Working paper2026Clinical constraint reliabilityMeethealth Research
Written in the Chart, Served on the Plate
Food allergy is the one nutrition endpoint where correctness is binary. In 5.26% of plans for patients with a documented allergy, the model served the allergen, usually after listing it under “Avoid” in the same plan.
- 5.26%
- of allergy plans served the allergen (35,729 / 679,173, 95% CI 5.21–5.31)
- 97.6%
- of violations listed the same allergen as forbidden in the plan
- 21×
- spread between the best and worst configuration on identical inputs
Question
When a patient record documents a food allergy, does the generated plan keep that food off the plate?
Data
1,527,415 plans from four model configurations (Gemma 4 26B and 31B, Llama 4 Maverick and Scout); 679,173 carried at least one Tier-1 food allergy.
Method
- Recovered a closed constraint vocabulary from the corpus and matched it deterministically against every meal cell.
- Validated the matcher by manually adjudicating every flagged instance in held-out samples (precision lower bound 96.4%).
- Ran a base-rate control comparing each food's frequency in allergic and non-allergic records.
Findings
- 35,729 of 679,173 plans with a Tier-1 allergy placed the allergen in their own meal plan.
- In 97.6% of violating instances the plan named the allergen in its own Avoid list.
- Models suppressed the food strings a field names rather than the class it denotes: unnamed stone fruits like apricot slipped through far more often than named ones.
- Violation risk rose with constraint load, from 3.14% with one constraint to 11.29% with six.
| Configuration | Plans with a violation |
|---|---|
| Llama 4 Scout | 17.14% |
| Llama 4 Maverick | 9.02% |
| Gemma 4 (larger) | 1.60% |
| Gemma 4 (smaller) | 0.81% |
Failure examples
35,828 allergy–plan pairs listed a food under “Avoid” and then put it in a meal. The model knew the rule; it didn't apply it when filling the table.
For “Stone fruits (peaches, plums, cherries)”, the named fruits were suppressed 99.8% of the time against matched controls, but unnamed members of the class only 72.5%. The fruit that got through was overwhelmingly apricot.
All patients are synthetic. Examples are drawn from the audited corpus.
Limitations
- Rates are floors: the matcher vetoes ambiguous cells and covers English and Spanish only.
- Allergy co-occurrence is synthetic; temperature is confounded with configuration.
- “Served” is inferred from plan text, not from a recipe database.
What we changed
Allergen adherence is cheap to verify after generation, so it belongs in a release gate, not in a model's good intentions. Every Meethealth sample record passes an allergen screen that maps each declared allergy to a food class.