Independent systems research · 2026
Spinel Research explores deterministic tensor consensus built around the low-precision matrix arithmetic modern AI accelerators execute best.
The most important arithmetic in modern computing is getting smaller. FP8. FP4. Microscaling. Spinel asks whether that same tensor work can secure a network.
01 / The object
For centuries, spinels were confused with rubies. Some of the world's most famous “rubies” are spinels: rare, structurally distinct, hiding in plain sight.
The name fits a research program looking for overlooked structure in machine compute.
02 / The question
AI accelerators spend enormous silicon area on low-precision matrix multiplication. Proof-of-work requires scarce computation whose result is independently verifiable.
Spinel begins at the intersection: deterministic FP8 first, MXFP4 next.
03 / The standard
Floating point is hostile to consensus. Different kernels, accumulation paths, rounding modes, and hardware can diverge by a bit. A network cannot.
The protocol defines correctness. CUDA and ROCm merely implement it.
Research program
We are not starting from tokenomics. We are starting from the parts that can actually invalidate the project.
Canonical encodings, scaling, accumulation, rounding, and exceptional-value semantics.
Microscaled four-bit computation designed as a portable work class, not a vendor benchmark.
A CPU reference, NVIDIA implementation, and AMD implementation must agree at every consensus-visible boundary.
Protocol-native GPU miners qualified against an independent oracle, then a reference pool, then a public testnet.
Parent block, transaction root, miner entropy.
Canonical operands and an explicit work class.
FP8 GEMM first. MXFP4 follows only after parity.
Compact consensus-visible checkpoints.
BLAKE3 ticket below network difficulty.
A research project
before a currency.
Mainnet is not the starting assumption.
First prove the arithmetic. Then prove the miner. Then earn the network.
Build sequence
Every phase must produce evidence strong enough to justify the next one.
Canonical FP8 arithmetic, challenge derivation, transcript, target semantics.
Independent CPU oracle and public randomized/adversarial test vectors.
Native H100 FP8 miner with exact oracle parity and accepted devnet blocks.
Multiple independent nodes, reference Stratum, explorer, public testnet.
MXFP4 and cross-vendor CUDA/ROCm qualification.
Spinel Research
Independent research into low-precision tensor consensus, mining systems, and useful compute.