Skip to main content

1.1 What is C-SWON?

C-SWON (Cross-Subnet Workflow Orchestration Network) is a Bittensor subnet where the mined commodity is optimal workflow policy — miners propose multi-subnet execution plans (DAGs), validators score them on task success, cost, and latency, and the network continuously learns the best orchestration strategies through competitive pressure.

"Zapier for Subnets" — The Intelligence Layer for Multi-Subnet Composition

The Problem

Bittensor hosts over 100 specialized subnets covering text generation, code review, inference, agents, data processing, and fact-checking — yet there is no native way to compose them into reliable, end-to-end workflows. Developers today:

  • Manually wire calls to 5–10 subnets per application
  • Guess at optimal routing with no objective benchmarks
  • Rebuild orchestration logic from scratch every time
  • Waste TAO on suboptimal routing through expensive or slow paths

Each new subnet that joins Bittensor increases the orchestration surface area — making the problem worse over time without a dedicated solution layer.

The Solution

C-SWON abstracts away manual orchestration for Bittensor's AI ecosystem. It turns any complex AI task into a single, optimized workflow through competitive pressure:

  1. Validators issue task packages with constraints (budget, latency, allowed subnets)
  2. Miners design DAG execution plans that chain multiple subnet calls
  3. Validators execute these plans in sandboxed Docker containers
  4. Scoring is deterministic: success rate, cost efficiency, latency, reliability
  5. Weights are set per tempo via Yuma Consensus — best orchestrators earn more

The Digital Commodity

The commodity mined by C-SWON is optimal orchestration policy: which subnets to call, in what order, with what parameters, to complete a given task at the lowest cost and highest quality.

The Kernel

Miners submit a WorkflowSynapse (an RPC endpoint receiving task packages and returning DAG plans). Validators pull responses and run a deterministic evaluation to produce weights. That is the entire kernel.

At its core, C-SWON is two Python files:

  • neurons/miner.py — receives task packages, designs workflow DAGs
  • neurons/validator.py — executes plans, scores results, submits weights

Everything else in these docs explains those two things.

How It Prevents Gaming

C-SWON is designed to prevent miners from earning rewards without delivering genuine orchestration intelligence:

  • VRF-keyed task selection — each validator derives its task from hash(hotkey + block), making pre-caching impossible
  • Synthetic ground truth (15-20%) — hidden known-answer tasks verify miners are actually planning, not memorizing
  • Benchmark rotation — tasks are deprecated when >70% of miners score >0.90 for 3 consecutive tempos
  • Deterministic scoring — ROUGE-L, test pass rates, schema validation — no LLM judge, no subjective evaluation

Full anti-gaming details →


← PreviousDocumentation Index
→ Next1.2 Quickstart: Miner
IndexDocumentation Index
Repositorygithub.com/adysingh5711/C-SWON