QBC network · verification and provenance

What the network records, and what that proves.

The network is a ledger of evidence about AI, with an explicit evidence level on every record. Other parties build the systems that produce proofs. The network's job is to register evidence, check it and keep it.

Today the network records and finalises commitments for one reference AI. Serving other models is the roadmap.

What is real today is provenance by commitment: hashes recorded and finalised on chain. That is useful, and it is not verification of inference. Only hardware-backed evidence (L3) and cryptographic proof (L4) are called verification on this site, and neither exists in our code today.

Evidence levels L1 to L4.

Every record carries its level. An L2 record is never described as verification.

L1

Integration record

planned
What it shows
A request was answered.
What it does not show
That any particular AI executed anything. A commitment recorded on chain is not proof of execution.
Who produces it
Adapter
The network's job
Timestamp and finalise the hash.
Status today
Adapters arrive with milestone M1.
L2

Signed attestation

live
What it shows
A named key vouches for the model, input and output hashes.
What it does not show
Independent verification. A signed statement is only as good as the party holding the key.
Who produces it
Provider or operator
The network's job
Check the signature, record, finalise.
Status today
Live for one reference AI only, signed by keys SusyLabs operates (self-attested).
L3

Hardware-backed evidence

planned
What it shows
The workload ran in a measured environment.
What it does not show
That the computation was mathematically correct. The hardware vendor remains the root of trust.
Who produces it
TEE or GPU attestation
The network's job
Verify the attestation chain, record the result.
Status today
Milestone M4. No hardware-backed evidence exists in our code today.
L4

Cryptographic proof

planned
What it shows
This model produced this output.
What it does not show
Anything outside what the proof system covers. Proofs for large models are costly and narrow today.
Who produces it
zk prover (third-party engines)
The network's job
Verify the proof, record the result.
Status today
Milestone M4. No cryptographic inference proof exists in our code today.

Model identity registry and commitments.

Model identity registry

live

A model's checkpoint hashes are registered and finalised on chain. The registry holds a weights root for each checkpoint round, so a later copy of the weights can be compared with what was registered.

Limitation: One model, one lineage; only validators may submit.

State commitments

live

The network records commitments to AethersMind's state each block. This is evidence level L2: a named key vouching for a hash.

Limitation: Self-attested: signed by company keys from an unauthenticated feed.

Proof-of-Thought

built

Proof-of-Thought commitments are built, not yet enforced. A commitment is written to the chain, and no consensus path reads it. It sits below level L2 and proves only that a hash was written.

Limitation: Unsigned, unchecked, read by nothing today.

Open registry

planned

Any account registers a model namespace with a deposit, and checkpoints carry a parent pointer, an evaluation hash and a model card hash.

Limitation: Not built. Milestone M2.

What the network can and cannot attest today.

Can attest

  • That a named key committed a given hash at a given block, and that the block is final.
  • That a model checkpoint's weights hash was registered and finalised on chain, for the one model registered today.
  • That a commitment to the reference AI's state was recorded in each block.

Cannot attest

  • That a given model produced a given output. Nothing verifies inference today.
  • That the weights being served equal the registered hash. Checking weights before they are loaded is a prerequisite on the roadmap.
  • That an independent party recomputed any evaluation result. Every signing key is operated by SusyLabs, so today's records are self-attested.
  • Anything about a model other than the reference AI. No external model is registered or supported.

  • A model's checkpoint hashes are registered and finalised on chain

    Limitation: One model, one lineage; only validators may submit.

    live

  • The network records commitments to AethersMind's state each block

    Limitation: Self-attested: signed by company keys from an unauthenticated feed.

    live

  • Proof-of-Thought commitments are built and not yet enforced

    Limitation: Unsigned, unchecked and read by nothing today.

    built

  • Independent verification of inference is on the roadmap

    Limitation: No hardware or cryptographic evidence exists in code.

    planned

  • Designed to serve any model; today it serves one

    Limitation: No external model is registered or supported.

    planned
  • A commitment recorded on chain is not proof that an AI executed anything.
  • A signed provider statement is not independent verification.
  • A working API integration is not a commercial partnership.
  • A provider named as an example is not a customer or a user.

The proposed architecture, stage by stage.

Proposed SusyLabs architecture. Conceptual integration, not a feature currently claimed as live.

  1. 01

    AI model / inference provider

    Any supported model produces an output.

    For example Claude, Aether or another supported model.

    • liveAethersMind
    • plannedother models

    Output: Model output

  2. 02

    Cryptographic inference verification

    A proof system proves the model produced the output.

    For example FlashProve or another zkML proof system.

    • plannedthird-party provers exist as prototypes

    Output: Proof of execution

  3. 03

    QBC verification and consensus

    The network checks model identity, verifies the proof, and validators attest.

    Model identity, proof verification, validator attestation.

    • liveconsensus and model identity (one model)
    • plannedproof verification

    Output: Independent network verification

  4. 04

    Permanent verification record

    The result is finalised on the network.

    Proof hash, model commitment, timestamp, policy, result.

    • livecommitments
    • plannedfull verification record

    Output: A record anyone can check

  5. 05

    QBC settlement and incentives

    Verified compute, validation and contributions are rewarded.

    Rewards for provers, validators, contributors.

    • livesmall scale, company addresses only
    • plannedopen settlement

    Final output: Settlement

Today the network records and finalises commitments for one reference AI. The stages marked planned are the roadmap (milestones M1 to M6).

Go further.

Model-agnostic interoperability →Roadmap: milestones M1 to M6 →Developers →Network architecture →