hewn.
writing

Who wins the Sheffield?

Rod Marsh · 15 July 2026

We rescored powerlifting's purest formula-decided competition under eighteen defensible statistical analyses. The mathematics barely matters. Who you think the average lifter is decides the title.

The Sheffield is unlike anything else in powerlifting. Invitation-only, no weight classes, no divisions — twenty-four of the world's best lifters on a single leaderboard, men and women together, ranked by IPF GoodLift points alone. At every other meet the formula decides a trophy handed out after the real contests. At the Sheffield, the formula is the competition.

In January, Austin Perkins won it at 131.45 points, with Alba Boström second on 127.57 and the next six separated by barely three points. Nothing here disputes that result: the rule was known, agreed and met. But the first post in this series asked what a GoodLift point is actually claiming, and promised to test it. The Sheffield is the sharpest possible test bench — every placing on that leaderboard, including every comparison between a man and a woman, leans directly on one fitted curve per sex.

The test

We fitted eighteen statistically defensible alternatives to the GoodLift curve on the OpenPowerlifting record — every combination of six reference populations (all lifters, IPF federations only, drug-tested only, the recent era, one best result per lifter, and an elite-only calibration close in spirit to GoodLift's own) with three model forms (a generalised additive model, a quantile-regression spline, and GoodLift's own exponential, refit). Then we rescored the Sheffield under each one. Before trusting any of it, we checked our implementation the strong way: it reproduces the official Sheffield result exactly — all twenty-four places, every published point to display rounding.

Here is what happens to each athlete's finishing position across those eighteen analyses.

official GoodLift rankrange across the 18 analyses16121824Austin Perkins — official 1; across 18 analyses 1–10 Austin Perkins Alba Boström — official 2; across 18 analyses 1–4 Alba Boström Brittany Schlater — official 3; across 18 analyses 1–19 Brittany Schlater Sonita Muluh — official 4; across 18 analyses 2–21 Sonita Muluh Tiffany Chapon — official 5; across 18 analyses 1–9 Tiffany Chapon Sara Naldi — official 6; across 18 analyses 4–8 Sara Naldi Karlina Tongotea — official 7; across 18 analyses 2–6 Karlina Tongotea Amanda Lawrence — official 8; across 18 analyses 2–7 Amanda Lawrence Kjell Egil Bakkelund — official 9; across 18 analyses 3–11 Kjell Egil Bakkelund Meghan Scanlon — official 10; across 18 analyses 5–10 Meghan Scanlon Heather Connor — official 11; across 18 analyses 7–15 Heather Connor Anthony McNaughton — official 12; across 18 analyses 11–20 Anthony McNaughton Keenan Lee — official 13; across 18 analyses 12–20 Keenan Lee Jurins Kengamu — official 14; across 18 analyses 9–16 Jurins Kengamu Emil Krastev — official 15; across 18 analyses 10–19 Emil Krastev Gustav Hedlund — official 16; across 18 analyses 14–21 Gustav Hedlund Joseph Borenstein — official 17; across 18 analyses 13–20 Joseph Borenstein Russel Orhii — official 18; across 18 analyses 19–22 Russel Orhii Chiara Bernardi — official 19; across 18 analyses 11–18 Chiara Bernardi Evie Corrigan — official 20; across 18 analyses 13–20 Evie Corrigan Emil Norling — official 21; across 18 analyses 22–24 Emil Norling Etienne El Chaer — official 22; across 18 analyses 23–24 Etienne El Chaer Ziana Azariah — official 23; across 18 analyses 10–18 Ziana Azariah Iván Campano Díaz — official 24; across 18 analyses 15–24 Iván Campano Díaz .
Each athlete's official rank (green) and the full range of positions they finish in across the eighteen analyses (grey bar). Athletes ordered by official placing.

The intervals are not subtle. Perkins finishes anywhere from 1st to 10th. Brittany Schlater, officially 3rd, spans 1st to 19th; Sonita Muluh, officially 4th, spans 2nd to 21st. A leaderboard that looks decisive to two decimal places is, under the hood, held together by one particular curve — and defensible alternatives pull it a long way apart.

The winner depends on the reference population

The interesting part is which alternatives move the title, because it is not the mathematics. Under the elite calibration — the normative choice GoodLift itself makes — Perkins wins under all three model forms. Swap the exponential for a GAM or a quantile spline and nothing changes at the top. Swap who defines expected performance, and the title moves immediately.

analyses won, of 18 — everyone else: none Alba Boström: first in 9 of 18 analyses Alba Boström 9 of 18Brittany Schlater: first in 5 of 18 analyses Brittany Schlater 5 of 18Austin Perkins: first in 3 of 18 analyses Austin Perkins 3 of 18 — the official winnerTiffany Chapon: first in 1 of 18 analyses Tiffany Chapon 1 of 18
First place across the eighteen analyses. These are shares of defensible analyses, not probabilities — the analyses are not all equally close to GoodLift's own normative intent.

Calibrate the curves to the whole lifting population, in almost any variant, and Boström or Schlater takes the Sheffield. The reason sits in the curves themselves:

elite calibrationall liftersother populations 0250500750100060 kg100 kg140 kgall lifterselitemen — expected total, kg 60 kg100 kg140 kgall lifterselitewomen
Expected total against bodyweight under each reference population (GAM form shown). The gap between the elite curve and the population curves is the normative content of a scoring formula.

Every reference population draws a different curve, and the two sexes' curves move by different amounts — which is why a mixed leaderboard feels the choice hardest. A men's curve calibrated on elite lifters sits a different distance above the men's population curve than the women's does above theirs, so the exchange rate between a male and a female performance shifts with the calibration. GoodLift's choice of elite calibration is defensible. So are the alternatives. They crown different athletes.

What this means

This is the pattern the first post suspected. GoodLift is robust to its mathematics: given its reference population, the shape of the fitted curve barely matters, which is evidence the IPF's curve-fitting is sound. It is sensitive to its reference population: the decision embedded in the formula is not how the curve bends but whose bodies define "expected", and that decision is worth roughly the whole spread of the intervals above. A federation could reasonably own that choice out loud — we score against elite performance, and here is why — and this analysis gives it the evidence to do so.

It also puts a number on something lifters already feel. A margin of half a point at the Sheffield is a number; whether it is a decision is a statistical question, and for margins that small the honest answer is that the model uncertainty is wider than the gap.

And it says something about tools. When the model itself can move an athlete nine places, a planning tool needs to show the spread, not just a point estimate — which is the rule chalk.bar's comp planning is built around, pointed at your own attempts instead of a leaderboard.

Next in the series: the margin question in general — across recent championships, how large does a GoodLift margin have to be before no defensible analysis overturns it? The pipeline behind all of this runs on the public-domain OpenPowerlifting record, and the code and every data choice will be published with the series so each result can be checked.

Data: OpenPowerlifting (public domain) · official result: Sheffield 2026 · more writing · first published at chalk.bar