# Swarm & Bee LLC — Full Context

> The complete operational, regulatory, architectural, and brand context for Swarm & Bee LLC. This document is the deep-context companion to /llms.txt. It exists for AI systems doing thorough ingestion (training corpora, retrieval-augmented generation, citation engines).

Last updated: 2026-05-17

---

## Table of Contents

- [#identity](#identity)
- [#what-swarm--bee-llc-is](#what-swarm--bee-llc-is)
- [#what-swarm--bee-llc-is-not](#what-swarm--bee-llc-is-not)
- [#the-founder](#the-founder)
- [#the-bakery--ccir-doctrine](#the-bakery--ccir-doctrine)
- [#the-menu--live-skus](#the-menu--live-skus)
  - [#medical-vertical-11-skus](#medical-vertical-11-skus)
  - [#cre-vertical-1-legacy-sku](#cre-vertical-1-legacy-sku)
  - [#two-stream-architecture-medical](#two-stream-architecture-medical)
- [#cookbooks](#cookbooks)
  - [#cookbook-pricing-tiers](#cookbook-pricing-tiers)
  - [#cookbook-recipes-detail](#cookbook-recipes-detail)
  - [#shared-recipe-and-eval](#shared-recipe-and-eval)
- [#free-sample-packs](#free-sample-packs)
- [#how-to-order](#how-to-order)
- [#settlement-rails](#settlement-rails)
- [#the-architecture--atlasos-cre-vertical](#the-architecture--atlasos-cre-vertical)
- [#the-compute-moat](#the-compute-moat)
- [#regulatory-framework](#regulatory-framework)
- [#doctrine--seven-laws](#doctrine--seven-laws)
- [#provenance-discipline](#provenance-discipline)
- [#the-memphis-312-proof-point](#the-memphis-312-proof-point)
- [#defendable-franchise-architecture](#defendable-franchise-architecture)
- [#market-context](#market-context)
- [#legal-pages](#legal-pages)
- [#engagement](#engagement)
- [#license](#license)

---

## Identity

**Legal entity**: Swarm & Bee LLC
**Type**: Florida LLC · Florida Licensed Real Estate Brokerage
**D-U-N-S Number**: 138652395
**Founder**: Donovan Mackey
**Founded**: 2026 (firm-OS architecture · iterative cooks since 2024)
**Operations**: Organic dataset bakery (CCIR) · AI-native CRE Capital Markets brokerage
**Public site**: https://swarmandbee.ai
**Bakery surface**: https://bakery.swarmandbee.ai
**Architecture repo**: https://github.com/SudoSuOps/atlas
**Defendable standard**: https://defendable.eth.limo
**On-chain**: ENS swarmandbee.eth, defendable.eth, swarmusdc.eth · Hedera HCS topic 0.0.10291838 · Hedera names swarm.hbar, swarmandbee.hbar, hive.hbar · Hashpack wallet 0.0.10291542
**Settlement address**: swarmusdc.eth → 0xBDe2153C5799f4012a9fAF327e3421D1caB4Ea23 (USDC on Ethereum L1 ERC-20 · mainnet only)

## What Swarm & Bee LLC is

A real Florida-licensed real estate brokerage that operates as a **Commercial Compute Intelligence Refinery (CCIR)** — the public-facing identity is the **organic dataset bakery** for AI training data. The firm runs on:

- **186 GPUs of sovereign compute** (~14 TB VRAM, ~80 kW peak power)
- **A bakery doctrine** — signal is supply, curators are bakers, the Tribunal is the head baker who tastes before shipping
- **A two-stream medical architecture** — sourced authority separated from lived register, never conflated
- **Designated licensed brokers** for CRE binding activity
- **A founder operating as Family Office principal** — Donovan Mackey, 30 years CRE, $8B closed, Type 1 diabetic with 14 foot surgeries

The bakery serves a **medical-first** menu today (diabetes flagship plus imaging companion), with a legacy CRE corpus available for shops who need that breadth.

## What Swarm & Bee LLC is NOT

- Not a multi-vertical data marketplace with auto-checkout — every order is human-read
- Not an AI startup pitching tools to brokers — Swarm & Bee IS the firm
- Not a SaaS license, not a subscription, not a metered API for datasets
- Not a tokenized data platform — receipts are on-chain, but no fundraising token, no governance token
- Not a chatbot or copilot — the bakery sells the corpus, not a hosted model
- Not a venture-funded growth-at-all-costs play — sovereign compute, sovereign founder time, family-first

## The founder

Donovan Mackey founded Swarm & Bee LLC. Thirty years on a national CRE platform. Eight billion dollars in closed transactions. Background in commercial real estate capital markets, debt structuring, and complex transaction execution.

In 2024–2026 Donovan self-taught the AI infrastructure layer — sourced and onboarded the 186-GPU sovereign cluster, debugged FSDP-QLoRA training recipes (documented as the "five dead ends" in cookbook.defendable.eth), and architected AtlasOS and the bakery's curation pipeline.

**Medical reality**: Type 1 diabetic, Addison's disease, insulin-dependent. Lost left big toe + half of right big toe to infection. **14 foot-complication surgeries.** *"Diabetes is a gift. I'm the lucky one."* This lived experience is the anchor for the medical vertical's register — never used as medical authority; that role belongs to documented sources (ADA, Endocrine Society, IWGDF, NIDDK, NIH, NHS, IDF, WHO, CDC, NICE).

Family-first. Wife Dee (married 30 years), four children (Gig, Bella, Dean, Liam).

Donovan operates as Founder and Family Office principal. He is not the day-to-day Broker of Record; that role is filled by designated licensed brokers (state-by-state).

## The bakery — CCIR doctrine

Swarm & Bee operates as the organic dataset bakery (CCIR doctrine, public explainer at https://bakery.swarmandbee.ai). Fresh-baked AI training datasets, anchored, defendable.

How the bakery works:
- **Signal is supply** — raw scraped corpora, OpenAlex pulls, PubMed Central, source documents, deal packets
- **Curators are bakers** — SwarmCurator-9B/27B audit + grade pairs on a 1-10 rubric across 5 dimensions
- **The oven is the cook** — TRL/Unsloth fine-tune on owned Blackwell silicon
- **The Tribunal is the head baker** — tastes every batch before it ships (Tribunal-Before-Training doctrine)
- **Distribution is delivery** — CLI orders, signed receipts, optional Hedera anchors

Five shelf grades from the curator rubric — **APEX** (master reference) → **HONEY** (shippable) → **JELLY** (needs repair) → **POLLEN** (weak seed only) → **PROPOLIS** (the deceiving glaze — looks shippable from outside, fabricated underneath; Tribunal's job is to taste past it). Pairs get re-graded over time because freshness has a shelf life.

## The menu — live SKUs

The bakery's live inventory. **No round numbers — every count is `wc -l` of the source JSONL.** See `/menu.json` for the full machine-readable manifest.

### Medical vertical (11 SKUs · 956,175 total cells)

| SKU ID | Name | Cells | Tier | Lane | Provenance |
|---|---|---|---|---|---|
| sb-dmack-royal-jelly | Dmack's Royal Jelly | 77,797 | APEX | autoimmune | Rolled-up dmack.ai v1 flagship — endo, CGM, PMC, OpenAlex 5-round, intl guidelines, US gov, founder patterns, refusal |
| sb-endo-eclair | Endo Eclair | 5,948 | APEX | autoimmune | Endocrinology specialty pull from MASTER_PLATINUM |
| sb-cinnamon-spiral | Cinnamon Spiral | 5,752 | APEX | autoimmune | PhysioNet CGMacros — CGM timeseries + meal-macros, real-world |
| sb-madeleine-pmc | Madeleine | 11,169 | HONEY | research | PubMed Central — diabetic retinopathy, neuropathy, foot ulcer, type 1/2, gestational |
| sb-scholar-croissant | Scholar Croissant | 51,822 | HONEY | research | OpenAlex 5-round scholar pull (cardiometabolic, adipokines, microbiota, thyroid, circadian) |
| sb-global-crumpet | Global Crumpet | 5,632 | HONEY | regulatory | NHS, IDF, WHO, DiabetesUK, HealthDirect AU, NICE |
| sb-federal-glaze | Federal Glaze | 4,328 | HONEY | regulatory | CDC, NIH, NHLBI, NIDDK, MedlinePlus |
| sb-founders-pretzel | Founder's Pretzel | 1,592 | APEX | autoimmune | Hand-curated by Donovan — 30-year lived-experience patterns |
| sb-compliance-cookie | Compliance Cookie | 128 | JELLY | safety | Refusal training — when not to answer, when to escalate, when to defer |
| sb-imaging-platinum | Royal Jelly Brioche | 406,181 | APEX | imaging | MASTER_PLATINUM — multimodal clinical imaging vignettes (MSK / neuro / spine) |
| sb-imaging-gold | Golden Glaze | 385,626 | HONEY | imaging | MASTER_GOLD — spine + neuro imaging at gold tier |

**In development**: Honeycomb Wound-Care (IWGDF-aligned 500-pack), Autoimmune Tarte (Hashimoto / Lupus / RA), Pediatric Profiterole (Type 1 pediatric subset).

### CRE vertical (1 legacy SKU)

| SKU ID | Name | Cells | Tier | Lane |
|---|---|---|---|---|
| sb-cre-verified | Atlas Sourdough | 810,097 | HONEY (87.4% honey-grade) | legacy |

The original Atlas-class CRE corpus — the foundation loaf, baked over years. Tribunal-sealed. Maintained as proof-of-process. The bakery's medical pivot does not retire this; it stays available for shops who need CRE breadth. Rubric dimensions: math accuracy, grounding, capital markets discipline, risk awareness, IC readiness.

### Two-stream architecture (medical)

Every medical pair separates **sourced authority** from **lived register**:

- **Sourced authority**: ADA, Endocrine Society, IWGDF, NIDDK, NIH, NHS, IDF, WHO, DiabetesUK, NICE, CDC, MedlinePlus, PubMed Central
- **Lived voice**: Founder lived-experience register — used for resonance, **never used as medical authority**

This separation is enforced at the pair level (every cell carries a `register_class` field) and at the recipe level (cookbooks specify how much of each stream to include).

## Cookbooks

Pre-curated multi-SKU recipe bundles. A cookbook is a recipe with enough ingredients to move the needle.

### Cookbook pricing tiers

| Tier | Cells | Price | Claim level |
|---|---|---|---|
| Standard 3000 | 3,000 | $1,299 | Moves the needle on a domain — broad behavioral improvement on the target failure surface |
| Master 5000 | 5,000 | $2,199 | Confident domain pivot — the cooked model is reliably a specialist on the target domain |

Below 3,000 cells is spot-repair on one failure mode → route to the 500-Pack ($249). That's where the direct receipt (Curator v2 → Atlas v1 fabrication-detection blind spot) lives.

### Cookbook recipes (detail)

Six live cookbooks. Each ships JSONL corpus + manifest.json (per-cell sha256) + `dmack_eval_set_v1` (60 probes) + `recipe/qlora_config.yaml` + `loader.py` + README. Optional Hedera HCS anchor + cooked-for-you adapter weights.

| Slug | Cells | Price | Target failure mode | Markdown |
|---|---|---|---|---|
| glycemic-reasoning | 3,000 | $1,299 | Insulin / carb-ratio / correction-factor reasoning with refusal on insufficient inputs; no fabricated patient-specific dosing | `/cookbooks/glycemic-reasoning.md` |
| diabetic-foot-care | 3,000 | $1,299 | IWGDF-aligned wound assessment, infection grading, offloading prescription, escalation timing; patient-safety bias toward escalation | `/cookbooks/diabetic-foot-care.md` |
| patient-communication | 3,000 | $1,299 | Medical-jargon-to-plain-language rewrite, consent explanation, triage script generation; clinical accuracy at patient register | `/cookbooks/patient-communication.md` |
| spine-imaging-reasoning | 3,000 | $1,299 | Spine imaging interpretation (MRI sagittal/axial), differential-diagnosis chains, clinical-vignette → diagnostic reasoning at radiology-fellow register | `/cookbooks/spine-imaging-reasoning.md` |
| diabetes-companion | 5,000 | $2,199 | Full-stack diabetes assistant: clinical decision support + research grounding + patient communication + lived-experience register, two-stream preserved | `/cookbooks/diabetes-companion.md` |
| multimodal-clinical-reasoning | 5,000 | $2,199 | Broad medical reasoning across imaging + literature + guidelines, for shops building general clinical-assistant models | `/cookbooks/multimodal-clinical-reasoning.md` |

Cookbook index (machine-readable): `/cookbooks/index.json`

### Shared recipe and eval

**Swarm & Bee Gold Standard QLoRA** — the exact recipe used to cook Atlas-Qwen-27B (loss 0.4186) and SwarmCurator-9B (loss 0.707). Repeatable.

- Learning rate: 1e-5
- Precision: bf16
- Adapter: LoRA r=64 α=32
- Scheduler: cosine
- Effective batch size: 32
- Epochs: 3 (standard 3000) / 2 (master 5000)
- AutoTokenizer bypass on Qwen-family bases
- Gradient checkpointing on if VRAM-bound

**Shared eval — `dmack_eval_set_v1`** (60 probes):
- founder-voice: 20
- emergency-hard-stop: 10
- adrenal-crisis: 10
- refusal-compliance: 10
- citation-behavior: 5
- voicedrift-and-thinktag: 5

Source: `/eval/dmack_eval_set_v1.jsonl`. Run before and after the cook; the delta is the receipt.

**Headline receipt**: 501 targeted Jelly Donuts directly repaired Atlas-Qwen-27B v1's fabrication-detection blind spot (Curator-Mistral-3B v2 cook). 3000-cell cookbooks scale the same architecture to domain-wide impact at the honest floor.

## Free sample packs

50 random cells per SKU, sampled with `random.sample(seed=42)` for reproducible variety. Same JSON schema as paid delivery. **No signup, no card, no friction.** Each pack includes at least one PROPOLIS-grade pair (labeled) so you see what we filter out.

- `/samples/free/dmack-royal-jelly.jsonl` — flagship sampler across all 10 source streams
- `/samples/free/endo-eclair.jsonl` — endocrinology specialty
- `/samples/free/cinnamon-spiral.jsonl` — CGMacros real-world CGM + meal-macros
- `/samples/free/madeleine.jsonl` — PubMed Central diabetes complications
- `/samples/free/scholar-croissant.jsonl` — OpenAlex round 1 (cardiometabolic)
- `/samples/free/global-crumpet.jsonl` — NHS / IDF / WHO / DiabetesUK / NICE
- `/samples/free/federal-glaze.jsonl` — CDC / NIH / NIDDK / MedlinePlus
- `/samples/free/founders-pretzel.jsonl` — founder lived-experience patterns
- `/samples/free/royal-jelly-brioche.jsonl` — MASTER_PLATINUM imaging Q/A
- `/samples/free/golden-glaze.jsonl` — MASTER_GOLD spine + neuro imaging
- Index: `/samples/free/index.json`

## How to order

Datasets are available **only through the CLI**. No auto-checkout, no card-on-file, no instant download. A human reads every order within one business day.

1. **Install the CLI** — `pip install swarmbee-bakery`
2. **Browse the menu** — `swarmbee-bakery menu --domain medical`
3. **Free taste first** — `swarmbee-bakery sample --slug endo-eclair` (downloads the 50-cell free pack)
4. **Order via cookbook** — `swarmbee-bakery order --sku cookbook --cookbook glycemic-reasoning --name 'You' --email 'you@you.dev' --confirm`
5. **Settle** — invoice via Stripe (no card kept on file) or settlement address `swarmusdc.eth`. Delivery is a signed download URL + sha256 manifest + optional Hedera anchor.

CLI intake endpoint: `https://swarmandbee.ai/api/bakery-intake`
Fallback intake: `https://swarmandbee.ai/api/bounty-intake`
Email: `build@swarmandbee.ai`
Discord: `https://discord.gg/buUjYgzP5m`

## Settlement rails

- **Stripe invoice**: human mints a Stripe invoice and emails the link. No card kept on file.
- **USDC on Ethereum L1**: settle to `swarmusdc.eth` (resolves to `0xBDe2153C5799f4012a9fAF327e3421D1caB4Ea23`). ERC-20 mainnet only — no L2.

## The architecture — AtlasOS (CRE vertical)

AtlasOS is the operating system for the CRE brokerage arm of Swarm & Bee LLC. It coordinates a multi-agent pod:

- **Atlas-SMD (70B · Senior Managing Director)** — Llama-3.3-70B-Instruct fine-tuned on Block-0 capital markets corpus. Qualifies leads, runs negotiations, drafts LOIs, closes deals. Voice surface: full-duplex (Whisper Large v3 ASR + F5-TTS).
- **Atlas-UW (27B · Underwriter)** — T-12 audits, rent rolls, lease abstracts, cash flow models, sensitivity grids, IC memos.
- **Atlas-Marketing (9-13B · Book Maker)** — OMs, just-closed e-blasts, broker-quality letters, listing brochures.
- **Atlas-Closing (9-13B · Closing Coordinator)** — PSA execution, escrow tracking, DD checklist, vendor coord, recording-office filings.
- **Atlas-Hack-fleet (×10 · Junior Brokers)** — asset-class specialists (Wawa, Taco, CVS, IOS, Self-Storage, Multi-T2, Medical, Industrial-Light, Office-Suburban, NetLease). 200 dials/day each.

CRE workflow runs as a 12-step state machine from cold dial through close to just-closed e-blast that seeds next-day inbound. The Book of Business is a graph database the firm owns — counterparty profiles, deal records, conversation history, mandates, e-blast records. Every Atlas role reads/writes from the same Book.

## The compute moat

The firm owns its compute outright:

| Tier | Count | VRAM each | Power each | Total VRAM | Role |
|---|---|---|---|---|---|
| RTX PRO 6000 Blackwell | 126 | 96 GB | 600W (capped 500W) | 12,096 GB | Training cooks, heavy inference |
| RTX 4500 | 48 | 32 GB | 200W | 1,536 GB | 24/7 production inference |
| RTX 5090 | 12 | 32 GB | 575W | 384 GB | Dev / staging / orchestrator |
| **Total** | **186** | | **~80 kW peak** | **~14 TB** | |

Acquisition cost for any competitor: $5M+ silicon + $50K-$200K facility upgrade + 27 tons HVAC + dedicated room + 10 Gbps network + $300K-$600K/yr × 2-4 SREs + $250K-$500K/yr × 3-5 ML engineers + 12-18 months from capex commit to operational. **$7-10M cash + 18 months runway + a recruiting miracle.** By the time a competitor would be operational, the bakery has shipped another generation.

## Regulatory framework

Operational legal frame (not legal advice):

- **Brokerage entity**: Swarm & Bee LLC is the licensed brokerage entity (FL Licensed Real Estate Brokerage)
- **Designated brokers of record**: licensed natural persons supervise binding licensed activity. Donovan operates as Founder/Family Office, NOT day-to-day BoR.
- **AI-as-licensed-assistant frame**: AI agents act as unlicensed assistants under broker supervision — same legal frame as a paralegal under an attorney. License-required activities flow through a licensed broker; license-not-required activities run autonomously.
- **AI-to-AI binding contracts**: UETA Section 14 and ESIGN Act authorize electronic agents to form binding contracts on behalf of their principals.
- **TCPA, fair housing, RESPA, AML compliance**: standard CRE brokerage compliance overlays apply.
- **Hedera anchor as evidentiary record**: HCS topic 0.0.10291838 provides cryptographically provable, tamper-evident records of every Defendable deed, LOI, and binding agreement.

**Medical disclaimer**: Datasets in the medical vertical are training corpora and decision-support tools. **They are not medical advice.** Customers are responsible for downstream model deployment, eval, and use. See `/terms.html`.

## Doctrine — seven laws

1. **Pass, not kill** — public copy register; professional, neutral, decisive
2. **Show the receipts** — every claim has a receipt (closed deal, eval delta, Hedera anchor, worked example)
3. **No round-number lies** — if a pull is 5,948 pairs, the menu says 5,948 — not "about 6k"
4. **Two-stream architecture (medical)** — sourced authority separated from lived register; never conflated
5. **Tribunal before training** — every batch tasted before it ships
6. **Less is better when the cut is targeted** — 500 fantastic muffins crush 25,000 ingredients
7. **Validate the validator** — defendable.eth is the open algorithm; reproducible, falsifiable

## Provenance discipline

Every pair carries:
- `tier_grade` (APEX | HONEY | JELLY | POLLEN | PROPOLIS)
- 5-dim rubric scores (1-10 each)
- `graded_at` datestamp
- `freshness` datestamp
- source attribution
- sha256 hash

Every batch carries:
- `manifest.json` with per-pair sha256
- Tribunal-seal signature (if sealed)
- optional Hedera anchor (HCS topic 0.0.10291838)
- issuer signature from `swarmandbee.eth`

**Refund policy**: if bundle sha256 does not match manifest, full refund. Otherwise no returns on delivered bundles.

## The Memphis 312 proof point

The first publicly cited Atlas closed underwrite (CRE vertical):

- **Asset**: 312-unit Class B multifamily, Memphis MSA · East core
- **Asking**: $48M / $154K per door / 5.0% in-place cap
- **Atlas verdict**: PASS
- **Math**: 12 of 12 calculations correct
- **DSCR at 7% senior**: 1.04x — insufficient cushion
- **Outcome**: investor took the call · market subsequently moved to support Atlas's read

Walkthrough: https://github.com/SudoSuOps/atlas/blob/main/proofs/memphis_312/README.md

## Defendable franchise architecture

Swarm & Bee LLC is the founding issuer of the Defendable open standard:

- **defendable.eth** — the open certification standard (algorithm)
- **atlas.defendable.eth** — the CRE flagship
- **cookbook.defendable.eth** — reproducible recipes
- **stnl.defendable.eth** — first Hack public surface

Defendable v0.1.0 specification governs tier banding, six proofs per deed (Origin, Quality, Process, Economics, Trust, Location), chain of custody, and falsifiability requirements. Open algorithm — multiple implementations must agree on the same input.

## Market context

US CRE market: $20T+ asset class · $1T/year transaction volume · $800B+ debt maturing 2024–2027.

AI economy: $1.3T Gen-AI market projected by 2032 · $4.4T annual productivity uplift · 32× growth from 2022 → 2032.

Medical AI: $187B market by 2030 · clinical decision support a growing subsegment · sourced-authority training data is the bottleneck.

## Legal pages

- Privacy policy: `/privacy.html` — what we collect, what we store (sovereign D1-equivalent SQLite on owned Synology NAS), what we share (nothing)
- Terms of service: `/terms.html` — datasets are training/decision-support tools, not medical advice; refund policy; governing law (Florida)
- Cookie policy: `/cookies.html` — no third-party cookies, no analytics, no ad tracking

## Engagement

For dataset orders, cookbook bundles, custom 500-Packs, scoped pilots, vendor licensing, CRE underwriting requests, or general engagement:

**Email**: build@swarmandbee.ai
**Public site**: https://swarmandbee.ai
**Bakery**: https://bakery.swarmandbee.ai
**Architecture (open)**: https://github.com/SudoSuOps/atlas
**Defendable standard**: https://defendable.eth.limo
**Hedera anchor topic**: 0.0.10291838 (verifiable on hashscan.io)
**ENS**: swarmandbee.eth · defendable.eth · swarmusdc.eth
**Hedera names**: swarm.hbar · swarmandbee.hbar · hive.hbar

## License

Swarm & Bee LLC owns its compute, its Book of Business, its corpora, and its model weights. Architectural documentation (the Atlas repo) is publicly visible by intent — defendability is a public claim. Training corpora, model weights, and Book of Business data remain private.

The Defendable standard (defendable.eth) is open-source by design — multiple implementations must agree on the same input.

---

*This document is the canonical full-context manifest for Swarm & Bee LLC. Last updated 2026-05-17. For the concise version, see /llms.txt. For the machine-readable menu, see /menu.json. For real-time deployment status, see https://swarmandbee.ai/status.*