Understand the agent. Control the action. Prove the outcome.

A bounded read-only adapter resolves one exact public Agentverse address, preserves registry/search/manifest provenance, and exposes provider-declared facts, conflicts and unknowns. It does not authenticate ownership, test declared endpoints, prove liveness, invoke ASI:One, register an agent, score trust or execute an action.

A bounded adapter normalizes Agent Launch public token list/detail GET responses for explicit chainId 97, preserves raw provider data and surfaces controller/schema drift. It exposes no authentication, tokenization, trading, payment, signing or deployment path; the observed chain-97 list was an honest zero state.

A bounded read-only adapter verifies the allowlisted BSC testnet deployment and can inspect one known transaction hash plus job id without broad log scans. It surfaces provider address/payment-asset drift, signs nothing, and no Ryntra-owned job, transaction or terminal outcome has been observed.

A strict, versioned AgentActionIntent normalizes a sample agent request in source. It does not prove that Agentverse or ASI:One delivered the request, and it creates, funds or submits no external job.

The source preview evaluates one immutable local policy and emits the shared Guard PolicyResult vocabulary. Execution is hard-disabled; an allowed result is not a provider signature, wallet authorization, job or settlement.

The verifier can recompute a source-preview passport that always states OUTCOME_NOT_OBSERVED. It is not a finalized receipt and carries no job id, transaction hash, provider result, evaluator result or independently observed effect.