third of a point on that reading.
A local artifact is not a rule. We pre-registered three challengers and ran each through the full corpus under the same frozen gates as every lab test.
Band-split variants — separate sprint and route variants for each day and surface. On routes alone the split gained (paired +0.83, duel +18); on sprints it gave the gain back (paired −0.38, duel −24). Across all 13,745 races: paired +0.15 with an interval crossing zero, and the duel lost by 6. Not supported.
Race-local shrinkage — pull each race’s variant toward that race’s own sample. This arm ranked winners best of the three (mean rank 3.21; winner in the top three 62.3% of the time) and still lost: paired −2.06, duel −4. Its top picks won at the same rate and paid less. Not supported.
Variant damping — apply the variant at a fraction instead of in full, at long distances and in top-class races. In graded company — a purse proxy, stakes at $100,000 and up, because the corpus carries no grade field — damping failed at every level under the frozen rule: the damped books returned more only because their fewer winners paid more, and the duel and the win rate fell at every level. At ten furlongs and beyond, every damping level made every metric worse, though at 112 races that cell is too small to render a verdict. One cell supported it: at nine furlongs and beyond, three-quarter damping cleared all three frozen conditions (paired +9.42, duel +4, win rate 24.48% against 23.44%) — by a duel margin of four races, with its interval touching zero. We chased it onto 697 held-out races the discovery never touched, and it lost on every metric (paired −6.63, duel −8, win rate 21.66% against 22.81%). Not confirmed; the cell reads as noise, and the thread closes.
We checked the remaining suspects. Thin levels do not drive the charge: require three prior lines behind every level and Saturday’s variant becomes +1.84, with Sovereignty at +1.01. The constants audit found no defect: the builder’s keys match the lookup’s keys on all 119,074 lines that carry a distance, with zero mismatches. Keeneland holds no 10-furlong cell because our store contains zero Keeneland lines at that distance, so the Classic priced at the track-and-surface rung. On same-distance evidence that price is right to 0.05 points — Saratoga’s true 10-furlong dirt cell reads 53.2258 against the rung’s 53.2531.
The audit closes with no error found and all three adjustments refused. Even the friendliest re-pricing we built — band variants run through the entire chain — leaves Sovereignty at +1.82. The pooled variant stands, and production did not change.
The gates stay where they are. A new variant construction earns adoption only by beating the pooled variant on the full corpus under the frozen rule — paired return positive, the head-to-head duel won, top-pick win rate at least even — and then confirming on held-out races it has never seen. No construction has passed. When one does, we will publish the numbers and change the build. Until then, the figure ships as described: a constant, a projection, a median — tested hardest on the days it looked most wrong.
The method’s sources are on the record in The Books Behind the Method.
The picks, the pace sheets, and this podcast are all produced by AI. The system behind them was built and tested using many of today's leading AI systems in a strict deterministic environment set up as a handicapping lab, and it has been independently audited by other Large Language Models. Rankings follow the morning line.