SwarmJelly-4B

The Royal-Jelly tier router. A 4B-parameter Qwen-class model fine-tuned to extract training pairs from receipts/verdicts and route each pair to its tier in the Royal Jelly taxonomy.

GGUF-quantized for CPU-runnable edge deployment. Lives on the smash spine alongside SwarmCurator-9B and the DefendableRouter.

Apex · Honey · Jelly · Pollen · Propolis. Less is better when the cut is targeted.

What this model does

1. Read a receipt + verdict from the DefendableLedger hash chain

2. Extract a training pair (prompt, response) that the Tribunal already judged

3. Route to a Royal Jelly tier:

   • 🍯 APEX · operator-grade ground truth (the master-baker reference set)
   • 🍯 HONEY · production-ready · shippable
   • 🟡 JELLY · solid corpus material · needs repair lift
   • 🟢 POLLEN · breadth coverage · weak seed
   • 🟤 PROPOLIS · the deceiving shine — looks shippable from outside, fabricated underneath; routed to SwarmFixer

4. Emit a SwarmJelly Pair (SJP- record)* that anchors to DefendableLedger

Status

🛠️ Reserved · GGUF weights pending publication. Model card is live; weights publish after the next Bakery cook cycle (anchored via sha256 to DefendableLedger before release).

Provenance

• Base model: Qwen/Qwen2.5-4B
• Quantization: GGUF · swarmjelly-4b-tribunal.gguf (production)
• Runtime: llama.cpp (python -m llama_cpp.server) — CPU-runnable for edge deployment
• Training corpus: Tribunal-graded receipt/verdict pairs from the live DefendableLedger production spine
• Anchor: DefendableLedger sovereign in-house · no external chain

Live deployment context

SwarmJelly-4B runs on the smash spine via llama.cpp on port 8089. As of 2026-05-25, the spine has produced 24 SwarmJelly pairs on the 82-record DefendableLedger hash chain:

APEX     · 1 record
HONEY    · 7 records
JELLY    · 6 records
PROPOLIS · 10 records
POLLEN   · 0 records (this cycle)

The PROPOLIS-heavy distribution is a healthy signal — the spine is catching false-honey cases at intake time, which feeds the SwarmFixer repair loop and proves the Tribunal is doing real work.

Companion models

• SwarmAtlas-27B · the CRE brain
• SwarmCurator-9B · the Tribunal grader

Companion datasets

• defendable-honey-signals-v0.1 · curated HONEY rows the bakery has shipped
• defendable-compute-proof-receipt-v0.1 · schemas + protocols this model operates within

Use cases

• Edge inference — runs on a HoneyBox appliance (Jetson Orin Nano · $249-$2K) on customer premises; raw data never leaves
• Local Tribunal pass — pair extraction + tier routing without cloud round-trip
• CI/CD hook — gate dataset commits at the merge boundary
• Standalone Royal-Jelly grader — point it at any prompt/response pair, get a tier verdict

Doctrine

• Tribunal before training
• No proof, no honey — PROPOLIS routes to SwarmFixer, never ships raw
• Less is better when the cut is targeted
• Apex · Honey · Jelly · Pollen · Propolis · 5 tiers, no merge, no shortcut

Operator + Contact

Swarm and Bee LLC · Florida · D-U-N-S 138652395 · DBA Swarm & Bee AI

• Email · build@swarmandbee.ai
• X · @swarmandbee
• LinkedIn · Donovan Mackey
• GitHub · SudoSuOps

License

Apache-2.0 · attribution to Swarm and Bee LLC.

---

Apex · Honey · Jelly · Pollen · Propolis. To the shed. 🐝