FITNESS MODEL▉

Efficiency. For every run since March 1, EF is speed in metres per minute divided by average heart rate. Conditions at each run's start come from the Open-Meteo historical archive (temperature, humidity, apparent temperature) and are cached in the database. EF is regressed on apparent temperature with an intensity term and each run is shifted to a 60°F reference. Easy and hard runs are separate series: EF rises with intensity, so the two are never compared with each other. Runs shorter than 1,500 m and treadmill runs without weather are excluded from the fit.

Fitness-fatigue. Each session's load is Edwards TRIMP: minutes in each heart-rate zone weighted 1 to 5, across swim, bike, and run. Fitness is an exponentially-weighted average of daily load with a 42-day time constant; fatigue uses 7 days; form is fitness minus fatigue. Seeded from June 1 so the August window carries real history. Negative form is expected in a build; deeply negative form with rising resting heart rate is the overreaching signal.

Race model. Garmin's predictor is compared with the two real races (5K on March 8, half on April 25) using its first snapshot after both, May 8, 2026. The bias is carried forward: corrected time = Garmin ÷ bias. Riegel-equivalent times from each race expose the gap between speed and endurance. The corrected numbers are only as current as the last calibration; a fixed-heart-rate benchmark or the Columbus Half will re-anchor them.