Glossary
Adaptive Learning
Updated February 28, 2026
Adaptive Learning updates recommendations as new behavioral and physiologic signals arrive, then recalibrates output to reduce the gap between expected and observed outcomes.
Inputs that shape each update
Adaptive models treat your system as a set of inputs and correction steps rather than a fixed rule set.
| Input channel | Signal source | Effect on recommendation |
|---|---|---|
| Body trend | 7 to 14 day weighted morning weight | trend correction |
| Nutrition quality | logging completeness and portion consistency | adaptation confidence |
| Training load | planned and completed sessions | carb and timing emphasis |
| Recovery state | sleep and symptom checks | safety ceiling and pause timing |
| Context flags | travel, missed logs, adherence context | model trust calibration |
These inputs are weighted by confidence. A high-quality stream moves outputs more than a noisy stream.
What adaptive learning adjusts
The engine adjusts three layers:
| Adjustment layer | Trigger | Typical action |
|---|---|---|
| Energy and schedule | trend error across two windows | shift calories and session fuel timing |
| Macro emphasis | change in training density | move carbohydrate and protein timing |
| Coaching cadence | unstable adherence or repeated misses | increase check-ins and simplify recommendations |
Worked trend adaptations
Fat-loss adaptation over stable adherence
A user holds 80 kg and logs 85% of intake. Baseline trend after 14 days is loss of 0.40% of body weight per week. Recovery is stable and sleep is solid. The system then sets a smaller correction because trend is in the expected zone and avoids changing calories.
When the same user then shows 0.95% weekly loss for two windows with rising fatigue, the model pauses deficit increase and restores some energy around training. That adaptation is driven by trend magnitude plus recovery risk, not weight alone.
Maintenance adaptation with hidden data shifts
A user targets maintenance and logs 95% adherence. Weight oscillates +0.15 to +0.30% after weekends but returns to baseline during weekdays, while sleep remains erratic. The model classifies this as noise plus context variability and adjusts only by tightening routine cues, not by forcing macro changes. On a second 14-day block, trend remains in tolerance and no further adaptation is applied.
Common false-positive triggers
| False-positive trigger | Why it misleads | Correction action |
|---|---|---|
| Scale noise from hydration shifts | fake short-term weight swings | widen trend window |
| Wrong workout duration metadata | inflated or muted activity cost | re-log session and lower confidence |
| Missing beverages or oils | understated intake | request re-entry and pause precision changes |
| Wearable drift from strap issues | inflated heart-rate or inactivity signal | require sensor check before next update |
| Menstrual or travel context | temporary body-water shift | apply context rule for that block |
| Manual edits that conflict with model logic | signal contradiction | pause auto-change until user confirms source of truth |
Safety pause criteria
Pause adaptation when two or more safety flags appear.
- Persistent sleep debt and repeated high soreness with no expected performance compensation.
- Unclear data quality in either intake logs or wearable streams.
- Rapid body-weight movement beyond expected variance for seven to ten days.
- Repeated medical language, severe dizziness, or sustained appetite collapse.
When a safety pause starts:
- Stop adaptive targets for 48 to 72 hours.
- Keep protein baseline stable.
- Reopen only one lever at a time after data confidence returns.
Adaptive Learning remains active as a directional system. It does not replace clinical care, and it never outranks user safety when quality signals conflict.
Learn the related mechanism behind adaptive calorie goals, which explains how this logic becomes daily intake bands.