It is easy to say a coaching engine "learns you over time." It is more useful to show the actual numbers. So we pulled three synthetic users out of the validation cohort and put their day-1 coaching state next to their day-60 state — the same persona, two snapshots, what changed in between and why.
Think of each comparison below as a before-and-after filmstrip frame pair: the left column is the day-1 frame (formula guess), the right column is the day-60 frame (measured reality), and the explanation is what the engine observed across the 60 frames in between. These come from deterministic simulations, which is why they are reproducible and why we can show them at all.
Pair 1 — The Calorie Target (Persona: chaotic adherence, cutting)
| Metric | Day 1 (formula) | Day 60 (measured) |
|---|---|---|
| Estimated TDEE | 2,180 kcal | 1,940 kcal |
| Calorie target (cut) | 1,680 kcal | 1,640 kcal |
| Confidence band | wide (formula) | narrow (measured) |
On day 1 the engine had only a population formula, which estimated this user's maintenance at 2,180 kcal. Across 60 days of logged intake and trend weight — even with chaotic, gap-filled logging — the energy balance math revealed the real figure was closer to 1,940. The formula had overestimated by roughly 240 kcal, which is the entire difference between a real deficit and a phantom one. By day 60 the target is anchored to the measured number, and the confidence band around it has narrowed from a guess to a measurement.
Pair 2 — The Wearable Discount (Persona: good adherence, maintenance)
| Metric | Day 1 (default) | Day 60 (calibrated) |
|---|---|---|
| Wearable active-cal trust | 80% (conservative default) | 68% (personalized) |
| Effective daily activity cals | ~480 kcal | ~410 kcal |
| Trend-vs-prediction error | not yet measurable | within tolerance |
Every user starts with the same conservative assumption: trust only 80% of the active calories a wrist wearable reports, because wearables consistently over-report. This persona's weight trend, tracked against the energy math over 60 days, showed the device was over-reporting by more than the default assumed. The engine narrowed the discount to 68% — specific to this user's device and body — so the day-60 activity figure reflects observed reality rather than a starting guess. Nothing about this calibration could have happened on day 1; it required 60 days of the trend disagreeing with the prediction in a consistent direction.
Pair 3 — The Training Progression (Persona: beginner, perfect adherence)
| Metric | Day 1 (template) | Day 60 (adapted) |
|---|---|---|
| Goblet squat working load | 12 kg (conservative start) | 28 kg (demonstrated) |
| Progression step size | fixed default | tuned to recovery signal |
| Volume tolerance | assumed (beginner template) | measured (logged sessions) |
On day 1 the engine had no performance history, so it started this beginner conservatively and deliberately — a light, safe entry load and a default progression step. Sixty logged sessions later, the engine had real evidence of how this user recovered between sessions and how they responded to load. The working weight more than doubled because the user demonstrated the capacity, and the progression step size was tuned to the recovery signal the logs actually showed rather than to a generic beginner template. The day-60 program is not a harder version of the day-1 program — it is a different program built from this user's own data.
What the Three Pairs Have in Common
Across all three filmstrip pairs the pattern is identical: day 1 is an honest, conservative guess shared by everyone; day 60 is a measurement specific to one person. The calorie target, the wearable trust, and the training load all started as defaults and became personal — not because the user filled in a better intake form, but because the engine spent 60 days watching what actually happened and corrected itself.
That is what calibrated coaching looks like in concrete numbers. It is also why the value compounds the longer you stay: the day-60 column is not the destination, it is just the first checkpoint where the measurement clearly beats the guess. The only way to generate your own version of these pairs is to give the engine 60 days of your real data and let it do the same work.