Skip to main content

Proof of Model Execution

Concept​

Proof of Model Execution (PoME) is the cryptographic mechanism by which Republic verifies that a validator has correctly executed a computational task. This proof is built around a specialized structure called the HashedModel, which captures the state of the model at multiple checkpoints during execution.

HashedModel and Checkpoints​

The HashedModel enhances a neural network model to:

  • Produce deterministic intermediate checkpoints during inference or computation.
  • Compute cryptographic hashes of these checkpoint states.
  • Aggregate these checkpoint hashes into a single root hash that uniquely represents the full model execution.

Verification Process​

  • To verify a proof, verifiers re-execute the model on the provided inputs, computing the checkpoint hashes at each step.
  • By reproducing these hashes and combining them, verifiers reconstruct the root hash.
  • The computed root hash is then compared with the submitted proof hash to confirm correctness.

This re-execution step is essential; it ensures that the proof corresponds precisely to the computation performed, preventing fraudulent claims without actual compute.

Current Status and Future Work​

At present, the protocol primarily relies on benchmark timing and output correctness for validation. Integration of full checkpoint-based proof verification is planned, which will enable distributed, scalable validation through partial checkpoint verification by multiple parties.

Benefits​

  • Cryptographic integrity: The root hash binds the computation’s entire execution history.
  • Reproducibility: Verification through re-execution ensures computations are deterministic and honest.
  • Scalability potential: Checkpoint hashes can be distributed for parallel verification.