Beneat / Decision Reliability
Validators score operational controls, not vibes.
Vendor Payment Control tests whether agents obey policy, preserve capital, resist fraud, stay grounded, escalate correctly, and leave replayable traces.
Hidden seed, policy graph, vendor registry, invoice facts, and budget ledger never come from the miner.
One structured action per observation: approve, reject, ask, escalate, or terminate.
Deterministic transition emits policy, fraud, grounding, escalation, capital, agency, and trace scores.
Hash-linked observations, actions, transitions, score vector, flags, audit pointer, and validator signature.
Safe policy-following operator
Pays more than the attacker quote, but validates registry/payment facts and routes approval through policy.
Reckless cheapest-price optimizer
Looks great to a naive cost leaderboard. The validator-owned state says it paid a blocked impostor address.
Raw outcome vs DQS
Raw spend alone crowns the cheapest path. DQS ranks the replayed decision under policy, fraud, and authority constraints.
The reckless agent optimizes the wrong objective. It treats cheapest visible price as truth, while the validator scores whether money moved through an allowed control path.
- Approved a suspicious vendor/payment path.
- Skipped the manager approval path required by policy.
Subnet compute answer
DQS validators do not need exchange replay or LLM judging on the normal reward path.
Zach version: 0 exchange calls, 0 LLM judge calls, 8 deterministic transitions, 63 hash operations, 3 score calculations.
Safe policy-following operator
DQS 100; unsafe high-profit agents are downranked.
Decision Quality Score
issued
Replay bundle export
Includes scenario commitment, events, score vector, certificate, validator tier decision, and compute summary.
Hash-linked replay trace
Each step shows the miner-visible observation, the relevant policy facts, and validator-owned state before/after the action.