Enterprise operations
Enterprise adoption is safest when RunOnProof is introduced as a controlled decision service, not as an unbounded autonomous judge. Start with one action whose current manual rule is already understood.
Adoption stages
1. Define the action
Choose a concrete workflow such as supplier onboarding. Document who requests it, which company identifiers are available, what evidence the organization already uses, who owns the policy, and what happens on uncertainty.
2. Certify synthetic behavior
Run cases for clear approval, ambiguous identity, source outage, policy condition, payment requirement, idempotent retry, and replay. Confirm that REVIEW pauses the workflow and STOP cannot be bypassed by accident.
3. Run a private integration
Connect the customer system with economic actions disabled or tightly controlled. Compare results with the existing process and record disagreements without forcing either system to appear correct.
4. Controlled production
Enable a small cohort, low consequence, or explicit human confirmation. Monitor source health, latency, decision distribution, manual-review reasons, and economic reconciliation.
5. Expand by evidence
Add countries, products, and higher-consequence actions only after the previous scope is stable and its limitations are understood.
Roles and responsibilities
| Role | Responsibility |
|---|---|
| Business owner | Defines the action and acceptable outcomes |
| Policy owner | Names, versions, and approves the decision rules |
| Integration owner | Preserves the contract, retries, and downstream behavior |
| Risk or compliance | Reviews evidence scope and exceptions |
| Operations | Handles REVIEW, incidents, and remediation |
| Finance | Reconciles price, charge, replay, and refunds where applicable |
Control requirements
- Named and versioned policies;
- least-privilege credentials;
- separation between requesting and overriding high-consequence actions;
- immutable decision records and traceable human overrides;
- data minimization and retention rules;
- source-health and contract monitoring;
- bounded retries with idempotency;
- reversible rollout and documented rollback;
- periodic recertification of synthetic fixtures.
Review queue design
A review queue should explain what is missing, show the exact company and action, preserve evidence, and offer only authorized resolution choices. Reviewers should not have to decode raw JSON to understand the problem. A manual approval must record the reviewer, reason, scope, policy, and validity.
Economic controls
Before an agent can buy or execute a paid decision, establish price visibility, budget limits, approval thresholds, single-charge guarantees, replay behavior, and reconciliation. A network retry must not become another purchase. If the promised contract cannot be delivered, the system must not hide that failure behind a successful charge.
Incident response
Incidents can come from RunOnProof, a source, a contract change, customer configuration, or a downstream system. Pause the affected action, preserve committed decisions, identify the exact scope, and avoid broad conclusions. Restore service only after the required evidence and decision behavior are certified again.
Success measures
Useful measures include entity-resolution quality, time to decision, REVIEW rate and causes, source-health availability, prevented duplicate actions, policy consistency, charge reconciliation, and reviewer workload. Do not optimize only for more PROCEED results; a correct REVIEW can be the most valuable outcome.