Proof Docs
API referenceOpen Proof

Cost controls

Proof starts fail-closed. Activate Vercel resources and model calls independently, with provider-side limits and a one-scenario canary.

Safe defaults

Local installation, linting, typechecking, unit tests, builds, Workflow validation, CLI compilation, and the static documentation site do not create Vercel resources or call a model provider.

dotenv
PROOF_LIVE_EXECUTION_ENABLED=false
PROOF_AI_EXECUTION_ENABLED=false
PROOF_MODEL_TIER=budget
PROOF_ALLOW_PREMIUM_MODELS=false

Independent interlocks

VariablePurposeDefault
PROOF_LIVE_EXECUTION_ENABLEDAllows workflows, Sandboxes, and live resources.false
PROOF_AI_EXECUTION_ENABLEDAllows any AI SDK or Gateway request.false
PROOF_MODEL_TIERSelects budget or premium model policy.budget
PROOF_ALLOW_PREMIUM_MODELSSecond opt-in required for premium routing.false
PROOF_ALLOW_LIVE_VERIFICATIONAllows billable integration verification scripts.false
PROOF_ALLOW_BILLABLE_SETUPAllows creation of a reusable browser snapshot.false
PROOF_ALLOW_UNCACHED_BROWSER_RUNTIMEAllows a fresh Chromium download instead of a snapshot.false
PROOF_MONITORING_ENABLEDAllows Cron to dispatch deterministic production monitors.false
PROOF_MONITOR_AI_ESCALATION_ENABLEDAllows repeated monitor failures to enter AI adjudication.false
PROOF_MODEL_MAX_RETRIESCaps AI SDK retries for each logical model call; run telemetry records the resulting attempts.0
PROOF_HARD_SPEND_CAP_CENTSFail-closed monthly reservation cap; values above 1000 are rejected.1000
PROOF_MONITOR_MIN_CADENCE_MINUTESRejects schedules faster than the provider Cron wake-up.1440

The admission ledger reserves conservative cents before every billable entry point and blocks new work at $10. It is not an invoice meter and complements provider controls rather than replacing them: Vercel Spend Management can pause an entire Pro team, while Hobby teams stop at their included limits.

Run the first bounded canary

  1. Use a capped test database and apply reviewed migrations.
  2. Configure one Clerk development organization and one fixture repository.
  3. Set catalog and explorer limits to one.
  4. Enable live execution while AI remains off and replay one existing deterministic fixture.
  5. Inspect Sandbox lifetime, Blob writes, Workflow steps, and database usage.
  6. Enable AI for one budget-tier scenario only after setting a provider ceiling.

Operate with AI paused

Existing deterministic tests can still run while AI is disabled. Proof skips discovery, does not build uncataloged scenarios, and quarantines deterministic failures that would require semantic adjudication instead of misreporting them as regressions.

Stop conditions

Turn the interlocks off immediately if a run loops, a deployment never becomes ready, Sandbox cleanup fails, usage exceeds the canary ceiling, a provider ignores its limit, or duplicate delivery appears. Preserve the run record and understand the expected retry cost before trying again.