How sturdy is a GoodLift point?

Best Lifter at a championship is routinely decided by less than a point of IPF GoodLift. Two lifters in different weight classes, both of whom did everything right on the day, are separated by a formula — and the margin between them is smaller than the width of a conversion table cell. Nobody protests, because the formula is the agreed rule of the game. But it is worth asking what the rule is actually claiming.

Make it concrete. At the 2023 World Classic Championships, the top two women on the GoodLift table — one lifting at 55.8 kg, the other at 62.9 kg — finished 0.08 points apart. On the bar that margin is worth a few hundred grams, and the smallest legal jump is 2.5 kg. Was that a real difference between two athletes, or an artefact of one particular curve? The score alone cannot tell you. This series exists to answer that question — not for one meet, but for the system.

Every bodyweight-adjusted scoring system — GoodLift, DOTS, or Wilks before them — is a statistical model. Each one fits a curve of expected performance against bodyweight, and your score is your total measured against that expectation. GoodLift's curve is an exponential, fitted to elite IPF results, with separate coefficients by sex, equipment and event. Fit the curve once, freeze the coefficients, publish the tables. From that moment a modelling exercise becomes law.

Here is the thing modern statistics has learned, sometimes painfully, over the last decade: for any real dataset there are many defensible models. Different smoothers, different error structures, different subsets of the data — each choice reasonable on its own, each leading to a slightly different curve. When dozens of analyst teams are handed the same data and the same question, they return a spread of answers, all competently produced. The response to this in fields from psychology to medicine has been the multiverse analysis: fit every defensible combination of choices, and see which conclusions hold across all of them.

That reframing matters, because it changes what a result would mean. We are not building a rival formula, and this is not an argument that GoodLift is wrong. It is a stress test — the kind any system that settles awards deserves and almost no scoring system has ever had.

The plan

The OpenPowerlifting archive holds the sport's complete competition record, in the public domain. Against it we'll fit a multiverse of statistically reasonable alternatives to the GoodLift curve, varying along two dimensions:

• Model form. Generalised additive models, penalised and monotonic splines, Bayesian versions of the same, Gaussian processes, and quantile regression — which asks whether the bodyweight relationship for elite performers differs from the one for the middle of the field.
• Reference population. All valid totals, or elite performances only. Every federation, or IPF and affiliates. Tested or open. The full era or recent years. Every performance, or one per lifter.

Then the crucial step: score real competitions — world and national championships, meets with razor-thin Best Lifter margins — under every member of the multiverse, and count what actually changes. Rank correlations get reported, but the emphasis sits on decisions: winners, podiums, top tens. A bootstrap layer on top estimates how much the curve moves simply because of which lifters happened to be in the data, separating that sampling noise from disagreement about the model itself.

What we suspect

Our prior — held loosely, as a prior should be — is that the mathematics will matter less than the reference population. Swapping an exponential for a spline probably moves scores very little. Swapping who defines expected performance — elite IPF finalists versus the whole lifting population — plausibly moves them more. If that's what the data shows, it would locate the real decision inside GoodLift, and it is a normative decision wearing a technical costume: not how should the curve bend, but whose bodies is the sport normalising against.

We could be wrong on every count, and that would be just as useful. If GoodLift's rankings hold across the full multiverse, the IPF gains something rare — empirical evidence that its scoring system is robust, not merely agreed upon. If rankings wobble only where margins are already tiny, the sport learns that a half-point Best Lifter margin is a number, while whether it's a decision is a statistical question with a quantifiable answer. Either way, powerlifting ends up knowing more about its own rules than any strength sport currently does.

The analysis will publish here as it lands, with the code and the data choices in the open so every result can be checked. And the spread we find doesn't only matter for trophies: it is the same kind of uncertainty a comp plan has to navigate when it recommends your third attempt. If honest uncertainty is the house style at chalk.bar — and it is — then the sport's scoring system is the natural first place to point it.