Cookbook · Patient Communication · 3,000 cells

$1,299 · standard 3000 · lane

Medical-jargon-to-plain-language rewrite, consent explanation, triage script generation. Maintains clinical accuracy at patient-level register.

The receipt

Receipt pending — recipe designed against the dmack.ai senior-hack review staging. NHS / MedlinePlus pulls are independently verified plain-language baselines (Flesch reading ease > 60 on source material). Cookbook-specific pre/post receipt publishes when first paying customer runs the eval.

Ingredients · 3,000 cells, deterministic

Order this cookbook

swarmbee-bakery order \
  --sku cookbook \
  --cookbook patient-communication \
  --domain medical \
  --name "Your Name" --email "[email protected]" \
  --settlement swarmusdc \
  --confirm

Settlement: Stripe invoice or USDC to swarmusdc.eth (0xBDe2153C5799f4012a9fAF327e3421D1caB4Ea23) on Ethereum L1 ERC-20 mainnet only.

Full recipe + schema + loader + eval pattern

The complete markdown recipe is at /cookbooks/patient-communication.md — includes the Gold Standard QLoRA config (used to cook Atlas-Qwen-27B loss 0.4186, SwarmCurator-9B loss 0.707), schema reference with quirks, drop-in Python loader, and the 60-probe dmack_eval_set_v1 grading pattern.

Preview the cookbook (first ~8KB rendered)

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

    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

    {

    "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

    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

    swarmbee-bakery cookbook patient-communication

    swarmbee-bakery order \

    --sku cookbook \

    --cookbook patient-communication \

    --name "Your Name" \

    --email "[email protected]" \

    --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.