Five years of data, rebuilt from raw sensor streams and fitted to two physiological models. Both say the same thing: your ceiling is unknown, because you have not run or ridden a maximal effort all year. Every 2026 number in this report is a floor, not a measurement — and the gap between the floor and the ceiling is the difference between a 1:55 half marathon and a 2:15 one.
Everything here was rebuilt from raw per-second streams. Three prior analyses were wrong because they trusted Garmin's summary fields, which are corrupted in specific and non-obvious ways.
Wall-clock timing. Garmin's duration and sumDuration both stop during auto-pause, so a "5 km" window over an interval session silently deletes the recoveries. Using directTimestamp instead makes every effort honest.
Grade adjustment. Every running sample was corrected with the Minetti (2002) energy-cost-of-gradient polynomial, so a hilly run and a flat run are directly comparable.
Power modelling. You have never owned a power meter. Cycling power was reconstructed sample-by-sample from the Martin et al. (1998) steady-state model — gravity, rolling resistance and aerodynamic drag — at 82 kg system mass, CdA 0.35, Crr 0.005. The acceleration term was dropped: it averages to zero over five minutes and was generating 2,900 W artifacts from GPS jumps.
Model fitting. Critical speed from d = CS·t + D′ on 3–20 min efforts; critical power from W = CP·t + W′ on 5/10/20 min. Both fitted per season, then validated against known maximal races.
A speed–duration curve is the outer edge of everything you managed in a season. Read the 2026 line carefully: it is not just lower, it is flat — and a flat curve is the signature of an athlete who never went hard, not one who is slow.
Plate I
Running — grade-adjusted speed–duration curve, by season
Best pace sustained for each duration · Minetti-corrected · wall clock · log scale
Plate II
Cycling — modelled power–duration curve, by season
Martin et al. (1998) physics model · 82 kg · CdA 0.35 · Crr 0.005 · outdoor rides only
| Season | Critical speed | D′ | Critical power | W/kg | Note |
|---|---|---|---|---|---|
| 2021 | 4:37.5 /km | 165 m | — | — | |
| 2022 | 4:37.8 /km | 322 m | 146 W | 2.03 | |
| 2023 | 4:03.0 /km | 108 m | 264 W | 3.67 | ▲ peak, both sports |
| 2024 | 4:16.6 /km | 215 m | 216 W | 3.00 | |
| 2025 | 4:15.6 /km | 192 m | 202 W | 2.81 | |
| 2026 | 4:52.3 /km | 52 m | 175 W | 2.43 | ▼ floor, not a measurement |
D′ — the finite distance you can cover above critical speed — is among the most stable parameters in exercise physiology. It does not fall 75% in eight months. What actually happened is that all your fast running in 2026 is 400 m repeats, and a wall-clock window across those includes the recovery; all your riding is deliberately capped at HR ≤ 152. The model can only fit the envelope you hand it, and you have handed it a season of steady efforts. Your true current critical speed is somewhere between the fitted floor of 4:52/km and the ~4:15/km your rep paces imply. That range is the single largest uncertainty in this report.
A model that can't reproduce known results is decoration. This one was tested against five maximal performances you actually produced.
Plate III
Fraction of critical speed held, by race
Actual race pace ÷ that season's fitted critical speed
Comparing pace within heart-rate bands removes fitness-independent confounders — terrain, weather, how hard you felt like going. It isolates the engine.
Plate IV
Running pace at matched heart rate
Median lap pace within each HR band · 2024 vs 2025 vs 2026
Regressing grade-adjusted 20-minute speed against the preceding eight weeks of training gives an uncomfortable answer.
| Training input | r | Reading |
|---|---|---|
| Bike hours | +0.435 | strongest single predictor |
| Swim hours | +0.375 | likely a consistency proxy |
| Long rides ≥ 90 min | +0.365 | monotonic dose–response |
| Total hours | +0.339 | |
| Run frequency | +0.136 | weak |
| Run hours | −0.130 | no relationship |
| Long runs ≥ 60 min | −0.077 | confounded by the 2023 marathon build |
Plate V
Long rides in the prior eight weeks vs running speed
Median 20-minute grade-adjusted pace achieved in each stratum
A 2026 systematic review and meta-analysis of seven randomised trials found no significant difference between cycling-only and running-only training for VO₂max (Hedges' g = −0.32, p = 0.16) or running performance (g = 0.02, p = 0.88). Studies substituted 20–50% of weekly run volume with cycling at roughly a 2:1 time ratio without performance decline. Your bike-led shape is well supported. The critical exception: none of those trials measured the ability to hold form for 21 km off a bike. Cycling builds the engine; only running builds the durability that converts it into a race split.
Three sessions of advice have leaned on your recovery metrics. That advice was not supportable, and this is why.
Plate VI
Data completeness, 394 days
Proportion of days with a usable reading
Built from the validated constant: race run pace = 76.3% of critical speed, with bike execution worth ±4%.
| Race-day CS | Run pace | 21.1 km | Requires |
|---|---|---|---|
| 3.42 m/s — today's floor | 6:23 /km | 2:14:41 | no change at all |
| 3.60 m/s | 6:04 /km | 2:08:00 | modest aerobic return |
| 3.80 m/s | 5:45 /km | 2:01:15 | six weeks of long work |
| 3.90 m/s — 2025 level | 5:36 /km | 1:58:10 | back to last year |
| 4.00 m/s | 5:28 /km | 1:55:15 | better than 2024 |
| Leg | Range | Central | Basis |
|---|---|---|---|
| Swim 1.9 km | 31 – 35 | 33 | 1:45/100m — three independent sources agree |
| T1 | 4 | ||
| Bike ~85 km | 3:05 – 3:20 | 3:10 | 19 Jul ride matched the course profile at 9.8 m/km |
| T2 | 3 | ||
| Run 21.1 km | 1:55 – 2:08 | 2:00 | 76.3% of race-day CS |
| Total | 5:38 – 6:02 | ≈ 5:50 | every minute of upside is aerobic |
Power required on the bike, from the same physics model: 171 W (2.37 W/kg) for 26 km/h, 184 W for 27, 198 W for 28. Your 19 July ride averaged 137 W at 25.6 km/h over 89 km — which is 3:19 for race distance at that effort, before taper.
A hard continuous 20 minutes running on the flat, and a hard 20 minutes on the bike. Half an hour of work that collapses a range currently spanning 4:15 to 4:52/km — the difference between a 1:55 and a 2:15 run split.
You are at two. Your best form has always followed three to five. Hold ≤ 152 bpm and finish with 15–20 minutes hard — that combination is what the durability literature specifies.
12.4 → 13.6 → 15.0 → 16.5 → 18.1 km, each step ≤ 10% over your 30-day longest. Three clear 90 minutes. Truly easy: HR ≤ 148, which right now means about 6:00/km.
Your longest unbroken swim in 2026 is 100 metres. The race is 1.9 km. One unbroken 1,500–2,000 m per week, no wall stops.
Six rep sessions since May have produced your best top-end ever and contributed nothing to what is broken. They are now eating the easy volume you need.
Not because it is proven to matter — it isn't, in your data. Because at 8% coverage the question cannot be answered, and six weeks of nightly data would answer it before the race.