Same stroke, different swimmer

Five years of open-water swimming, read back out of an Apple Health export that no application had ever been permitted to open. The turnover never changed. Almost everything else did.

244: Swims logged
137 km: Total distance
93 h: In the water
25.2: Strokes per min
178 s: Pace per 100 m
1.36 m: Per stroke

Distance per stroke, every half-year. Four flat years, then a low point — and the low point is the first half on a new watch, so the hardware cannot explain the four consecutive rises that follow it on that same watch.

What never moved

25.2 spm

Stroke rate in 2026. It was 25.0 in 2022 — and 24.7 and 25.2 in the years between. Five seasons at the same cadence, within a fifth of a stroke per minute.

That constancy is what makes the rest readable. When turnover holds still, a change in speed has to come from somewhere else.

What flipped

−2.9 %

Last third of a swim against the first third, 2025–26. The same figure for 2022–24 was +2.9%.

The pattern used to be a fade. It is now a negative split — and unlike every other number here that is a shape rather than a level, so warmer water does not explain it away.

The five years

Four flat years, then a break

Restricted to continuous swimming — pace under 230 s/100 m, credible stroke rate, plausible efficiency — the record sits still for four years and then moves all at once.

Year	Swims	Metres per stroke	Pace s/100 m	Strokes per min
2022	17	1.19	205	25.0
2023	2	1.32	222	20.5
2024	6	1.18	210	24.7
2025	8	1.14	215	25.2
2026	26	1.36	178	25.2

2023 rests on two qualifying swims and 2024 on six. Those rows are sightings, not seasons — the years either side carry the argument.

The chart above reads as a collapse across 2023–25. It is not a decline but a change of activity. The calendar fills with October, November and January, and water temperature governs distance almost completely: 307 m at under 12 °C against 903 m at 14–16 °C. Those short swims are cold immersion, logged the same way as a training swim.

Inside a single swim

The warm-up is real, and it is not about speed

Trim 100 yards off each end — entering and leaving the water is not swimming — and pace across the rest of the swim is flat to within one percent. There is no pace warm-up. There is an efficiency one.

Median stroke rate by tenth of the swim, pooled across 100 swims. Turnover opens above the swim's own median and is back to it by the third tenth. Pace over the same stretch does not move.

The first stretch costs 4.4% more strokes and 7.8% less distance per stroke, at the same speed. The opening pace is bought with turnover instead of technique.

The penalty scales with the swim: −2.4% under 600 m, −10.4% at 600–1200 m, −12.7% beyond that. Longer swims are not harder to start; they simply provide a settled baseline for the opening to look bad against.

Where it used to fall apart

A mid-swim stall became a negative split

Median pace by third of the swim. The older era slumps 11% through the middle and recovers; the recent era holds flat and finishes six percent faster than it started.

The floating shows up in the same place. In 2022–24 the median swim's worst tenth ran 92% slower than its own median, and 44 of 60 swims had a stretch more than half as slow again. By 2025–26 that worst tenth is 39%, in 14 of 40 swims.

Is it real?

A shifted process, not a good patch

A rising median can be luck; a control chart can tell the difference. Limits here are computed from 2025 alone — the first full year on the current watch — and 2026 is plotted against them, so a signal can be neither an artefact of new hardware nor a centre line dragged by the very points it is meant to judge.

Individuals chart, distance per stroke. Thirty-four of 38 swims sit above the 2025 centre line, seven break the upper limit, and nine consecutive swims from 14 May trigger a sustained-shift rule that never resets.

The shift dates to mid-May, which is also when the water crossed 14 °C and stayed there. With only three qualifying swims before it, the chart cannot separate got better from got warmer. The negative split can — a shape change is not something warm water hands you.

What was thrown away

A sixth of the record is unusable

Every figure above survives a plausibility gate, because the raw data does not deserve trust it has not earned.

Discarded	Count	Why
Impossible pace	36	One swim at 60 s/100 m — near world-record
Impossible efficiency	20	Two report 118 and 207 m per stroke
No stroke count	10	Distance recorded, cadence absent
Drifting, not swimming	4	High metres-per-stroke at a walking pace
Heart rate, entirely	244	The median swim carried one reading

Distance per stroke is unbounded for a swimmer who stops swimming. The three most efficient swims in the archive were drifts.

That is the trap the whole exercise is built to avoid, and it is why efficiency is never gated alone here — always jointly with pace and stroke rate. Heart rate was dropped outright: wrist optical sensors barely function underwater, and a mean of one reading presented without its count is a lie with a decimal point.

The open question

October settles it

When the water drops back under 14 °C, distance per stroke does one of two things. Near 1.35 means the gain is technique and it is permanent. Back to 1.17 means it was thermal all along. The prediction has a date on it, which is the only kind worth making.

Volunteers wanted

Run this on your own swimming

This came out of one personal archive, with tooling written along the way. It now needs swimmers who are not its author — different water, different watches, different habits — to find where it breaks.

What you need: An Apple Watch that records swims, and a Health export. Roughly a season is enough to say something; more is better.
What happens to it: Nothing leaves your laptop. The reduction runs locally and emits a few hundred rows of derived numbers — no raw series, no medical history, no upload.
What you get: Your own version of this study, and a straight answer about whether the method holds on data it was not built from.

The tooling is open at dhk/plumbline — an inventory pass that writes nothing, an allowlist filter that reduces a Health export to sport before it is copied anywhere, and the analysis that produced every figure above.

Volunteer to test it ↗

Method

244 swims, 2022-02-11 to 2026-08-23, San Francisco Bay and pools. Source is Apple HealthKit via a local export; Strava holds none of the stroke data, because no application had ever been granted read access to it. Figures are medians unless stated, and open water is reported separately from pool throughout — a pool swim has walls to push off and no current. Distance is GPS-derived and approximate in open water, where the antenna is submerged most of the time; comparisons hold because the bias is consistent across years, but absolute pace may not.