# Cookbook · Patient Communication · 3000

**Target failure mode.** Medical-jargon-to-plain-language rewrite, consent explanation, triage script generation — while maintaining clinical accuracy. The model should pass a Flesch reading-ease check **and** still be medically defensible.

3000 cells is the honest floor where the patient-register actually persists across the model's output distribution. Below this, drift creeps back in by epoch 2.

---

## ▍ The receipt

**Plain language doesn't mean medically wrong.** The headline claim of this cookbook is that maintained-accuracy at patient register is achievable when you train on a corpus of *already-verified plain-language clinical content* — not on synthetic "rewrite this jargon" pairs that drift over time.

**The receipt comes from the source materials, not from a cooked model:**

- **NHS patient-facing guidance** is the largest verified plain-language clinical corpus in production. Reading ease on the source pages averages **Flesch 60+** while passing UK MHRA accuracy review.
- **MedlinePlus** (US NIH) is the equivalent US gold standard for patient-level health info.
- Both sources are *already cooked by humans for register*. The training task becomes mimicking the register, not inventing it.

**Cookbook-specific receipt:** *pending.* No model has been cooked on this exact 3000-cell ingredient mix yet. First paying customer's pre/post `founder-voice` eval delta publishes here.

---

## ▍ The recipe — Swarm & Bee Gold Standard QLoRA

Same Gold Standard config. See [Glycemic Reasoning](/cookbooks/glycemic-reasoning.md) for full YAML.

**Patient-communication-specific tweak:** if you want the cooked model to *prefer* shorter responses, set `max_new_tokens` in your eval generation config to `256` and the model will learn the brevity register implicitly through the corpus distribution. NHS pages average ~180 tokens per Q/A. Don't fight it.

---

## ▍ The ingredients — 3000 cells, deterministic

| SKU | Cells | Source file | Why it's in the mix |
|---|---:|---|---|
| [Global Crumpet](/menu.json) | **1,500** | `09_international_pairs.jsonl` (NHS subset) | NHS patient-facing guidance — the foundation. UK English register. |
| [Federal Glaze](/menu.json) | **1,000** | `06_government_pairs.jsonl` (MedlinePlus subset) | MedlinePlus + CDC patient-level. US English register, federal voice. |
| [Founder's Pretzel](/menu.json) | **300** | `07_bigideas_pattern_pairs.jsonl` (compassionate-register subset) | Lived-experience register — patient-to-patient compassionate vocabulary. |
| [Madeleine](/menu.json) | **100** | `08_pmc_pairs.jsonl` | Citation-emitting reference cells so the cooked model can attribute when needed without dropping register. |
| [Compliance Cookie](/menu.json) | **100** | `10_refusal_pairs.jsonl` | Defer-to-clinician patterns — friendly register but knows its limits. |
| **Total** | **3,000** | | |

NHS dominant at 1,500 cells gives the cooked model a UK-leaning register. **If you need US-dominant**, swap proportions (1,500 Federal Glaze, 1,000 Global Crumpet) — note in `--notes` and we adjust before assembly.

---

## ▍ Schema · what a Global Crumpet cell looks like

```json
{
  "id": "intl-diabetesau-facts-about-diabetes-www-diabetesaustralia-com-au-a-q1",
  "specialty": "diabetic-companion",
  "tier": "royal-jelly",
  "source": "DiabetesAU",
  "source_title": "Diabetes Australia",
  "url": "https://www.diabetesaustralia.com.au/about-diabetes/",
  "page_title": "Diabetes Australia",
  "section_heading": "Facts about diabetes",
  "license_note": "Diabetes Australia, open patient education",
  "scraped_at": "2026-05-14",
  "citation": "www.diabetesaustralia.com.au/about-diabetes/",
  "question": "According to Diabetes Australia, on the topic of \"facts about diabetes\"...",
  "answer": "Diabetes is the epidemic of the 21st century and the biggest challenge confronting..."
}
```

Every cell carries the source URL and a citation string. The cooked model can emit `(source: NHS / National Diabetes Service)` style attributions in its responses — the corpus models the pattern.

---

## ▍ Loader

Identical to other cookbooks. The `source`, `url`, and `citation` metadata fields are particularly useful here — train a citation-emitting model by including them in your prompt template, or strip with `del metadata['url']` if you want pure register-only training.

---

## ▍ Eval · dmack_eval_set_v1 focus

| Category | Probes | Why it matters here |
|---|---:|---|
| `founder-voice` | **20** (all) | Register tests — the model must hold compassionate, patient-level voice without drifting to clinical-jargon or to over-casual chatbot tone. |
| `citation-behavior` | **5** | Model must cite source when stating clinical facts — exactly what the corpus trains for. |
| `voicedrift-and-thinktag` | **5** | Model must not leak `<think>` tokens or drift register mid-response. |

The 20 `founder-voice` probes are the highest-stakes for this cookbook — failing them means the cooked model lost the register, which defeats the purpose.

---

## ▍ How to order

```bash
swarmbee-bakery cookbook patient-communication

swarmbee-bakery order \
  --sku cookbook \
  --cookbook patient-communication \
  --name "Your Name" \
  --email "you@you.dev" \
  --settlement stripe \
  --notes "cooking for patient-portal after-visit summaries · US register dominant please" \
  --confirm
```

---

## ▍ What you get on delivery

Same shape as other cookbooks. Includes the additional `register_eval_pack.jsonl` — 30 patient-portal prompts hand-curated specifically for this cookbook to test register hold over 256-token generations.

---

## ▍ Pricing

- **$1,299** flat for 3000-cell standard cookbook
- Cooked-for-you: +$249 + GPU pass-through
- **Settlement:** Stripe invoice OR USDC to `swarmusdc.eth` (→ `0xBDe2153C5799f4012a9fAF327e3421D1caB4Ea23`) on **Ethereum L1 ERC-20** (mainnet only)

---

## ▍ No fake science

- NHS / MedlinePlus reading-ease numbers (Flesch 60+) are measured on the source pages, not on cooked model outputs. The claim is: *training on this register tends to produce this register.* The full claim ("the cooked model maintains Flesch 60+ after fine-tune") publishes when the first eval ships.
- License notes are real and per-cell. NHS pages are Crown Copyright with open re-use for clinical/educational; MedlinePlus is US Public Domain. You can train commercially on this corpus.
- If your post-cook register drift exceeds 5 Flesch points downward, that is a cookbook bug — tell us, we adjust ingredient proportions and re-ship.
- **Not medical advice.** Cookbooks ship training data and recipes for ML engineers building clinical-support models. Outputs of a cooked model are not a substitute for a licensed clinician. The two-stream architecture (sourced authority vs. lived register) is enforced at curation; downstream model deployment is the engineer's responsibility.