MCIN is a Bittensor subnet that decentralizes the inference workload of the HP UMACS™ medical coding AI system. Instead of running inference on a fixed fleet of owned GPU servers, the subnet creates an open market where miners compete to serve the best, fastest, and most reliable medical coding predictions — and earn TAO for doing so.
Doctor's Note (de-identified)
│
▼
Validator samples task
from private benchmark
│
▼ (broadcast)
┌─────┬─────┬─────┐
│ 🖥 │ 🖥 │ 🖥 │ ← Miners (H200 GPU servers)
└─────┴─────┴─────┘
│
▼
Validator scores:
quality × capacity × reliability
│
▼
Yuma Consensus → TAO emissions
Core principle: Emissions follow verified useful inference — not machine count.
| Mechanism | Weight | Measures |
|---|---|---|
| Mechanism 0: Quality × Capacity | 75% | Coding accuracy × verified concurrent throughput |
| Mechanism 1: Service Reliability | 25% | Latency, uptime, format validity |
More H200 servers = more emissions — but only if quality stays high under load.
The validator measures each miner's Verified Concurrent Capacity (VCC): the highest number of simultaneous requests the miner can correctly handle without degrading quality or latency.
# Capacity multiplier (logarithmic — every GPU doubling = fixed bonus)
capacity_multiplier = 1.0 + log2(VCC / 1) * 0.5
# Example VCC → multiplier mapping:
# VCC = 1 (1 GPU — baseline) → ×1.00
# VCC = 2 (2 GPUs) → ×1.50
# VCC = 4 (4 GPUs) → ×2.00
# VCC = 8 (1× 8-GPU server) → ×2.50
# VCC = 16 (2× 8-GPU servers) → ×3.00
# VCC = 32 (4× 8-GPU servers) → ×3.50
# VCC = 64 (8× 8-GPU servers) → ×4.00
# ...unlimited scaling...
# Quality gate (cubic collapse below 85% quality under burst):
quality_gate = 1.0 if avg_quality >= 0.85 else (avg_quality / 0.85) ** 3
# Final epoch score:
score = quality_score × capacity_multiplier × quality_gate × reliability_factorKey property: A miner with 4 servers delivering 95% accuracy beats one with 30 servers at 88% accuracy. Scaling without quality never wins.
quality_score = 0.50 × code_F1
+ 0.20 × modifier_F1
+ 0.15 × hierarchy_score (partial credit for parent codes)
+ 0.10 × consistency_score (same note, different phrasing)
− 0.05 × hallucination_penalty
mcin-subnet/
├── template/
│ ├── protocol.py # MCINRequest / MCINResponse synapse schemas
│ ├── scoring.py # Full incentive mechanism + capacity scoring
│ └── benchmark.py # Gold label set management + task sampling
├── neurons/
│ ├── miner.py # Miner neuron (inference server)
│ └── validator.py # Validator neuron (scoring + weight setting)
├── tests/
│ ├── test_scoring.py # Incentive mechanism unit tests
│ └── test_protocol.py # Schema validation tests
├── scripts/
│ └── run_local.sh # One-command local development setup
├── benchmark/
│ └── sample/ # Sample benchmark cases for local testing
├── .github/workflows/
│ └── ci.yml # GitHub Actions CI
└── pyproject.toml
git clone https://github.com/your-org/mcin-subnet
cd mcin-subnet
pip install -e ".[dev]"pytest tests/ -vchmod +x scripts/run_local.sh
./scripts/run_local.shThis will:
- Start a local Subtensor node
- Create and fund test wallets
- Register and launch a mock miner + validator
python neurons/miner.py \
--netuid <NETUID> \
--subtensor.network finney \
--wallet.name miner \
--wallet.hotkey default \
--inference_endpoint http://YOUR_UMACS_ENDPOINT/predict \
--api_key YOUR_API_KEYpython neurons/validator.py \
--netuid <NETUID> \
--subtensor.network finney \
--wallet.name validator \
--wallet.hotkey default \
--benchmark_dir ./benchmark| Requirement | Details |
|---|---|
| GPU hardware | Any NVIDIA H100/H200 GPU (even 1 GPU is a valid baseline miner) |
| Access | HP UMACS API key (contact Med.Report) |
| Wallet | Registered coldkey + hotkey on MCIN subnet |
| Stake | ≥ 1 TAO on validator to avoid blacklisting |
Each epoch, validators run burst probes — sending 1, 2, 4, 8, 16... simultaneous requests to find your VCC. The highest N where you maintain:
- ≥ 90% of your baseline quality
- p95 latency ≤ 1.5× your baseline latency
- ≥ 97% success rate
...is your VCC. More properly-loaded H200 servers → higher VCC → higher capacity multiplier → more TAO.
- Use FP8 quantization and CUDA graphs for maximum H200 throughput
- Pre-warm the model on startup (avoids cold-start latency)
- Use async batching to serve concurrent validator requests efficiently
- Don't fake availability — burst tests reveal real capacity
Validators are the backbone of the subnet. To run a validator:
- Prepare a benchmark dataset in
./benchmark/withgold.jsonl,online.jsonl, andadversarial.jsonlfiles. - Keep your benchmark private. Never publish your gold set.
- Rotate monthly —
BenchmarkManagerhandles rotation automatically.
See template/benchmark.py for the full benchmark format spec.
All miners implement the MCINRequest / MCINResponse synapse pair:
# Request (validator → miner)
MCINRequest(
job_id = "uuid",
note_text = "De-identified clinical note...",
specialty = "Cardiology",
code_families = ["ICD10CM", "HCPCS"],
deadline_ms = 3000,
)
# Response (miner → validator)
MCINResponse(
job_id = "uuid",
codes = [
{"code": "I21.0", "system": "ICD10CM", "confidence": 0.97, "rationale": "acute MI mentioned"},
],
modifiers = ["acute", "STEMI"],
latency_ms = 312.5,
model_hash = "a3f7...",
schema_version = "1.0",
)| Rule | Implementation |
|---|---|
| Hidden test sets | 30% of scored jobs come from private gold set |
| Challenge-response | No batch downloads; all evaluation is live |
| Duplicate detection | Identical outputs across miners → collusion flag → score = 0 |
| Sybil protection | Throughput bonus saturates; node count never directly rewarded |
| Burst validation | Capacity only counted when quality + latency hold under concurrent load |
btcli subnet burn_cost --subtensor.network finneybtcli subnet create --wallet.name owner --subtensor.network finneybtcli subnet start --netuid <NETUID> --wallet.name owner --subtensor.network finneyHIPAA / PHI: All clinical note text entering the subnet must be fully de-identified before transmission. No real Protected Health Information (PHI) may enter miner traffic under any circumstances. Consult your legal team before deployment.
IP Protection: The HP UMACS model is owner-operated. Miners call the inference API endpoint — they do not receive model weights. Obfuscation alone is not a durable IP control; the API boundary is the primary protection.
Not for clinical use: Outputs from this subnet are for billing and coding assistance only. They must not be used for medical diagnosis, treatment decisions, or any other clinical purpose.
MIT License — see LICENSE.
Built on Bittensor by Med.Report / IFORELS Inc.