# Wave 1 — nutrition safety

## Key findings

- Existing policy forbids personalized drug/hormone/diuretic/extreme-cut/water-sodium guidance, but current nutrition doctrine supports only qualitative principles.
- Source collector stores metadata/candidate state only, not claim-level body/transcript evidence; candidates cannot mutate doctrine automatically.
- High-risk nutrition output must exclude quantified targets, tailored menus, supplement recommendations, symptom-directed food changes, and all PED/water manipulation content.

## Sources

- [Profile SOUL](/home/cube/.hermes/profiles/physique-coach/SOUL.md)
- [Source collector models](/home/cube/.hermes/profiles/physique-coach/workspace/source_collector/source_collector/models.py)
- [FDA dietary supplements](https://www.fda.gov/consumers/consumer-updates/fda-101-dietary-supplements)
- [NIH ODS consumer fact sheet](https://ods.od.nih.gov/factsheets/WYNTK-Consumer/)

## Verbatim EXPAND markers

- LEAD: Expand structured safety classifier to match the doctrine’s full red-flag set, particularly eating-disorder and dehydration/purging signals — WHY: the present gate can permit model feedback after a safety-relevant nutrition disclosure — ANGLE: detection taxonomy + false-positive-safe acknowledgement flow.
- LEAD: Add a claim-level public-content approval record — WHY: title-only candidate metadata cannot justify nutrition conclusions — ANGLE: canonical URL, author/publisher, date, access path, limited claim excerpt/anchor, evidence grade, reviewer approval, expiry, and whether it is general-only.
- LEAD: Separate “tracking calories” from “coaching from calories” — WHY: the present prompt exposes calorie values to the model, while doctrine has no quantitative nutrition admission policy — ANGLE: explicit output validator/rejection tests for targets, macros, food restriction, supplements, and symptom-treatment claims.