Cere The Sovereign Context Protocol

The protocol for verifiable AI compute.

Agents can use private data only within permissions signed by the data owner. Each operation creates a verifiable activity record that validators can attest and the protocol can settle.

Identity & delegation OWNS · DATA-WALLET YOUR WALLET OWNS THE VAULT · AGENTS HOLD DERIVED PERMISSIONS layer 05 SR25519 / SECP256K1 signed delegation · scoped + time-bound Global Agent Registry CONSENTS · GAR ED25519 VERIFY ON EVERY WRITE · CONSUMERS FAIL CLOSED 04 immutable versions scope-of-record · audits read before pay Independent validators PROVES · DAC RANDOMISED SUBSETS · QUORUM ≥ 2/3 · ROTATE PER ROUND 03 hashes, never plaintext quorum attestation · validate_node pass/fail Settlement PAYS · FEE-HANDLER DEBIT DEPOSITS · CREDIT PROVIDERS · TREASURY + RESERVE SPLIT 02 ClusterEraPaid / era extrinsics · era finalisation on chain Cere Network L1 ANCHORS · SUBSTRATE DDC-STAKING · DDC-CLUSTERS · DDC-CUSTOMERS · FEE-HANDLER 01 bonds slashable GAR MULTICAST · CONSENT → ALL DDC CLUSTERS
The problem

AI compute is a black box.

You get an answer and a bill. What ran, on which data, at what cost: unverifiable.

Opaque

The execution trail stops at the vendor's API. No receipts, no replay, no independent audit.

Centralized

Model, data, and metering all live with the same vendor. The auditor is the auditee.

Disconnected

Your context is locked inside apps. Agents that need context often have to hand it to whoever runs the model.

Proof of Compute turns AI from a black box into an auditable economic network.

ComponentStatusWhat you can do
Cere mainnetLive since 2021Use the L1 settlement and governance foundation
Consent registry (GAR)LiveRecord and verify signed consent
Dragon 1 storageDDC mainnet since 2024Store data on DDC
Dragon 1 agent computeTestnetBuild and test agent execution
Agent marketplaceIn build · pre-launchNothing public yet; first builders being onboarded
SCP standardDraft v0 · specs being writtenThe five interfaces are on this page; public drafts follow

// who is who: SCP = open standard · Cere = chain underneath · Dragon 1 = flagship cluster · marketplaces = built on top

How the protocol works

From signed access to paid execution.

signrecordattest · DACsettleWalletidentityGrantsigned scopeGARconsent of recordValidatorsproof of compute$CEREauto-settle

A grant is signed once, recorded by GAR, attested by validators against captured activity, then settled in $CERE. Proof of Compute, end to end.

01

Data lives in vaults and streams

Identity is a wallet. The vault, partitioned into scopes, belongs to it. Enterprise or personal, the data never moves to the model vendor.

data := vault[scope]
02

Agents get permissioned access

A grant is a signed agreement: which scopes, for how long. The Global Agent Registry verifies every access against it, and revoking is a signature change.

GAR.verify(grant) → ok
03

Compute runs next to the data

Agents execute on cluster compute, serverless and GPU-backed, inside the permission boundary instead of behind a vendor API.

run(agent, scope) @ cluster
04

Every operation is captured

DAC emits a signed activity record per operation: data touched, compute burned, result hash. Off-chain payloads, on-chain integrity.

dac.capture(op) → record
05

Validators attest it happened

Independent validators attest captured execution records against independently observed activity. Only attested records are eligible for settlement.

quorum.attest(record) ≥ 2/3
06

The money clears

Each window closes automatically: customer deposits debited, operators, validators, and the protocol treasury paid in $CERE.

settle(era) → paid
Choose your path

Four ways in.

Build agents

Publish agents that run against private vaults under signed grants.

start building →

Run compute

Operate a bonded cluster or node and earn from validated usage.

launch a cluster →

Use agents on private data

Run agents on personal or enterprise context without surrendering the underlying data.

explore vaults →talk to us →

Govern

Stake, vote, and help set network parameters on-chain.

participate →
What it makes possible

Five building blocks for verifiable AI execution.

01Private data computeagents run under signed access 02Proof of Computeexecution becomes verifiable 03Signed activity logsevery operation leaves a record 04Automatic settlementverified usage clears automatically 05Marketplace layeragents, attribution, and payouts
Receipts

See a receipt.

What the chain of evidence actually looks like, from signed grant to settled era. Illustrative samples: field shapes follow the reference implementation, not live records.

{
  "agent_service_pub_key": "ed25519:8f3c…a1d2",
  "user_pub_key":          "ed25519:5d9e…44b7",
  "created_at":            "2026-06-10T08:14:02Z",
  "expires_at":            "2026-09-10T08:14:02Z",
  "revoked":               false,
  "metadata": { "scopes": [ { "context": { "vaultId": "v-7c21…", "scopeId": "calendar" } } ] },
  "signature": { "algorithm": "ed25519", "signer": "5d9e…44b7", "value": "0x9b41…" }
}
// signed by the wallet · verified by GAR on every write · status derived at read
{
  "record_id":   "dac-019283…",
  "agent":       "ed25519:8f3c…a1d2",
  "vault_scope": "v-7c21… / calendar",
  "operation":   "read + inference",
  "compute":     { "cpu_ms": 412, "gpu_units": 0.83 },
  "result_hash": "blake2:cf02…77aa",
  "signatures":  ["agent-derived-key", "orchestrator", "agent-runtime"]
}
// multi-signed at source · batched → single extrinsic · payload off-chain, hash on-chain
attestation round · cluster dragon-1 · era 412

  assigned validators : 7  (randomised subset, rotates per round)
  cross-checked vs    : independently observed activity
  agree               : 6 / 7  →  quorum ≥ 2/3  ✓ PASS
// recorded on ddc-clusters · unattested records never reach payout
event ClusterEraPaid { cluster: dragon-1, era: 412 }

  customer deposits   →  debited
  node providers      →  credited per validated activity
  validators          →  paid for inspection
  protocol treasury   →  share accrued
// closes the window · split set per cluster by on-chain governance

// math, not promises: every step above is a signed, inspectable object before anyone gets paid

Use cases

Where verifiable compute lands first.

Enterprise private AI

Agents on enterprise data without the data leaving the boundary. Enterprises can be billed in familiar cloud-style terms while the protocol settles verified usage underneath.

Robotics and autonomous ops

Drone and robotics fleets where every agent decision, escalation, and payload ties to measurable compute and an auditable trail.

Agent marketplace

Publish once, run on private vaults, get paid per proven execution. Verification, attribution, and cost handled by the protocol.

Personal agents

Your own agents on your own data, with memory that survives provider switching. Revoke with a signature.

Why SCP is different

The missing execution layer.

Most AI infrastructure focuses on models, hosting, or orchestration. SCP sits between private data, agent permissions, activity capture, and settlement.

Not just compute

It records what happened: every run leaves a signed, attestable trail.

Not just storage

Agents run against scoped vaults instead of copied datasets.

Not just payment rails

Settlement depends on attested activity, not platform-reported usage.

Not just a marketplace

Marketplaces are built on top; the protocol stays neutral underneath.

Who gets paid

Usage becomes settlement.

Compute, storage, and data activity are measured, verified, and paid out automatically to the participants who power the network.

// indicative split, set per cluster by on-chain governance. developers are paid at the marketplace layer: launch your own marketplace on the protocol and set your own creator split.

Stake

Node operators stake $CERE to join a cluster on-chain through the staking pallets.

Bond

Providers post $CERE bonds against the SLAs and data-availability commitments they declare. Miss them and the bond is slashed.

Settle

Independent validators verify usage and the protocol pays operators automatically. No invoicing, no reconciliation.

Govern

$CERE governs infrastructure commitments on-chain. Trusted parts replicate freely; staked parts scale with the network.

Threat model

Assumed: some nodes misreport or fail; no single validator is trusted. Detected: validator inspection catches misreported activity, out-of-scope compute, and missed SLAs. Consequence: unattested activity never pays out, bonds are slashed, and repeat offenders are removed from the cluster set.

Clusters

Cere is the protocol. Clusters are instances.

// want to run a node with real $CERE rewards? join the Founding 50 →

Run your own cluster, or build on someone else's. Some participants operate infrastructure; many more just use it. Dragon 1 is the flagship cluster, one instance of many by design.

Permissionless

Run the cluster software, register it on-chain through the ddc-clusters pallets, post your bond, declare your SLAs. No permission needed.

cluster software →

Governed

Propose a cluster through OpenGov for shared resources and incentives. Public, transparent, voted by $CERE holders.

propose a cluster →

Supported

Want help with architecture, region, SLA, or compliance? We help teams design and launch tailored clusters.

talk to us →

// every cluster is registered, bonded, and inspected on-chain. browse them on the chain explorer.

The standard

One spec, any implementer.

SCP defines five interfaces for conformant participants: identity and delegation, vault access, event streams, portable memory, and conformance. Cere Protocol and Dragon 1 are first implementers, not owners: the standard is open, and intended to move to an external standards body as it stabilises.

// stage: draft v0 · five surfaces scaffolded · identity-and-delegation leads · an open change process ships with the public drafts

01

Identity and delegation

Wallet-owned identity and signed, scoped, time-limited grants. The load-bearing surface.

delegation
02

Vault access

How agents read and write a user's vault, partitioned by scope.

vault
03

Event streams

Real-time context as a first-class, subscribable surface.

streams
04

Memory and state

Portable agent memory that survives provider switching. The reference implementation calls these containers cubbies.

cubbies
05

Conformance

What a conformant participant must demonstrate, plus audit and lineage.

audit
Build on Cere

Claim a vault in five lines.

Publish an agent once, plug into any user's vault under the grants they sign, and never operate their infrastructure. Reference SDK in TypeScript today, with the CLI and a Go DDC client alongside.

import { VaultSDK, KeypairWallet } from "@cef-ai/vault-sdk"

const sdk = new VaultSDK({ wallet: new KeypairWallet(seed) })
await sdk.vault.ensure()  // claim the vault, bound to your key
await sdk.vault.scope("calendar").publish(event)
// → accepted: signed by your wallet, verified against GAR,
//   activity record captured for attestation

// quickstart: install sdk → claim vault → request grant
//             → publish event → inspect the activity record
npm i -g @cef-ai/cli

cef keypair generate   # your wallet is your identity
cef build              # signed ES module + manifest
cef publish            # listed under your key
import "github.com/cerebellum-network/cere-ddc-sdk-go"

// content-addressed reads and writes against
// your DDC bucket: erasure-coded, client-keyed,
// brokered by the vault's S3 credentials
Participate

Two governance tracks.

$CERE holders vote on-chain over protocol parameters, treasury spends, and cluster updates. Separately, the Cere Network DAO operates as a Swiss association with member voting through its General Assembly.

01

Hold and stake

Stake $CERE to secure the network and back the bonds infrastructure providers post against their commitments.

stake →
02

Vote on-chain

Propose and vote on referenda via Polkassembly, one CERE one vote: protocol parameters, treasury spends, cluster protocol updates, across graduated tracks.

govern →
03

Join the DAO

The Cere Network DAO is a Swiss association: members vote one member, one vote in its General Assembly. A separate track from token voting.

become a member →
04

Operate a node or a cluster

Contribute compute to any cluster through the ddc-staking and ddc-nodes pallets, post a bond, earn a share of validated usage. Or launch a cluster of your own.

clusters →
05

Build and publish

Publish agents once, signed under your own key, or launch your own marketplace on the protocol. Any conformant marketplace can carry your manifest; users bring their own vaults.

build →
06

Claim your vault

Claim a vault, connect agents under grants you sign, revoke with a signature. Your data and your agents' memory stay yours when you switch providers.

ecosystem →
FAQ

Straight answers.

What is Proof of Compute? +
The chain of evidence behind every agent action: the grant that authorised it, the DAC record of data touched and compute burned, and the validator quorum that attested it. Attested execution records, integrity anchors, and settlement gating. It is not a zero-knowledge proof of execution and does not prove model internals; it proves what was authorised, captured, and attested before anyone gets paid. AWS gives you a bill; the protocol gives you the receipt.
Is SCP a blockchain? +
No. SCP is a standard. The Cere Protocol implementation uses a Substrate chain, but another implementation could use a different substrate.
Is it a product? +
No. SCP is not a product and not a single-vendor standard. The first marketplace built on it is in launch readiness, and anyone can build others.
What is live today? +
Cere Network mainnet (since 2021) and the Global Agent Registry are live. The flagship cluster, Dragon 1, has served DDC storage on mainnet since 2024; its agent-compute layer runs on testnet. In development since 2019.
Is there a DAO? +
Yes, and it is distinct from on-chain voting. On-chain, any $CERE holder votes referenda on Polkassembly: one token, one vote, over protocol parameters and the on-chain treasury. The Cere Network DAO is a Swiss association (Baar, Zug) with its own track: members vote one member, one vote in its General Assembly, which amends the articles, approves financials, and admits Observers. Token holders can apply to become members. The first referendum, a token burn, is on-chain for anyone to verify.
How is this different from a normal AI API? +
A normal AI API asks you to trust the vendor's logs and billing. SCP records execution activity, verifies it through independent validators, and settles usage through the protocol. The auditor is no longer the auditee.
Do enterprises need to hold $CERE? +
Not necessarily, by design. Enterprises can be billed in familiar cloud-style terms while verified usage settles through the protocol underneath; cluster operators, validators, and protocol participants settle in $CERE. Fiat onboarding is in active development.
What prevents fake compute records? +
Activity records are signed at source, and independent validators attest them against independently observed activity before settlement. Unattested records never pay out, misreporting slashes the cluster's bond, and repeat offenders are removed.
Can I run my own cluster? +
Yes, that is the point. Two tracks: permissionless, run the cluster software and register it on-chain through the ddc-clusters pallets with a bond; or governed, propose a cluster through OpenGov for shared resources and incentives. Dragon 1 is the flagship cluster, not the only one by design.
Who owns the data? +
The user. The wallet owns the vault, agents hold derived permissions only, and the data is encrypted and client-keyed.

Math, not promises.

Verifiable AI compute is ready to build on. Choose your path.

Build an agent

One SDK, any conformant cluster. Publish once under your own key.

get the sdk →

Launch or operate a cluster

Permissionless or governed. Registered, bonded, and inspected on-chain.

launch a cluster →

Claim your vault

Your data, your agents, your signature. Switch providers without losing memory.

explore the ecosystem →

// questions first? talk to us

Sovereign Context Protocol