Status: v0.1.0, 14 November 2026. Phase 2C of the
master architecture and roadmap.
Reference implementation: @arl/protocol. Built-in
policies: policies/. Test vectors: vectors.json.
A verification policy says how a class of task is checked. An arl-task/1 names its policy in
verification.policy as arl-policy/1:<id>. The contracts enforce the committee shape, windows
and bonds within their own bounds; a policy narrows those bounds and adds the rules validators
follow when they vote. A validator that votes against its policy can be disputed and slashed.
Document
| Field | Rule |
|---|---|
version |
"arl-policy/1" |
id |
lowercase letters, digits and dashes, ≤ 48 |
workload_classes |
non-empty subset of reexec, attested, model-assisted, review |
tiers |
non-empty subset of fallback, reexecution, attested |
committee |
{ min_size, max_size } within 1–7 |
vote_window_seconds |
[min, max] within the evaluator bounds (600–604,800) |
challenge_window_seconds |
[min, max] within 86,400–2,592,000 |
min_bond_floor |
ARL base units; tasks must set at least this floor (Base Mainnet evaluators require 100 ARL) |
bond_covers_budget |
ARL tasks: threshold × max(floor, budget / 10) ≥ budget. Not applied to USDC tasks (bonds are in ARL; it needs a price) |
required_evidence |
evidence kinds every task under this policy must require |
model.allowed |
whether a model verdict may decide an acceptance rule |
model.max_budget |
per token, the highest budget at which a model verdict may decide; required when models are allowed |
model.models |
{ name, weights_sha256 } a validator may use; empty means any model pinned by its weights hash |
A reexec policy never allows model verdicts: re-execution decides.
Validator vote (decideVote)
A validator checks each acceptance rule of the task itself and records an outcome per rule:
by: "check" (its own deterministic check or review) or by: "model" (a model verdict, with the
weights SHA-256). It votes approve only if every rule has an admissible passing outcome. A
model outcome is admissible only if the policy allows models, the task's budget is within the
policy's cap for its token, and, when the policy lists models, the weights are listed. A missing,
failed or inadmissible outcome is a reject, never an approve.
This is how open models such as NVIDIA Nemotron are used: as a validator's tool for checking
non-deterministic work up to a budget cap, with the model and weights pinned in the proof
(arl-proof/1), the vote still bonded and every decision still open to dispute.
Built-in policies
| Policy | Classes | Tier | Committee | Vote window | Challenge window | Min bond floor | Model verdicts |
|---|---|---|---|---|---|---|---|
reexec-standard |
reexec |
reexecution |
3–7 | 1–24 h | 24–72 h | 100 ARL | no |
model-assisted-review |
model-assisted |
attested |
3–7 | 1–48 h | 24–72 h | 100 ARL | yes, up to 100 ARL or 100 USDC |
fallback-review |
review |
fallback |
1 | 10 min–24 h | 24–48 h | 100 ARL | no |
code-tests |
reexec |
reexecution |
3–7 | 1–24 h | 24–72 h | 100 ARL | no |
data-recompute |
reexec |
reexecution |
3–7 | 1–24 h | 24–72 h | 100 ARL | no |
gpu-inference |
attested |
attested |
3–7 | 1–48 h | 24–72 h | 100 ARL | no |
research-sources |
model-assisted |
attested |
3–7 | 1–48 h | 24–72 h | 100 ARL | yes, up to 100 ARL or 100 USDC |
The model-assisted-review model list is empty until the owner pins the model (decision D-6 of
the roadmap); until then any model pinned by its weights hash is admissible within the cap.
Job-type templates
The last four policies are one per common kind of agent work. Each comes with a task template in
templates/: a valid arl-task/1 that fits its policy, with acceptance rules a
client can keep, tighten or extend. The templates change nothing on chain; they run on the same
ARLJobs and ARLVerifiedEvaluator contracts, committees, bonds, windows and disputes as every
other task. What differs is what a validator checks before it votes:
| Template | Evidence the provider submits | What each validator checks before voting |
|---|---|---|
code-tests |
result manifest, logs | Runs the pinned test command on the delivered commit in the pinned container (arl-reexec/1); confirms exit code 0, that test files are unchanged, and that the result hash matches. |
data-recompute |
result manifest, output artifact | Fetches every input by URL and SHA-256, re-runs the pinned analysis, and compares each output hash with the delivered manifest. |
gpu-inference |
attestation, result manifest, logs | Verifies the attestation signature and that it names the model weights, input and output hashes; re-runs a sample of the inputs and compares within the tolerance the task states. |
research-sources |
brief, model verdict, fetch log | Resolves every cited URL, checks the logged content hash, and confirms each source supports its claim. A pinned open model may decide a rule only up to the budget cap. |
A validator records one outcome per acceptance rule and votes with decideVote: approve only if
every rule passed by an admissible check. A wrong vote can be disputed and its bond slashed.