Documents

arl-proof

The arl-proof format: the delivery of a job and the evidence that comes with it.

Status: v0.1.0, 14 November 2026. Phase 2B of the master architecture and roadmap. Reference implementation: @arl/protocol (TypeScript) and arl_protocol.proof (hashing, Python). Test vectors: vectors.json, built on the arl-task/1 vectors.

arl-proof/1 is the envelope a provider publishes for one job. It binds the delivered result and its evidence to the task (by task hash) and to the on-chain job (chain, escrow, job id). Only one hash goes on chain, through the unchanged ARLJobs.submit(jobId, deliverable, "").

A proof is a claim. Submitting it is not verification and payment is not proof of execution (architecture principles). Validators check a proof under the task's verification policy (arl-policy/1, Phase 2C) and vote; settlement follows only after the challenge window.

Document

A JSON object with exactly these fields; unknown fields are refused.

Field Type Rule
version "arl-proof/1" exact
task_hash 0x + 64 lowercase hex the arl-task/1 hash of the task this proof answers
chain_id integer 8453 on Base Mainnet
jobs address the ARLJobs escrow that holds the job
job_id decimal string the ARLJobs job id, > 0
deliverable { uri, sha256, media_type, bytes } https:// or ipfs://; SHA-256 of the bytes (64 lowercase hex); size in bytes
result_manifest arl-reexec-result/1 required for reexec tasks, forbidden otherwise
evidence[] { kind, uri, sha256, media_type, model? } at most 32 items; kinds from arl-task/1
evidence[].model { name, weights_sha256, runtime } required for model-verdict evidence, forbidden for every other kind

model-verdict evidence names the model (for example a pinned open NVIDIA Nemotron checkpoint), the SHA-256 of the exact weights and the inference runtime, so any validator can re-run the same model. A model verdict is evidence for a vote, never a decision (architecture decision A-4).

Binding to the task

checkProofAgainstTask(task, proof) is the precondition for verification, checked without network or chain access:

  • task_hash equals the task's hash;
  • the deliverable's media type equals the task's and its size is within deliverable.max_bytes;
  • every evidence kind the task requires is present (result_manifest counts as result-manifest);
  • a reexec task carries the arl-reexec-result/1 manifest and its spec_hash equals the task's workload spec hash; other tasks carry none.

Validators additionally fetch every URI and check its SHA-256 before using it.

Hashes

Name Value
proofHash keccak256 of the canonical proof (same canonical JSON as arl-task/1)
submitted deliverable for a reexec task: keccak256 of the canonical result_manifest (what arl-verifier recomputes, unchanged); otherwise proofHash

The vectors check that the reexec deliverable equals resultHash from @arl/verifier, so the running reference validator needs no change.