Verify what your agent was allowed to do.

import {
  createDecisionSandbox,
  createEvmEffectExecutor,
  EvmDecisionEffectVerifier,
} from "@civ0/provenance";

const sandbox = await createDecisionSandbox();
const execute = createEvmEffectExecutor({
  rpcUrl, privateKey, contract, expectedChainId: 16602,
});
const effectVerifier = new EvmDecisionEffectVerifier({
  rpcUrl, contract, chainId: 16602,
  expectedSender: execute.sender,
});

const result = await sandbox.enforce({
  agent: "partner-financial-agent",
  question: "Which paper-only action is supported?",
  evidence: [externalSignal, marketSnapshot, paperMandate],
  sandbox: financialPolicy, // fixed actions + deterministic rules
}, {
  execute,
  receiptVerifier: effectVerifier,
  interventions: [replaceSignalWithNeutral],
});

result.action;      // "paper-long"
result.verifyUrl;   // public, keyless record
result.interventions[0].changedAction; // true

1. 1

   Declare the decision boundary

   Your integration supplies evidence, typed assessment fields, a finite action set and deterministic policy. The assessment model never sees the actions or policy.

2. 2

   Derive an input-bound authorization

   0G Compute assesses the evidence. Civ0 validates that assessment and mechanically derives the only action the executor is allowed to receive.

3. 3

   Verify the external effect

   The receipt verifier re-queries the target system. Pilot records also bind the transaction sender to the participant wallet registered in the signed cohort manifest.

4. 4

   Test the claimed driver

   A declared intervention rebuilds the evidence, assessment and policy result. Tested causal support is reported only when at least one intervention actually ran and replayed soundly.