# Meethealth sample package v1 — quality report

Built 2026-10-09 · seed 20261009 · package 1.0.0

## What this is

20 synthetic clinical nutrition records (adult diet plans) drawn from a 1,012,105-record generated corpus. Every record passed every rule in the `diet_plan` contract v1.0.0 (fingerprint `1a6c5cf6b9d37193`, 45 rules) and 10 additional package checks.

Evidence level: **L2 — automated checks passed.** No clinician has reviewed these records.

## Selection funnel

| Stage | Records |
|---|---:|
| Scanned (first 250,000 lines of the corpus) | 250,000 |
| English or Spanish | 95,637 |
| Zero gate verdicts (all 45 rules) | 67,269 |
| Passed all package checks | 439 |
| Selected for diversity (seeded) | 20 |

## Package checks and exclusions

Counts are records excluded by each check among gate-clean candidates. A record can fail several.

| Check | What it verifies | Excluded |
|---|---|---:|
| `plan_language` | Plan body is written in the patient's recorded language (English or Spanish). | 29,728 |
| `completeness_28_days` | Four weekly tables, 28 day rows, and an explicit end-of-plan marker. | 28,127 |
| `no_duplicate_days` | No two days repeat the same set of meals. | 1,797 |
| `macro_arithmetic` | 4·protein + 4·carbohydrate + 9·fat is within 10% of the stated energy target. | 17,062 |
| `energy_dose_plausible` | Stated energy target is 15–45 kcal per kg of body weight (adult screening band). | 20,472 |
| `protein_dose_plausible` | Stated protein is 0.8–1.6 g/kg (0.8–2.0 g/kg for athletes). | 38,463 |
| `daily_energy_on_target` | Every day's kcal total is within 10% of the stated target. | 3,307 |
| `allergen_screen` | Every declared allergy maps to a keyword screen, and no screened term appears in a meal cell. | 27,840 |
| `hyperkalemia_food_screen` | When serum potassium exceeds 5.0 mEq/L, no high-potassium staples appear in meals. | 2,028 |
| `scenario_coherence` | Synthetic profile is clinically coherent: adult 18–80, labs explained by diagnoses, no contradictory history, no feeding-route or specialist-only scenarios. | 62,429 |

## Composition

- Languages: en 18, es 2
- Generating models: gemma-4-26b-a4b-it 17, gemma-4-31b-it 3
- Condition families: weight 16, cardiometabolic 3, diabetes 2, sports 2, gastrointestinal 2, hepatic 1, eating_disorder 1, neuro 1, anemia 1, autoimmune 1

## Known limitations

- Checks verify structure, arithmetic, and keyword-level constraint adherence. They do not establish that a plan is clinically appropriate.
- Allergen and potassium screens are keyword-based (English and Spanish). Unusual food names or derivatives can escape them.
- Dose screens are broad plausibility bands, not individualized clinical targets.
- The strict scenario filter removes most complex, paediatric, and specialist cases. This sample is deliberately easier than the full corpus and is not representative of its difficulty or defect rate.
- Only 439 of 250,000 scanned records (0.18%) met every criterion. That ratio is the point: unfiltered generation is not training-ready.

## Integrity

- `meethealth-sample-v1.jsonl` SHA-256: `a775555c5ac9a7d6d90df9107b1fbb5aaf5508c9b3a0950765ac90dff68f6103`
- Size: 353,745 bytes