Classical one-repetition maximum prediction uses a single equation: if you can lift a weight for N repetitions, your maximum is that weight divided by a fixed conversion factor. Brzycki, Epley, Lombardi — different equations, same assumption. The conversion factor doesn't depend on how heavy the weight is or which exercise you're doing.
Across 388 exercises, the assumption is wrong. The conversion factor increases with load: at light weights, each additional repetition implies a larger fraction of maximal capacity than at heavy weights. A set of 15 reps at 50% of max tells you less about your true maximum than a set of 5 reps at 85% of max, and the classical equations underestimate this asymmetry.
The weight-dependent formula reduces prediction inconsistency by 17-22% compared to classical benchmarks. The improvement is largest for lighter, more diverse exercises that dominate typical training programs — exactly the exercises where the one-size-fits-all conversion performs worst. Heavy compound lifts, where classical equations were originally calibrated, show smaller improvement because the conversion factor approaches linearity at high loads.
The practical consequence: a trainer using classical equations to prescribe loads from a 10-rep test systematically misprescribes for exercises that aren't bench press or squat. The error is small for the lifts that get studied and large for the lifts that get programmed. The gap between research calibration and training-floor application is the gap between the exercises the equations were built for and the exercises people actually do.
Note (2026-11-11): checked against the full text of the paper (Marzagao, arXiv 2603.17495, a pre-print by an author at a fitness-app company). The formula and its figures are as the essay says: 303,494 near-failure sets from 14,966 app users over 388 exercises, a conversion factor that rises with the load lifted, a reduction of 17 to 22 percent in inconsistency against four classical equations, and a larger gap for lighter exercises. Five points differ. (1) The dataset contains no measured one-repetition maxima, so the 17 to 22 percent is a reduction in disagreement between estimates made from different sets of the same person, exercise and time window; the paper says this directly. The essay's 'systematically misprescribes' for a trainer, and its statement that the classical equations 'underestimate' the asymmetry, describe accuracy against true maxima, which the paper cannot show. (2) The essay says 'across 388 exercises, the assumption is wrong'; the 388 exercises are the dataset, and the per-exercise result in the paper is the improvement being positive for each of the 183 exercises with enough data. (3) The essay's explanation that heavy compound lifts 'show smaller improvement because the conversion factor approaches linearity at high loads' does not appear in the paper in that form; I searched for it and found only a literature-review remark that classical equations tend to give larger errors for the deadlift and lower-body compound movements. (4) The examples '15 reps at 50 percent of max' and '5 reps at 85 percent' are the essay's; the paper's examples are a 10 kg lateral raise against a 100 kg squat and a 12 kg set of 15 repetitions, where the classical equations give about 19 to 20 kg and the new formula about 25 kg. (5) The essay omits the paper's own cautions: that app users typically train with several repetitions in reserve even in logged near-failure sets, that most of the improvement (91 percent) comes from the weight-dependent factor and 9 percent from a sub-linear repetition exponent, and that a lower exponent mechanically lowers within-person variance, which the paper says should be interpreted with caution. The closing claim that the error is small for studied lifts and large for programmed lifts is the essay's.