When is a margin real?

This series has established two things. A GoodLift score is one fitted curve's opinion of a performance, and there are many defensible curves. And when the stakes are highest — the Sheffield, where the formula is the whole competition — those curves crown different athletes. The natural next question leaves any single meet behind. Forget who should have won; how large does a GoodLift margin have to be before the modelling stops mattering?

Flip risk

Take two lifters in the same contest whom GoodLift separates by some margin. Call the pair flipped if at least one of our eighteen defensible analyses puts them in the other order. We measured this for every pair in the top ten of every IPF-family raw contest since 2018 with a meaningful field — 10,924 contests, around 423,000 pairs — and read it two ways. The stricter reading counts flips only among the elite-calibrated analyses, the ones that accept GoodLift's own answer to who defines expected performance and vary only the shape of the curve. The broader reading counts a flip under any of the eighteen, normative disagreements included.

There is no cliff — no magic margin where uncertainty ends. But the decay is steady, and it gives the sport its numbers. A margin under half a point is close to a coin flip: about half of such pairs reverse somewhere in the multiverse, and two in five reverse even granting GoodLift its own calibration. Flip risk falls below 5% at about 4 points on the strict reading and 6 points on the broad one, and below 1% at roughly 8 and 12 points respectively.

How often real contests live inside the noise

Those widths would be trivia if real contests were usually decided by twenty points. They are not.

The median winning margin is 3.2 points. One contest in five is decided by less than a single point, where flip risk runs near 40% even on the strict reading. Take the narrower of the two widths — the one that concedes GoodLift every normative choice it makes — and 58% of winning margins sit inside it. Take the broad width and it is 73%. Most of the Best Lifter awards handed out in the last eight years were decided by a margin the modelling could plausibly have swallowed.

What to do with a number like that

Not, to be clear, throw out the formula. Any single-formula system — DOTS, Wilks, any replacement anyone might propose — has a noise width; publishing one canonical table is a choice about governance, not evidence the uncertainty is zero. The failure mode isn't using GoodLift. It's treating a 0.4-point margin as if it were a 40 kg one.

So, rules of thumb, honestly stated. Under a point, the order of two lifters is a modelling artefact — celebrate both. Inside about four points, the result stands on GoodLift's normative choices as much as on the lifting; that is the region where a federation might reasonably say the award is decided by our published formula, which is a rule, and rules settle ties. Past about eight points, stop arguing about statistics — the stronger performance won under every defensible description of the sport we could construct.

This is also, in the end, the house point. A score is a number; a decision needs the number to clear its own uncertainty, and now the uncertainty has a size. That habit — never a number without its width — is what chalk.bar is built on, whether the number is a GoodLift point or your own projected total.

The caveats travel with the series: contests here are top-tens within sex at each meet rather than per-division awards, the corpus is not a random sample of lifters, and the code and every data choice publish with the series so each figure can be checked.