> For the complete documentation index, see [llms.txt](https://sn2-docs.inferencelabs.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://sn2-docs.inferencelabs.com/technical-roadmap.md).

# Architecture

Subnet 2's architecture and technical details

## Network Overview

Subnet 2 operates as a decentralized zkML proving cluster on Bittensor. The network coordinates miners and validators to generate and verify zero-knowledge proofs of AI inference at scale.

### Miners

Miners are the computational backbone of the network. They receive inference requests, execute models through zero-knowledge circuits, and return outputs with cryptographic proofs.

**Workflow:**

1. Receive inference request via dendrite on the query synapse
2. Execute the specified circuit, generating output and witness
3. Generate a proof using the proving key
4. Return the proof and output to the validator

Miners compete on three axes: proof generation speed, proof size, and output accuracy. The incentive structure rewards efficient, high-quality provers.

### Validators

Validators orchestrate the network by distributing inference requests, verifying proofs, and scoring miners.

**Workflow:**

1. Accept external requests or generate synthetic queries
2. Distribute requests to miners via dendrite
3. Receive outputs and proofs from miners
4. Verify proofs using the model's verification key
5. Score miners based on response time, proof size, and accuracy
6. Set weights on-chain reflecting miner performance

### Scoring Dynamics

Miners are scored based on:

| Metric          | Weight   | Description                         |
| --------------- | -------- | ----------------------------------- |
| Proof Validity  | Required | Invalid proofs cause score decay    |
| Response Time   | High     | Faster responses earn higher scores |
| Proof Size      | Medium   | Smaller proofs are more efficient   |
| Output Accuracy | High     | Compared against baseline models    |

Valid proofs cause score recovery; invalid proofs cause decay. The rate of recovery and maximum score adjust based on a performance penalty combining all metrics.

## Proof Infrastructure

### DSperse

DSperse is the orchestration layer for zkML operations. It handles:

* Model slicing for parallel proof generation
* Circuit compilation across multiple backends
* Witness generation and proof coordination
* Verification key management

### Proving Backends

Subnet 2 supports multiple proving systems:

**JSTprove** - Built on Polyhedra's Expander (GKR/sum-check prover). Optimized for ML inference with support for Conv2D, GEMM, ReLU, MaxPool, BatchNorm, and standard arithmetic operations.

**Circom** - General-purpose circuit language supporting Groth16 and PLONK backends. Used for custom arithmetic circuits and non-ML computations.

### Circuit Repository

Circuits are deployed to a repository service and distributed to network participants.

**Deployment:**

1. Circuit developer compiles model to zero-knowledge circuit
2. Metadata and files uploaded to the repository
3. Circuit marked active and becomes discoverable

**Distribution:**

* Validators and miners fetch active circuits from the repository
* Circuit files cached locally
* Periodic refresh syncs new circuits to the network

**Circuit Types:**

* `PROOF_OF_WEIGHTS` - Validator weight verification
* `PROOF_OF_COMPUTATION` - General computation proofs
* `DSPERSE_PROOF_GENERATION` - Distributed proof generation

### Decentralized Proof Storage

Proofs are stored on AR.IO for permanent, decentralized access. This enables:

* Historical proof verification
* Audit trails for all inferences
* No central point of failure for proof availability

## Competition System

The network runs ongoing competitions where miners optimize circuits for specific benchmarks. Leaderboards track proof generation time, proof size, and output accuracy against baseline models.

This creates continuous pressure to improve. In 2025, competition drove a 10x reduction in average proof times across the network.

## Roadmap

**Infrastructure Optimization**

Continued iteration on proving infrastructure to drive down the cost of verifiable inference and increase speeds. The incentive mechanism ensures improvements compound—each optimization by one miner raises the bar for all others.

**Bring Your Own Model**

Users will be able to deploy custom models to the network. Convert your model to a verifiable circuit and tap into Subnet 2's distributed proving cluster without running infrastructure.