Per-run spend caps, receipts,
and reconciliation for AI runs
Uncontrolled AI execution is a financial and operational risk. Decisionproof is an execution governance / settlement integrity control layer for AI runs — with per-run spend caps, receipt-backed cost settlement, and a reconciliation loop that rolls every run forward or back, deterministically.
Sandbox private beta — currently access-by-approval — new self-serve sign-ups paused — not the B2B Design Partner offer
Design Partner Program (B2B KR Only): Invite-only / Closed
KR B2B pilot access is handled separately by written scope and manual invoice. Cross-border design partner onboarding is not open before Phase 1 completion.
Per-run spend caps
Submit each run with a per-run spend cap (reservation.max_cost_usd). Budget is reserved before work starts — runs that would exceed the cap are rejected, not truncated. This is the maximum USD reserved for a single run; it is not your monthly, account-level, or workspace-level budget.
Receipt-backed settlement
Cost settles only when the result artifact is written and its metadata is captured. No receipt means no settlement.
Automatic reconciliation
A continuous loop scans for lease expiry, stuck finalize stages, and missing receipts, and rolls each run forward or back without silent money drift.
Audit-ready exports
Run metadata, state transitions, receipt records, and result references are logged. Exports are governed by the applicable retention tier: Sandbox includes Hot online access for 30 days by default; Cold Archive and Deep Archive are available only where included in the customer's plan or contract.
API-key auth
Create and revoke keys from your dashboard. No SDK required. No OAuth flows to manage.
Tenant data isolation
Your run data is scoped to your workspace. API keys only authorize access to your own tenant's runs and receipts.
Not an agent framework. The control layer you run around it.
Decisionproof is not an agent framework, prompt orchestration layer, model router, or model-quality evaluator. Agent frameworks, durable execution engines, and orchestration toolkits cover how an AI run is constructed and resumed. Decisionproof sits one layer up: it governs what a run is allowed to spend, how completion is proven, and how cost settles when a worker crashes, a lease expires, or a retry collides. Initial v1.0 integration is LiteLLM-first; other frameworks should route through the supported proxy path unless separately agreed.
Treat each run as an accountable economic transaction. Decisionproof handles the money, the receipts, and the recovery — so your AI stack clears cleanly, even on the bad days.
Ready to try it?
Sandbox is currently in private beta. Access is by approval only; new self-serve sign-ups are paused while we complete production-readiness review.
Request beta access