Status: v0.2.0, 14 November 2026. Phase 5 of the
master architecture and roadmap.
Reference implementation: @arl/protocol
(checkEvidence). Test vectors: vectors.json.
One JSON file that lets anyone check, with only that file and a Base RPC, what was asked, what was delivered, how it was judged and how the money moved.
{
"version": "arl-evidence/2",
"settlement": { "version": "arl-evidence/1", "…": "on-chain record from arl-verifier bundle" },
"task": { "version": "arl-task/1", "…": "…" },
"proof": { "version": "arl-proof/1", "…": "…" },
"policy": { "version": "arl-policy/1", "…": "…" },
"policy_hash": "0x…"
}
Checking a bundle
- On chain (unchanged, existing tool):
arl-verifier verify-bundle <rpc> settlement.json <fromBlock>re-reads every record ofsettlement(job, configuration, submission, votes, decision, finalization, escrow settlement) and exits 1 on any mismatch. - Offline (
checkEvidence), that the four documents agree:- the task, proof and policy each parse under their spec;
policy_hashmatches the policy; - the task follows its policy (
checkTaskAgainstPolicy) and the proof fits the task (checkProofAgainstTask); - the proof names this chain, escrow and job id;
- the job's on-chain description is
arl-task:<task hash> <uri>, its budget is the task's, and itsspecHash, tier, threshold and committee size are the task's; - the submitted deliverable is the proof's (
submittedDeliverable); - votes come only from the committee, once each, and every approval names the submitted deliverable; a committee approval has at least the threshold of approvals;
- finalization and escrow settlement agree.
- the task, proof and policy each parse under their spec;
The result is the outcome (completed, rejected or pending), the number of approvals and
rejections, and overturned: whether the settlement differs from the committee's decision, which
only a ruled dispute can cause.
What a valid bundle shows
That a funded job was opened for exactly this task, the provider submitted exactly this proof, the
committee judged it under exactly this policy, and the escrow paid or refunded accordingly. For a
reexec task the approvals are re-executions of a pinned spec; for other classes they are bonded
judgments under the policy's rules. It does not prove that a non-deterministic result is
semantically correct; that is what the policy, bonds and the challenge window are for.
Storage
Bundles are content: store them where tasks and proofs live (IPFS or HTTPS) and reference them by SHA-256. The storage service itself is decision D-7 of the roadmap.