Vol. I — Testnet Edition Base Sepolia · Open Repository
CLOUDANA
A Journal of Useful Computation — published as a working experiment
Testnet
edition

The proof and the product are the same operation.

Cloudana is a decentralized compute network in which the work that secures a provider's reward is the computation a user paid for. No wasted puzzle. No "just trust the provider." One matrix multiplication does both jobs — and this page performs it, live, below.

computing A · B block 0/0
×
Seed σ · fetching latest base sepolia block… Transcript · real SHA-256 chain over computed blocks Proof z · SHA(σ ‖ nonce ‖ transcript ‖ H(A) ‖ H(B)) awaiting transcript…
Plate I A cuPOW miner at demonstration scale (n=32, r=8, d=8), executing in the reader's browser via WebCrypto. Not a simulation.
Edge API · live check checking…
Settlement chain Base Sepolia · testnet
Verification model orchestrator-attested
zk verifier Groth16 · in development
§ 1

One multiplication, two outputs.

The hardness of Cloudana's proof comes from the full transcript of intermediate computation blocks — not from the answer.1 You can shortcut the answer with low-rank tricks; you cannot shortcut the transcript. So the proof is only valid if the provider genuinely did the work.

And that work is the user's matrix multiply — so the same operation yields both a proof you can't fake and a result somebody actually wanted. That single property separates useful work from every hash lottery that preceded it.

1. After Komargodski & Weinstein, Proofs of Useful Work from Arbitrary Matrix Multiplication (2025). Our construction is cuPOW; the module is public in the repository.

§ 2

Method: how a certificate is mined.

Encode

Derive low-rank noise E, F from the block-hash seed σ. Form A+E and B+F so the work cannot be precomputed.

Multiply & record

Compute the product in r×r blocks, hashing every intermediate block into a transcript. This is the part that cannot be shortcut.

Prove

z = hash(σ, transcript, H(A), H(B)). A certificate is valid when z meets the network's difficulty target.

Decode

Recover C = A·B cheaply. C returns to the user as the answer. Reward is paid only for work backed by a real workload.

Table 1 — cuPOW certificate · testnet specimen№ 00001
Matrix n512 × 512
Block r8 × 8
Transcript blocks4,096 hashed
Wall time1.84 s
Seed σ · latest Base block hash0x8f3a…c41e · unpredictable, 5-min validity
Proof z · difficulty 16 bits0000b9f31e4a7d20c685f9e0b3a1d47c28e5f60a
Backed by workloadtrue — reward gated on a real, paid job
Replay protectionz recorded in usedZ — a certificate spends once
source: pouw/src/cupow.ts — illustrative specimen, not live network data
Verified
§ 3

Experiment: reproduce our results, now.

Good papers invite replication; this one ships the apparatus. Press run: your browser multiplies two real 64×64 matrices in 8×8 blocks, chains every block through SHA-256, seeds σ from the latest Base Sepolia block, and grinds a nonce to the difficulty you choose. Then it re-verifies its own certificate independently — the same redundancy check our testnet orchestrator performs.

idle — choose difficulty and run block 0/64
difficulty
Seed σ
Transcript
Nonce attempts
Hash rate
Elapsed
Proof z
Verification
demo scale n=64, r=8 · source: pouw/src
Plate II The reader's replication bench. Difficulty selectable; verification independent.
what a node seesauto-detect
GPU
detecting…
WebGPU
CPU threads
Memory
Measured rate
run the miner ←
TIER — · AWAITING SCAN
The plug-and-play principle: hardware describes itself, Cloudana classifies it. A real node agent measures far more than a browser can. Register real hardware →
Plate III Instrument self-description — the browser-visible slice.
§ 4

Economics: no hard cap — a soft cap set by usage.

Emission (M CLD/yr) Burn · 2% of payments Supply
burn > emission soft cap — set by usage y0 y6 y12
Fig. 1 Traced from the emission model — 8% → 15%/yr decay → 1.5% tail; burn at 2% of volume on an illustrative adoption curve. Crossover timing depends on real usage.
  • EmissionStarts at 8%/yr, decays 15%/yr, never below a permanent 1.5% tail — providers are always paid to keep the network alive.
  • Burn2% of every workload payment is destroyed. Users pay real value for real compute, so scarcity tracks usage, not speculation.
  • EquilibriumHigh usage → burn outpaces emission → deflation. Low usage → the tail keeps the lights on. The network self-regulates.
WORKLOADuser pays in CLD ENCODEσ-seeded noise · A+E, B+F MINEtranscript forms block by block VERIFYz meets difficulty · replay-locked SETTLEmint → providerburn 2% → everyoneresult C → back to the user
Fig. 2 The full loop. The same multiplication carries the job and the proof down this line — the mark turns from red (work) to blue (verified) at the moment of verification. Plates I & II run the MINE and VERIFY stations for real.
§ 5

Status, stated plainly.

Today, a trusted orchestrator verifies each certificate by re-running it — a normal way to launch, but not "completely decentralized," and we won't call it that. Full trustlessness arrives when a zkSNARK verifies proofs on-chain and the orchestrator leaves the critical path.

We'll claim decentralized verification when it's true, not before. In this field, saying exactly where you stand is the rarest feature — so it's ours.

Mainnet ships when all four gates clear — not earlier
Third-party audit of the emission & burn contractsopen
Legal review of CLD's classificationopen
zkSNARK transcript verification live on-chainopen
End-to-end proof: real workload → certificate → mint → result deliveredopen

Correspondence.

Q.01What is Proof of Useful Work?

A consensus mechanism where the computation that proves a provider did work is the same computation a user paid for. cuPOW derives proof hardness from the full transcript of a matrix multiplication — so the proof is only valid if the provider genuinely computed the user's result.

Q.02How is the work verified?

On testnet, an open-source orchestrator re-verifies each certificate. On mainnet, a Groth16 zkSNARK verifies proofs on-chain in O(1), removing the orchestrator from the critical path.

Q.03What gives CLD value?

Every CLD reward is backed by a paid workload — users pay real value for real compute, and 2% of every payment is burned. Emission is tied to work performed and scarcity to usage, not to a liquidity pool or speculation.

Q.04Is Cloudana fully decentralized?

Not yet, and we say so. Testnet is trust-minimized and orchestrator-coordinated. Full decentralization arrives when zkSNARK verification is live on-chain — that's the mainnet milestone, and we won't claim it before it's true.

Q.05How do I provide compute?

Install the node once. It auto-detects your CPU, GPU, RAM and disk, classifies its tier, and self-registers on-chain. No spec forms, no bidding marketplace.

Q.06Is there a hard cap on CLD?

No — by design. A useful-work network must always pay its providers, so emission decays to a permanent 1.5% tail instead of halting. The 2% burn on payments means high usage turns supply deflationary: a soft cap set by the market, not a hardcoded number.

Flash your hardware. Cloudana figures out the rest.

Install once. The node detects your CPU, GPU, and memory, classifies itself, and registers. No spec forms. No bidding. Just useful work.