FORECAST
Garmin: 4:24/km.
Two days before the race, Garmin's forecast pointed to roughly 4:24/km. The runner was feeling positive and explicitly reconsidered whether to chase a sub-1:30 finish.
CASE STUDY · CALIBRATED DECISION SUPPORT
When the forecast says faster.
A race forecast is useful evidence, not an instruction. Before the 2026 Isar-Lauf, the interesting question was not whether an AI could predict a finish time. It was how forecast, recent training context, uncertainty and the runner's own ambition should become an executable race plan.
Official outcome: 18. Isar-Lauf · 21-km race · 27 September 2026 · 1:33:03.8 · 19th male finisher.
THE DECISION CHAIN
Forecast → context → human decision → outcome.
UNCERTAINTY
A forecast was not the whole context.
The pre-race discussion also contained uncertainty about freshness and how aggressively to translate a strong forecast into the opening kilometres. The conversational layer was used to reason about that trade-off rather than simply echo the fastest number available.
EXECUTION
Controlled first. Faster later.
On race day, the runner reported following the planned controlled approach with PacePro: no early overpacing, stable execution, then a late acceleration when the race still allowed it.
THE OUTCOME
Two records, two different roles.
Official race timing: 1:33:03.8 for the 21-km race at the 18. Isar-Lauf in Bad Tölz.
Device record: the separate Garmin/Runalyze activity recorded approximately 21.37 km in 1:33:08.
The official result is the race outcome. The GPS activity is a device record. They are useful together, but they are not interchangeable.
WHY THIS MATTERS
Decision support is not prediction worship.
The useful behavior was not producing an even faster target than the wearable. It was keeping forecast, uncertainty, recent context and the runner's own judgment visible at the same time. The final pacing decision remained human; the system's job was to make that decision better informed and easier to execute.
PROVENANCE
Keep the evidence layers separate.
The pre-race forecast came from Garmin. The pacing discussion and uncertainty were conversational decision support. PacePro helped execute the race. The runner supplied the subjective race report. Garmin/Runalyze supplied the device activity. The official result supplied the canonical finish time. No single source contains the whole story.
LIMITATIONS
What this case does not prove.
This is one observational race in one runner. There is no counterfactual showing what would have happened with a different pacing strategy. The result does not establish that AI improved performance, validate a race-prediction model or separate the contribution of training, Garmin/PacePro, conversational guidance and the runner's own judgment.