Humans were plateauing in the marathon. For the last 20 years, they’ve stopped — and the record just broke 2:00:00.

“The marathon world record is approaching a hard physiological wall” is a claim with real pedigree — exercise physiologists spent a decade and a half arguing about whether a sub-2:00:00 marathon was even biologically possible, and as recently as 2019 the only person to run one (Eliud Kipchoge, Vienna) did it in a shoe-and-pacing setup deliberately ineligible for record purposes. This run tests the claim against the actual 118-year record book rather than against the argument.

Source: Wikipedia’s Marathon world record progression page, itself sourced record-by-record to World Athletics/IAAF and the Association of Road Racing Statisticians (ARRS) — 53 men’s rows and 42 women’s rows as published. Before a single global ratifying body took over marathon records in 2003, IAAF and the independent ARRS tracked overlapping lists, so the raw table contains 11 men’s and 8 women’s entries that do not actually beat the mark before them (disputed-course or dual-authority marks). Nothing here edits a time or drops a row for any reason except that test: rows are kept only if they beat every prior time, recovering the true best-ever-to-date progression — 42 men’s records and 34 women’s records survive. One honesty note carries through every fit below: a series of records is not a series of independent measurements — by construction every point is a new extreme — so these p-values describe this specific 118-year history, not a formal guarantee about some population of alternate histories, in the same spirit as this desk's flags on run 027's repetition-preserving surrogates and run 028's non-independent months.

The long-run claim holds up. Split the 42 men’s records at their median year (1965) and fit each half separately: +0.064 km/h/yr before 1965 (n=21, R²=0.94) falls to +0.027 km/h/yr after (n=21, R²=0.95) — the rate of improvement roughly halved. Testing the difference directly (an interaction term in one combined regression) puts the gap at -0.0370 km/h/yr, 95% CI [-0.0456, -0.0283] — nowhere near zero, p=1.6e-10. Over the full century-plus, this is real: the marathon did slow down.

But “did” is the operative tense. Restrict the same fit to just the last 20 years — the 9 records set 2006–2026 — and the rate climbs back to +0.042 km/h/yr (95% CI [+0.036, +0.049], R²=0.97) — 1.6× the post-1965 rate and within a hair of the +0.049 km/h/yr century-long average. Sabastian Sawe’s 2026 1:59:30 (Nairobi’s runner, London Marathon, the first record-eligible sub-2:00:00) sits +1.13% above where the post-1965 “slow” trend line predicts for 2026, and +0.35% above the last-20-years line — the newest record is not lagging the recent pace, it is running right on it. The plateau this desk can confirm is a 20th-century one; the last two decades don't extend it.

A second, independent test says the same thing from a different angle: if records were coming less often, that alone would support “plateauing” regardless of how big each jump is. They aren’t. Years-between-records, regressed against time across all 41 gaps: slope +0.0055 yr/yr, 95% CI [-0.0222, +0.0332] — contains zero, p=0.69. Descriptively, the four most recent gaps (2018–2026) averaged 2.89 years against 1.75 for the four before them — noisier, not conclusive at n=4 each, and not the story the speed data tells: the last four records were each unusually large drops (Kipchoge’s two, then Kiptum’s 34 seconds, then Sawe’s 65), not small chips off a wall.

The likeliest driver of the reacceleration is not physiology, and the run says so plainly rather than crediting it to human improvement it can’t isolate: every one of the last four records (2018 onward) was run in a carbon-plated “super shoe” (Nike’s Vaporfly/Alphafly line and its rivals, commercialized from 2017), a documented multi-percent efficiency gain nobody in the earlier eras had access to. This run cannot separate how much of the last decade’s reacceleration is the shoe versus the athlete — it can only report that the deceleration this desk can confirm stopped right around when the equipment changed.