Chatbot Feedback handles adherence, appetite, and stress questions with scenario logic and explicit confidence tags. The useful test is not whether the answer sounds polished. It is whether the recommendation matches the available data and knows when to hold back.
01Adherence check-in tree
| Scenario | Signal set | Default action |
|---|
| Adherence drift | repeated missed logs and skipped meal checks | switch to simplified response mode |
| Appetite instability | persistent hunger or late-night overreach | prioritize meal sequencing and protein anchors |
| Stress pattern | sleep drop plus reduced training quality | pause aggressive targets |
02High confidence versus revise versus defer
| State | Trigger | Output |
|---|
| High confidence | complete logging and stable context | direct actionable suggestion |
| Revise | partial data or slight context conflict | present options and trade-offs |
| Defer | low signal quality or safety boundary reached | ask for confirmation and offer hold mode |
03Reliability and bias constraints
| Failure risk | Guardrail |
|---|
| Over-correction bias | cap daily change and require confirmation |
| Recency bias from one noisy week | use trend window instead of one-day change |
| Substitution hallucination | include source basis and fallback alternatives |
| Unsafe urgency pattern | escalate to safer recommendation style |
04Query examples
| Query pattern | Example | Response class |
|---|
| Adherence | "I missed two days of logging" | revise with one-step recovery template |
| Appetite | "I am ravenous at night" | adjust meal architecture first |
| Stress | "Recovery is poor and motivation is down" | defer and suggest temporary maintenance |
Medical diagnosis is out of scope. For urgent symptoms, use a local care pathway.
For reliability, privacy, and retention behavior, route to AI Coach, adaptive learning, and privacy settings.
