Who counts as elite?

When we rescored the Sheffield under eighteen defensible analyses, the finding was that the winner depended on the reference population — elite calibration kept the official result, population calibration overturned it — and we were explicit that eighteen analyses were a floor, not a ceiling. This update raises the floor. Three new model forms join the grid: a hierarchical model that pools each lifter's repeated performances (checked against a full Bayesian fit — details at the end), a Gaussian process, and a monotone spline. And because "elite" was doing so much work in the first result, it now gets the multiverse treatment itself: six defensible definitions, from the top 5% of the record to the athletes who actually stood on championship platforms. Six forms, eleven reference populations, 66 universes, every competition rescored under each.

Two methodology corrections ship with the wider grid, found while building it: every curve now scores only within its reference population's observed bodyweights (flexible forms extrapolated badly on small populations), and the lifter sampling behind the fits is now fully deterministic. Both touch the original eighteen cells. Their headline results are unchanged — the winner tallies match the published post exactly (Boström 9 of eighteen, Perkins 3, as before) — but individual scores shifted within noise, which is why the earlier post's figures stay pinned to its own run.

What got stronger

The first post's happier finding — that GoodLift is robust to its mathematics — survives the harder test. Under both percentile definitions of elite, the official winner takes 12 of 12 cells: every form, including the three new ones, agrees. If the argument is about curve-fitting technique, there is no argument. The IPF's mathematics is fine.

What got sharper

The argument is not about curve-fitting technique.

Define elite by percentile and Perkins wins every cell. Define it by who lifts at championships and the title changes hands. The IPF's published evaluation of GoodLift drew its data from world and European championships from 2011; build exactly that reference population — 20,186 championship performances — and Boström takes the title under 4 of 6 forms. World-championships-only calibration agrees (4 of 6), and restricting to open worlds — in case Masters and Sub-Junior fields were doing the work — leaves the flip standing (4 of 6). Across all 66 universes, the official winner is no longer the most likely one: Boström wins 44% of analyses, Perkins 32%.

One affinity worth naming: the IPF's evaluation grouped athletes into layers by distance from the class record; the record-relative population here (within 15% of the bin's best) is the same logic. The definitions the sport has actually reached for sit on both sides of this split.

The margins behind these flips are one to four points — inside the flip-risk noise we published before running any of this. And the disagreement is not confined to the top: championship calibrations reorder the whole board, averaging a Kendall's τ of 0.49 against the official ranking and moving athletes 4.1 places on average, where the universe that best reproduces the official ranking (top-1%) sits at 0.70 and 2.6 places. Rank intervals widen to match: Schlater spans 2–24 across the multiverse, Muluh 1–24.

Why championship calibration favours the women

The mechanism is an exchange rate. A mixed leaderboard compares each athlete to their own sex's expected curve, so what matters is not where a calibration sets the bar but whether it moves the two bars differently. It does.

Championship fields sit below the percentile elite for both sexes — but not equally. The men's world-championship curve sits at 88.0% of their top-5% curve; the women's at 84.8%. That 3.3-point differential is scoring ground the women gain under championship calibration, and it is the whole story of the flip: under the GAM, Boström scores 154.6 to Perkins' 153.7 against world-championship fields, and the two swap places again (152.4 to 150.7) when the calibration is open-worlds only — a reminder that at these margins the title is living inside model uncertainty, not outside it.

Age divisions widen the differential but don't create it: including Masters and Sub-Junior worlds costs the women's curve 2.6 points against 0.9 for the men. The tempting explanation — that women's age-division fields are relatively weaker — is wrong: their age-division lifters sit closer to their open medians than the men's do (96.7% vs 95.8%), and both sexes' championship fields are about a third age-division entries (36.8% and 35.7%). The differential lives in where that mass lands across the bodyweight range, not in how strong it is — the kind of structural detail no single fitted curve, ours included, should be assumed to have settled.

The thresholds, revisited

The margin post's rules of thumb came with a warning that its multiverse was a floor. Here is the floor rising. Under sixty-six universes, the "no defensible analysis disagrees" margin no longer exists inside the fifteen points we measure — mechanically, more analyses mean more chances for one to flip any pair, so that reading degrades by construction as a multiverse grows. The statistic that survives is the sturdier one: the margin at which fewer than 5% of pairs are flipped by at least a tenth of the universes now sits at 7.5 points. The variance decomposition has evened out too: with six elite definitions in play, model form (25%), reference population (16%) and their interaction (14%) each carry a real share of the assignable ranking disagreement — and roughly half resists assignment to any of them, which is its own honest finding: no single lever explains the multiverse.

The series opener promised a bootstrap layer, and here it is: resample each reference population with replacement and refit (100 resamples), and the curves move by just 0.2% on the whole record and 0.5% on the championship pool — an order of magnitude smaller than the differences between calibrations. Sampling luck is not the story; the choice of reference is. The championship flip itself is probabilistic rather than fated: under resampling of the championship pool, Alba Boström takes the title in 59% of resamples.

What this settles, and what it can't

These 66 analyses can't tell the sport which definition of elite is right, because that was never a statistical question. What they settle is that the question matters: the percentile definitions and the championship definitions are all defensible, they produce different Sheffield champions, and no amount of better curve-fitting will arbitrate between them. A federation could own this choice explicitly — we score against the top percentile of the whole record, or we score against championship fields — and either sentence would be honest. What isn't tenable, after this, is treating the choice as a technicality.

The receipts, as ever: the hierarchical form's fast approximation was validated against a full MCMC fit (maximum divergence 4.9% within the supported bodyweight range); the monotone spline and curve fits carry logged sample caps; every scoring curve is evaluated only within its reference population's observed bodyweights; and sixty-six universes are still a floor. The code and every data choice ship with the series, on the same public-domain record anyone can check — and the habit this series keeps demonstrating is the one chalk.bar sells: never hand an athlete a number without showing how far it can move under assumptions as good as yours.