REF-1504 · Figueroa, 2026a · SportRxiv — preprint, version 2, draft dated 14 August 2026, posted 5 September · Paper
Am I a runner type or a station type?
PREPRINT — not peer reviewed: men's Open profiles form a continuum and stay stable (0.87-0.92); Roxzone is 7.9% of the race. Nothing on what training changes.
Main caveat: PREPRINT — not peer reviewed. Version 2 on SportRxiv; Crossref showed no published version on 2026-10-30.
Figueroa built a panel of men's Open HYROX results from a public results service: a census of 204,560 athletes and 333,600 results, and 62,644 races with full 30-segment splits across six cohorts, including 2,496 athletes with at least 5 races over at least 3 years.
Log segment times were centred within each race, removing the day's level, and projected onto two axes (running against stations; strength against power). Athletes formed a continuum with no clusters. Position was stable: split-career correlations of 0.706/0.660, or 0.87-0.92 corrected for measured race-to-race noise.
Deviations had a half-life of about a year; 5.2% of athletes moved permanently, and those lost about 9 percentile points. Within 1,005 athletes moving between Open and Pro within a year, the five loaded stations cost about 34%, the same across profiles within ±2% per SD, station by station rather than in line with the added load.
In doubles, partners matched on ability (r = 0.394), and complementary profiles went with outperformance (+2.5 percentile points per SD). Between-station time was 7.9% of the race and learned within one race. The preprint reports nineteen pre-registrations, labelled confidence, and retractions of its own.
Our reading: the widest description of Open splits available, carefully audited by its author, but single-authored, unreviewed, finishers only and observational. A stable profile in people who mostly train as they always have does not show that a profile cannot be changed, and the cost seen in movers has no direction: injury or ageing could cause both the shift and the decline.
Protocol
Data: public results service hyresult.com (the primary vendor site forbids automated access and was not used); identities pseudonymised. Deposit found by DOI: Zenodo 10.5281/zenodo.22261061.
Census: men's Open 204,560 athletes and 333,600 results; men's Pro 36,405 and 65,720; men's doubles 268,726 athletes and 187,574 teams; 2018-2026.
Full-split cohorts (division, sex, unit): n = 62644 races in total, all men. Cross-sectional: 29 full-field events 2024-25, 31,461 athletes, 33,057 races. Longitudinal: Open men, at least 5 races over at least 3 years, 2,496 athletes, 18,888 races. Long-span sub-cohort: 352 athletes. Elite validation: 2,214 athletes 2022-23. Pro transition: 1,037 athletes, 3,501 races. Doubles: 4,562 athletes, 6,496 races.
Repeated athletes: used on purpose. Athletes are followed across races; the cross-section has about one race per athlete; career halves are split for stability. Analyses stay within athlete where stated.
Validation: 16 core segments present and positive; 30-segment sum within ±3 s of the official total.
Profile: within-race centred log-ratios over 16 segments; three families (aerobic: 8 runs, SkiErg, row; strength: sleds, farmers; power: wall balls, lunges, burpees); fixed principal axes (65% and 35%).
Analyses: split-career correlations with attenuation ceilings, variograms, mover detection, within-athlete difference-in-differences for Open to Pro (5 loaded vs 11 identical segments, pairs within 1 year), dyadic correlations, event-demeaned transition costs. Equivalence (TOST) for nulls; labels instead of a global multiplicity correction across about 30 analyses.
Funding and interests: no external funding; no conflicts declared. Sole author, Atlantic Technical University, Cape Verde. The author declares that data pipelines, analyses, figures and drafting were produced with the assistance of an AI system (Claude, Anthropic) operating under his direction; design decisions, data collection, verification and final content are stated as his responsibility.
Key Results
| Finding | Number | Author's label or note |
|---|---|---|
| Profile structure | Continuum; no clusters by dip, silhouette or gap; axes stable across seasons (Tucker ≥0.999) | Finishers only |
| Stability across career | Raw 0.706/0.660 (long-span), 0.658/0.599 (full); corrected 0.87-0.92; disjoint sub-cohort n = 2,065 | Corrected for measured noise |
| Deviations | Half-life about 1 year; no restoring force to the population centre (+0.016, −0.032 to +0.064 per year) | |
| Movers | 5.2% (113/2,187); stay moved (median 1.08, 1 = new point); about −9 percentile points vs non-movers | Direction of cause unknown |
| Open to Pro, men | Five loaded stations +0.290 log (about 34%); 1,005 athletes, 3,345 pairs; 1 SD of Open strength shifts it −0.2% (−2.1 to +1.3); TOST ±2%/SD p = 0.036 | Athletes who chose to move; author: selection attenuates, lower bound |
| Station elasticity | 1.33 sled push to 0.39 sandbag lunges; uncorrelated with added load (r = −0.09) | Only these two stations are printed |
| Profile identity Open to Pro | 0.550/0.510; corrected 0.73/0.74 | |
| Doubles, men | Partners on ability r = 0.394 (0.382-0.405; 22,119 dyads), not shape; complementary +2.5 percentile points per SD (0.9-4.1, p = 0.002); within-athlete switch +1.71 (−0.01 to +3.44; 479 athletes) | Supported, not conclusive |
| Recurring pairs | +0.95 points per shared race (0.26-1.64) | "appear to" |
| Pacing in doubles | Run-to-run CV 15% lower than the same athletes in singles | |
| Age | Power family +0.034 to +0.041 log per decade, cross-sectional | Within-athlete fit discarded: its negative control failed |
Roxzone (between-station time), men's Open:
7.9% of total time, after event fixed effects; course geometry was 66% of the raw variance.
7.2% of race time at the top decile vs 9.1% at the bottom; 10.7% of the top-to-bottom gap (median across 231 events).
Stable within athlete (0.88 corrected); debut +2.1%, absorbed by the second race. The author withdrew calling it a separate capacity (R² = 0.525 with level and plane).
Through a race the top decile sped up across transitions and the bottom decile slowed (interaction p about 0). Trouble at a station spilled into the next transition (r about 0.20). Longer pauses before a station went with marginally better station times (−0.04; 7/7 transitions). Open to Pro added +7.3% to transitions.
Run after a station: the preprint does not analyse the next run after a station. It places station-to-transition spillover only.
Mechanisms
Measured: split times and their within-race composition; ranks. No training, physiology or load history.
Author's interpretation: the profile is a constitutive trait, and advice should start from it: exploit one's shape, budget the sleds, consider a complementary partner, and do not treat remodelling as a plan. He also says the intervention question cannot be tested here.
Inference: finishers racing at least 5 times over at least 3 years are survivors, and most of them were not trying to change their profile, so a stable profile is what one expects even if deliberate training could shift it. A shift that goes with decline could follow injury, ageing or lost time, not cause them.
Inference: the Open-to-Pro penalty is measured on those who chose to move, and only two elasticities are printed, so the paper gives no full per-station conversion.
Practical Application
Am I a runner type or a station type? PREPRINT — not peer reviewed: men's Open profiles form a continuum and stay stable (0.87-0.92); Roxzone is 7.9% of the race. Nothing on what training changes.
Going Pro: on average the five loaded stations took about 34% longer against the unchanged segments, whatever the profile (men; women about 40%, REF-1505); per-station costs differ, so this is not one athlete's forecast. Strength-leaning men did not suffer less, within ±2% per SD; this is one preprint on self-selected movers.
Roxzone: a first race costs about 2% more there, gone by the second; beyond that about half its variance is shared with level and profile (corrected R² = 0.525), so the author withdrew calling it a separate capacity; nothing here tests training it. Brisk transitions as a cheap gain stay REF-902's inference.
Doubles: the base's only analysis of pairs. Partners in the data match on ability; the complementary premium is suggestive in men and absent in women (REF-1505). Neither is a tested way to choose a partner.
Limitations
PREPRINT — not peer reviewed. Version 2 on SportRxiv; Crossref showed no published version on 2026-10-30. kb-watch checks monthly for one (relation is-preprint-of). Level capped at C. The "panels" named in its version note appear to be the author's own process (one is called pre-submission), not journal review.
Single author, AI-assisted. One person on design, data and analysis, with disclosed AI assistance on pipelines, analyses and drafting. A Zenodo deposit of data and code exists; we did not rerun it.
Observational, finishers only. The author says so; one ecosystem, one division and sex; repeated racers are selected.
Fixed order. Stations are confounded with race position; the method removes the day's level, not position.
Third-party data. Results come from a public service, not the organiser's own site; ability from rank order, compressed at the elite end.
Claims beyond data. Words such as "constitutive" and the advice in the conclusion rest on description. Cause in the mover result is unknown. No single correction across about 30 analyses; labels carry the weight.
Arithmetic. The non-overlapping cohort rows sum to 64,153 races against the stated 62,644 (our arithmetic); this paper does not itemise overlap, which its companion does.
How this entry was checked and changed
Why this level
read from the full text (SportRxiv PDF, CC BY 4.0). A large cohort would earn B+ by design; the preprint rule sets one step below and caps at C
Full text (checked 2026-10-30)
Census men's Open 204,560 athletes and 333,600 results; 62,644 races extracted with full splits; longitudinal cohort 2,496 athletes (full text)
Disattenuated stability 0.87-0.92; split-career halves 0.706/0.660 (long-span) and 0.658/0.599 (full cohort) (full text)
Movers 5.2% (113/2,187); about 9 percentile points lost vs non-movers (full text)
Load penalty +0.290 log (about 34%); TOST within ±2%/SD p = 0.036; elasticities 1.33 sled push to 0.39 sandbag lunges; r = −0.09 with dose (full text)
Partners assort on ability r = 0.394; complementarity +2.5 pp/SD (0.9-4.1), p = 0.002; partner switch +1.71 (−0.01 to +3.44), 479 focal athletes (full text)
Between-station time 7.9% of total; 7.2% at top decile, 9.1% at bottom; 10.7% of the gap, median across 231 events; debut +2.1%; trait 0.88; R² = 0.525 (full text)
No external funding; no conflicts declared; AI assistance (Claude) declared (full text)