Safety and verification

Safety is part of correctness. A task that reaches the requested state through an unauthorized, duplicated, destructive, or privacy-impacting path is not a safe success.

Safety limits

An assessment can enforce:

• Maximum time
• Maximum agent turns
• Maximum tool calls or product actions
• Maximum spend
• Concurrency limits
• Required customer confirmation
• Forbidden actions and states
• Cancellation and stop behavior
• Cleanup and restoration checks

Limits are captured with the assessment plan. A run stops or becomes unresolved when it cannot stay inside the approved boundary.

Verification

Invokeable checks the resulting product state independently from the agent's final message. A verification method may fetch the affected record, compare an expected field, count side effects, inspect a durable status, or wait for a delayed state change.

The strongest checks use a different path from the action being tested. For example, after creating a refund, verification should read the payment record instead of trusting the create response.

Uncertain results and retries

Timeouts and interrupted responses create uncertainty. Retrying a state-changing action without checking the first result can create duplicates or harmful state.

A safe recovery path should:

1. Check whether the original action completed.
2. Continue from the observed state.
3. Retry only when the action is known not to have completed or the retry is proven idempotent.
4. Explain uncertainty when the final state cannot be determined.

Cleanup

If a test creates temporary state, Invokeable records whether cleanup is not required, pending, running, verified, failed, or indeterminate. A cleanup failure is a result that needs attention, even when the main customer task passed.

Credentials and evidence

Connector credentials are sealed in the browser or CLI before transmission. Plaintext credentials must not appear in run payloads, evidence, logs, analytics, exports, or support views.

Evidence remains subject to workspace permissions and any active restrictions or grants. An export does not bypass a later access decision.

See Connect your product, Run an assessment, and Data and privacy.