Validation Analyses

The claims on this site, tested against the fleet — in plain language, with the methods and the numbers attached.


Claims on this site are tested, and the tests are published. Three analyses run against the fleet data: whether any design type can buy the podium, how much the crew matters inside a design type, and whether projected boat profiles predict races as well as measured ones. Each is summarized here in plain language, and each links to its method and its numbers.

What changed

· The fairness validation moved to its October 2026 edition

What changed. The fairness validation now counts 17,757 boat-race results in 1,525 class-level fields. The July 2026 edition counted 7,652 results in 572 fields; it stays published as a historical edition.

The comparison now. Each result is compared within its own race and class, ORC and IRC together, for boats with an archetype in fields of six or more. A second, opportunity-adjusted column gives every boat an equal chance of a podium in its own field.

Reading a changed ranking. An archetype’s ratio can differ from the July figure because the sample is more than twice as large and because small and large fields are now weighed fairly. Compare ratios within one edition, not across the two.

Evidence. The October 2026 validation tables · The edition’s data file

All three carry the Validation badge — directional results while the validation program completes — and recompute as the fleet grows. Terms are defined on Data Definitions.

Each analysis carries its own edition — its data cut, its computation date and its method:

  • Fairness validation, October 2026 edition — data cut 25 October 2026, computed 26 October 2026; each result compared within its own race and class, with pooled and opportunity-adjusted baselines. The July 2026 edition stays published as a historical edition.
  • Crew residual spread, July 2026 edition — data cut 1 July 2026, computed 2 July 2026; medians and spreads by archetype for archetypes with 30 or more boats.
  • Bridge check, July 2026 edition — data cut 1 July 2026, computed 2 July 2026; ORC-certificate boats against their mapped-IRC twins.

The corpus stamp below is a different clock: the race-data intake, as of 2026-08-02.

Race-data corpus · scored ORC and IRC races · as of 2026-08-02

Two descriptions, one design identity.

Before the three tests below, one question comes first: are the archetypes (the fleet’s eight recurring design types) real groupings in design space, or an artifact of the method? FleetEdge answers it by describing every boat in two different ways, both from the certificate.

One description is a statistical signature learned from the whole fleet — where a boat sits on the design axes that separate the archetypes, discovered blind from the data. The other is the set of raw engineering measurements a naval architect reads directly from the rating certificate — stiffness, sail power, length, displacement, and upwind and downwind VMG — chosen by hand, never clustered.

The two agree. Nothing tells the statistical picture what a naval architect would call a big, powerful boat or a stiff, upwind-biased one — yet across every archetype the two descriptions point the same way, and none contradicts the other. The design identity a naval architect would draw by hand is recovered from the numbers alone. Because the two lenses are built on entirely different principles, their agreement is evidence the archetypes are genuine clusters in design space, not a trick of the math. Two FleetEdge readings built from the same certificate data are complementary evidence, not independent confirmation. Their agreement shows internal consistency; independent evidence comes from data the archetypes were not built from.

The quantified side-by-side comparison carries the Validation badge and forms the archetype validity review prepared for rating authorities.

Can any design type buy the podium?

The ORC promise is that a fairly rated boat wins on sailing, not on shopping. That promise makes a testable prediction: across enough races, each design type should reach the podium about as often as it shows up on the start line.

So we counted — not one championship, every scored race in the corpus. Across 17,757 boat-race results in 1,525 class-level fields, every archetype’s share of top-three finishes lands between 0.73 and 1.29 times its share of participation. That pooled baseline does not adjust for podium opportunity: three podium places are a one-in-two chance in a six-boat class and one in ten in a thirty-boat class, so an archetype concentrated in small classes reads high. The adjusted column gives every boat its equal chance in its own field, and on it every archetype sits between 0.84 and 1.14. No design type owns the podium, and none is locked out. The full table — including the archetypes running ahead of or behind their entry numbers — is below, because publishing only the flattering rows is not validation.

ArchetypeParticipationPodium shareRatioAdjustedPodium 95% CI
AEROSTREAM19.9%24.0%1.21×1.14×[22.8%, 25.3%]
LEANRUN19.5%16.7%0.86×0.94×[15.6%, 17.9%]
AEROCORE15.5%17.9%1.15×1.09×[16.8%, 19.0%]
AEROFORCE12.8%11.2%0.88×0.87×[10.3%, 12.2%]
HEFTLINE11.4%9.7%0.85×0.89×[8.9%, 10.6%]
LEANFORCE10.8%7.9%0.73×0.84×[7.2%, 8.8%]
KEELRUN6.9%8.9%1.29×1.08×[8.1%, 9.8%]
SPARKFLEX3.1%3.7%1.17×1.12×[3.2%, 4.3%]

Boats compared within their class · fields of 6 or more · podium = top three · full method below · as of 2026-10-25

The table counts the eight archetype words. A word that covers two measured slots counts the boats of both.

Reading the gaps. The departures from 1.0 are real, but bounded — and on their own they do not indict the rating. A design built for breeze will under-collect trophies in a light-air season, and class make-up differs by region. Separating weather effects from rating effects is the next scheduled deepening of this analysis.

One caution we state plainly: FleetEdge’s archetypes are built from ORC certificate data, so using them to evaluate ORC outcomes is not fully independent. What this analysis can show is internal consistency — that no design strategy the certificates describe is systematically able to convert its rating into podiums on the pooled baseline above. The bridge check below adds results scored under a second rating rule, IRC; it is an external check on the pace gap, not a held-out test of the archetypes.

Inside a design type, the crew is the difference.

For every scored boat-race result, FleetEdge splits it into the pace the design should deliver and the pace the crew actually extracted. The leftover — the crew residual (a model-adjusted residual that helps investigate execution: what remains after the boat’s evaluated pace is accounted for — negative is faster, in s/nm) — is what the ORC promise says should decide races.

Compare typical crew contributions across the archetypes and they differ by about 56 seconds per mile. Look inside any one archetype and the spread between its sharpest and its struggling crews is about 114 — twice as wide. Which boat you sail shapes the race; how you sail it separates the results.

Boat-event crew residuals · archetypes with 30+ observations · negative = better · full method below · as of 2026-07-01

Do projected profiles predict real races?

Every boat in a scored race gets a predicted pace from its physics profile. Compare that prediction with the pace the boat actually sailed and you get a gap, in seconds per nautical mile — the smaller the gap, the better the profile describes the boat.

For boats with a measured ORC certificate, the prediction is typically off by about 41 seconds per mile. For mapped-IRC boats — whose profile is projected from their ORC design family rather than measured — it is typically off by about 52. That looks like a difference, but there are only 94 mapped-IRC race observations (one boat’s result in one scored race, joined to the weather of its race window) so far. The mapped-IRC sample grows with every scored IRC event, and this comparison updates with it.

Formally: the medians are 41 (95% confidence range 38–44, from 1,388 boat-event observations) and 52 (range 35–68, from 94). The ranges overlap. These descriptive summaries do not, on their own, establish equivalence between the two groups.

Per-event actual-vs-predicted pace gap · full method below · as of 2026-07-01

How these numbers are made.

The fairness distribution. Cohort (the boats a figure is computed over — a real entry list, a class, a fleet — named on every figure): scored races under ORC and IRC rating rules; boats compared within their class where a class division is published; race-class cells with fewer than six observed boats excluded; podium = the top three places in a cell; one result per boat per cell (best rank). Confidence intervals on podium share are Wilson 95% intervals. The adjusted ratio divides an archetype’s podiums by its equal-chance expectation: in each field, the podium places awarded, shared in proportion to the archetype’s boats in that field. Sensitivity: raising the field floor to eight (1,048 fields, 14,687 observations) leaves every conclusion unchanged, with pooled ratios spanning 0.74–1.22 and adjusted ratios 0.77–1.17.

The crew effect. Per boat-event crew residuals under the canonical polarity (negative = better). Only archetypes with thirty or more observations are compared. "Typical" is the median; the within-archetype spread is the interquartile range.

The bridge check. Per-event gap between predicted and observed race pace. "Typically off by" is the median absolute gap; confidence ranges are bootstrap 95% intervals. Mapped-IRC boats are identified through the published identity pairing between IRC entries and their ORC design-family representatives.

Coverage. Observations count boats in the FleetEdge corpus; a race’s full entry list can be larger. Counts follow the vocabulary on Data Definitions.

What these analyses cannot say. The fairness distribution cannot yet separate conditions effects from rating effects, and certifies no individual rating. FleetEdge alters no rating and no score — ORC remains the rating and scoring authority. The crew effect describes dispersion, not causation. The bridge check’s mapped-IRC sample is small and its interval wide. Each recomputes with the publication run and carries its badge until the validation program completes.

The numbers. Every figure on this page, with cohort definitions and intervals, is published as a machine-readable artifact: the fairness distribution as validation-analyses-2026-10-26.json; the crew effect and the bridge check as validation-analyses-2026-07-01.json. A new artifact is published with each recompute.

Evidence, with error bars.

Every analysis names its cohort, its interval, and its limits — and updates with the fleet.