The Lab Ledger · Racing

What We Tested

Every idea on this page — from books, from authors, from track tradition and wives' tales — was tested on the lab's own data before it was allowed anywhere near the product. This is the full ledger: what earned a place, what failed, and what cannot be tested yet. All picks, sheets and pages on this site are AI-produced. This page is a paper record; it places no wagers and is not betting advice.

How to read it: every verdict carries its scope — the races it was measured on. A claim tested on 362 races and a claim tested on 35,925 races are not the same weight of evidence, and the page says which is which. The companion page, The Books Behind the Method, covers the shelf the ideas came from; this page covers what the tests found.

What works — adopted, and in the product

The morning line as the ranking anchor

The simplest idea in handicapping: the morning-line order is the ranking. Every challenger the lab built — pace models, speed models, blended lines, a fitted statistical model — had to beat it head-to-head on the full data. None did. A ranking search over the 35,925-race Corpus v2 (2023) and the confirmation runs on Corpus v3 (13,763 races, 29 tracks, 2023) left the morning line standing as the anchor (ML-BASE v1.0). Read the number honestly: on Corpus v3 the filed selection book returned −17.82% against the favorite's −20.59%. The anchor loses less than the favorite. It does not make money, and this page does not claim it does.

Charles Carroll — the speed figure's foundation

From Handicapping Speed (1991): adjust every running line against the track's own constants, and project a horse off its own prior lines. That method is our speed figure — each horse's figure is the median of its last three own lines, adjusted per track — printed on every sheet since October 29, 2026. An audit of the constants on November 5, 2026 found no defect. The figure informs; it never outranks the morning line.

The Sartin lineage — each horse as its own control

No single book: a school of pace handicapping built on one principle — rate a horse against its own demonstrated lines, not against a class ladder. That projection principle sits inside the same figure build as Carroll's constants and has been live in the product since October 29, 2026.

Randy Giles — the seed of Race Shape

From Extreme Pace Handicapping (2008): a race's pace pressure can be typed before it is run. Giles's types were tested over 207 races on the October 2026 production record, then closed and rebuilt as our own six types — Solo Lead, Solo Buried, Soft Front, Pocket Trip, Meltdown, Dual Early — with points-only scoring. Race Shape is printed as annotation. Pace describes a race; it never reorders the ranking.

James Quinn — maiden races are the least handicapable

From The Complete Handicapper (2012): some race types resist handicapping more than others. The split measured real on October 21, 2026: maiden races were the decisive leak in both records tested — a −68.7% return on 24 filed legs in the live record, −60.4% in the rehearsal record (292 live-era races settled leg by leg). The finding entered the protocol in v2.6 as the maiden exclusion and survives today inside the cell rules that decide which races carry a selection.

Mark Cramer — the test that left a tool behind

From Value Handicapping: Cramer's overlay claim failed its test (see below), but the clean line built to test it — our own probability line from 13 owned inputs, no vendor rating inside — was kept as OUR_LINE v1. On its out-of-sample split it matched the morning line at parity (top pick won 28.4% against the line's 29.3% over 116 races): not good enough to anchor the ranking, good enough to serve as the protocol's fallback line. Own-line ranking later lost the anchor job to the morning line outright.

The five-factor consensus — printed, and graded annotation only

The morning line's top four, scored on five factors: the morning line itself, our rating weight, our speed figure, early pace percentile, and post position. The post-position factor is the one candidate that won its seat in a head-to-head column test (paired +1.9072 points against the four-factor control, Corpus v2, November 3, 2026). On Corpus v3 (13,763 races, 29 tracks, 2023) the consensus beat the official pick head-to-head — paired +2.36 points, duel +553 — but its money estimate, +2.87, carries a 95% interval of [−4.44, +10.22]. The interval crosses zero, so the consensus is graded annotation only: it is printed on every card, and it claims no wagering edge.

GN-POST — a figure construction for retrospective work

Adopted November 5, 2026 as the lab's retrospective figure construction only: each past line's figure is nudged toward the market's closing-odds percentile for that race before the horse's median is taken. Its morning-line form failed its first gate on Corpus v3 (paired −0.5062) and was not adopted. Production filing is unchanged; this one grades old races, it does not pick new ones.

What doesn't work — tested on the full data, not supported, closed

Market fusion — blending the closing odds into our line

The published research said a blend of a fundamental model and the market beats either alone. Our version, tested November 3, 2026 on Corpus v2 (25,654 races, 2023): the fit pushed our line's weight to the floor of the grid (w = 0.05), and even there the blend scored slightly worse than the market alone on held-out races (winner log-loss 1.5697 against 1.5677). The betting arms that appeared to pass were proven to be an artifact of how the corpus was built, not a fusion effect. Our line adds nothing the close does not already know. Closed.

Tom Brohamer — track-model matching

The claim: score each horse against the pace profile of the track's past winners, and back the closest match. Tested November 4, 2026 on Corpus v2 (27,562 races): the matcher trailed the favorite by −8.12 paired points and lost the head-to-head duel by −3,126. The winner profile is real as a description — winners do cluster in the mid-ranks — but the market's choice won more than twice as often as the matcher's (32.7% against 15.0%). A profile of winners is not a way to find one. Closed.

Tom Brohamer — the one-figure wonder screen, as a rule

The measurement held: horses whose only credential is one strong pace figure won 12.25% against the 15.32% their morning lines implied, across 41,441 starters (Corpus v2). But tested as an actual selection rule — remove the wonders, take the next horse — it failed its gate: the replacements won too seldom, and the duel was lost by −216. A true warning is not automatically a working rule. Nothing was adopted. Closed.

William Scott — APV and PCR money-class ratings

From Total Victory at the Track: rate horses by earnings per start (APV) or by a class-points table (PCR) instead of by the market. Tested October 21, 2026 on the production record. APV's top pick won 24.2% against the morning line's 32.9% (232 races, 1,680 starters). PCR was worse: top pick 16.3% against 30.5%, return −48.5% against −19.4%. Both rejected as rankers. Scott's form screen — no competitive race in the last three lines — measured real (0.73× the implied win rate) and survives as a warning flag, not a rating.

Gary Hall — two claims from a 2010 pace article

Claim 1, the first-turn compromise: an early-speed horse breaking outside with speed inside it is, Hall wrote, severely compromised. Tested on 10,366 route races (Corpus v2): the flagged horses did run below expectation (−1.88 points), but at roughly a third of the claimed strength, and not the required 3 points worse than either control group. The market had already priced most of it. Claim 2, combined pace in honest-pace races: the combined-pace leader trailed the favorite in honest races too (paired −2.15 points; −4.02 across all pace shapes, 27,288 races). Both claims: not supported, closed November 4, 2026.

The favorite-value claim — a win that was withdrawn

Bet only favorites our own line rates well clear of their price. The first test, November 3, 2026, said supported: the strongest arm returned +25.0% on 401 bets (Corpus v2, 25,654 races). An audit then proved the profit came from a construction artifact — races with many unscored starters inflated our probabilities, and every threshold book had concentrated its bets in exactly those races. On the repaired line, no book stood at volume. The supported verdict was withdrawn, and the claim survives only as a live paper watch with pre-set kill criteria — its one route back. Listed here because the final state of the claim is: does not work as published.

Mark Cramer — the overlay claim

The claim: when your line says a horse is much likelier than its price, bet it. Tested out of sample on October 21, 2026 with the clean line: horses our line rated at least 10 points above the market's implied chance won 11.7% of the time — below the 12.7% the market itself implied — and returned −58.3%. The claim failed inverted. Our "overlays" were our errors. Overlay selection on our line is prohibited in the protocol. Closed.

Ray Taulbot — the 12-day recency screen

From Thoroughbred Horse Racing: Playing for Profit (1949): a horse that raced within the last 12 days is sitting on a peak. Tested October 21, 2026 (362 races, 3,197 starters): horses back inside 12 days won at 0.98× their market-implied rate. Whatever truth the angle held in 1949, the market prices it now. Three further Taulbot angles tested on November 4, 2026 came back unsupported or untestable. Closed.

Wives' tale — second or third start off a layoff

The tale: a horse needs its first race back, then fires second or third time out. Tested in the October 21, 2026 form-cycle battery (362 races, 3,197 starters) and again on 241,298 starters in the later layoff-features run: starts-since-layoff bands all measured between 0.91× and 1.01× of market-implied. There is no second-start pop and no third-start peak. The real effect runs the other way — 90 or more days off is a genuine penalty (0.62×) — and the protocol carries that as a knock, not a play. Closed.

Wives' tale — first start off the claim

The tale: a sharp barn claims a horse and wins with it first time out, especially a high-percentage trainer. Tested in the lab's October 2026 angle battery: the qualifying angles did not validate — 10.8% wins against an 11.5% random expectation on the covered set. The claim's own source literature agrees: published figures for first, second and third starts after a claim never returned better than 0.81 to 0.92 per $1, even for trainers winning above 20–25%. The claim is priced. Closed.

Wives' tale — the beaten favorite bounces back

The tale: a horse that lost as favorite last out is due to make amends today. Tested as a consensus column on November 7, 2026: the first half of the year was negative (paired −0.7983), and the small full-year edge on Corpus v3 — $77.06 across 13,049 races — was more than explained by a single track (Woodbine, +$82.30, 106.8% of the total). Corpus v2's full-year figure was negative (−0.0861). One track's noise is not an angle. Closed.

Wives' tale — horse for course

The tale: some horses just love a track, so weight their past wins at today's track. Tested as a consensus column on November 7, 2026: on Corpus v3 the duel was lost (−2) and the second half was negative; on Corpus v2 the column was negative everywhere (paired −0.4780). The apparent v3 edge decomposed to 239.7% one track (Charles Town). A horse's record at a track is class and form, and the morning line already carries it. Closed.

Boxing and keying the consensus four

The idea: if the four consensus horses contain the finish, some arrangement of them — a box, a key, a price filter — must pay. Ten structures tested on Corpus v3 on November 7, 2026. The four-horse exacta box returned −26.3% over 13,410 races; the trifecta box −34.7% over 12,717; every keyed or filtered variant lost more than the plain box. One structure had shown +26.5% on the live week's 221 races; on the corpus it was the worst of its family at −41.7%. The four do contain the exacta 50.4% of the time — and the pools price that containment at a loss no arrangement escapes. Closed.

Connections, three ways — elite barn, rolling form, meet form

The idea that the trainer and jockey columns hide an edge has been tested in every form the lab could build. The elite-connections edge was adopted on November 3, 2026 and removed the next day: the full recheck on the corrected Corpus v2 found its basis reversed (duel spread +1.46 against a +2.27 baseline). A rolling connections column earned no seat in the November 3, 2026 column tests. And the meet-form column — the connections' return per $2 at this track during this meet — tested November 7, 2026 on 13,049 races (Corpus v3): paired −0.1084, duel −14, with Corpus v2 agreeing in sign. Three constructions, one answer: the morning line already prices the connections. Closed.

Trip trouble — the troubled-trip excuse

The idea: a horse whose last chart comment shows a troubled trip — steadied, wide, blocked — is better than that race looks. This was the first genuinely new information class the lab tested in this program; the flag was built from chart comments for 275,712 starters. As a consensus column (Corpus v3, races from March 2023, 11,540 head-to-head): paired −0.1545, duel −20. Corpus v2 agreed in sign. Whatever the flag knows, the market — which reads the same charts — and the sheet already knew. Closed November 7, 2026.

Six home-built ranking methods, one program

In November 2026 the lab built six ranking methods of its own — pace efficiency, consistency, form trend, class-adjusted performance, pace-shape vulnerability among them — and ran every ranking rule derived from them on both corpora. None was supported on either. Two examples, Corpus v3: the consistency rule trailed by −3.03 paired points (duel −1,776); the class-adjusted rule by −1.35 (duel −1,237). The program is complete and closed: on the current information, no re-weighting of what the sheet already knows beats the morning line.

Unproven — untested, untestable at scale, or still open

Steven Davidowitz — the key race

The concept: a race that produces multiple next-out winners was a strong race, so bet its other graduates next time. A recount on November 7, 2026 (Corpus v3) found a point-in-time key-race value computable for 42.55% of starters, covering 7,467 races at the lab's coverage bar — 533 races short of the 8,000 needed to test. The limit is not archive depth: about half of all starters fail because fewer than three of their stablemates from that race ran again before today, and horses run about once a month. The claim is untested, not refuted — it stays parked on a measured ceiling.

Barry Meadow — blinkers on, first start after a claim

From The Skeptical Handicapper: the interaction of blinkers added with a first start for a new barn. Pre-registered November 5, 2026 with frozen gates. The extraction passed its coverage floor (82.3% of starters classifiable across the archive), but the test cell itself holds 147 starters against a frozen 150-starter scale gate. Verdict: untestable at scale under its frozen terms. It reopens only if the archive grows the cell past the gate — nobody lowers the gate after the fact.

Odds movement — does the odds path know anything?

Whether the way a horse's odds move from open to close adds information the morning line lacks has never been tested here, because the record did not exist. The lab now collects odds paths daily under a frozen specification (OP-SPEC v1.1), and no gate may be evaluated until the file holds at least 2,500 qualifying races across at least 90 days and 15 tracks. At current card rates those minimums cannot be met before February 2027. Until then: unproven, and nobody peeks early.

Public opinion — the crowd as a factor

A tally of positive public backing — forums, social posts, published picks — as a sixth consensus factor. It runs in the site's other labs. In racing it exists only as a pre-registered prospective shadow test (frozen November 6, 2026): the shadow selection is computed and recorded, production filing never changes, and a decision review triggers at 150 shadow races where the factor is active — with a 30% coverage floor below which the verdict is untestable at scale. The record is young. Unproven.

LOGIT_ALL — a locked paper candidate

A logistic model over eight pre-race features, locked on November 4, 2026 — coefficients frozen, never refit. In the ranking search it beat the favorite by +2.76 paired points on the holdout and missed the +3.00 adoption margin by 0.24. A forensic audit then showed the edge was carried by its post-position term: with the draw measured point-in-time the paired figure was +1.42, and with the draw neutralized, +0.34 (Corpus v2 holdout, 18,080 races). Its prospective judge ran from the November 4 card and froze at 216 tracked races: the model's book returned −37.4% against the favorite's −41.3% on the same races (paired +3.85, duel net +4). Better than the favorite; still losing; never adopted. It remains what it was locked as — a paper candidate record.

The ledger stays open. New information classes get tested the same way everything on this page was: pre-registered, on the full data, with the scope printed next to the verdict. The shelf behind these tests is on The Books Behind the Method.