Protocol R0

Tensor work as chain work.

The first Spinel protocol is intentionally small. Chain state derives a deterministic tensor challenge. Miners execute canonical low-precision work, fold a compact transcript, and search for a digest below difficulty.

Candidate pipeline

A legible consensus path.

R0 is a research specification. Parameters shown here are design targets and can change before devnet.

TensorPoW R0

Work must be challenge-bound, difficult to shortcut, parallel enough for modern accelerators, and cheap enough to validate independently.

The protocol defines correctness. CUDA and ROCm merely implement it.
Block contextparent · tx root · nonce domain
→
Challengederive operands + work class
↓
Tensor workFP8 first · MXFP4 next
→
Fold transcriptcheckpoint state + coordinates
↓
TicketBLAKE3(challenge, transcript)
→
Proofticket < target

Consensus objects

TensorWorkProof { version work_class matrix_shape scale_contract challenge_commitment transcript_commitment winning_coordinate ticket proof_payload }

The exact binary format is not frozen. The point of R0 is to make every consensus-visible field explicit before optimizing the miner.

Precision work classes

ClassRoleHardware targetStatus
FP8 E4M3 R0Initial canonical tensor work.Hopper, Blackwell, compatible implementations.Design
MXFP4 R0Microscaled four-bit tensor work.Blackwell, CDNA4, future accelerators.Research
Future classesProtocol-versioned numerical formats.Added only with public parity evidence.Unscheduled

Difficulty and chainwork

Different work classes may have radically different raw throughput. A future multi-class network therefore cannot equate one FP8 operation with one FP4 operation by marketing FLOPS. Chainwork normalization must be an explicit economic and security rule, measured against attack cost and verified work.

Verification hierarchy

We are designing three verification layers:

  • Miner-local parity: optimized kernel against the CPU oracle.
  • Pool share verification: bounded-cost checks suitable for high submission rates.
  • Block verification: consensus validation that ordinary full nodes can sustain.

A design that requires every validating node to own a frontier GPU is not acceptable.

Launch sequence

Devnet

Private, disposable, adversarial

Spinel-controlled miners and nodes. No economic value. Breaking changes expected.

First
Public testnet

Outside miners and pool operators

Open binaries, source, test vectors, explorer, and reference Stratum implementation.

After parity
Mainnet

Only after a public launch window

No private founder-mining period. The launch schedule and miner are published in advance.

Unscheduled