Configure Piece safety

How Lead360 helps

Piece Safety documents the boundaries that reviewers expect across privacy, sensitive data, customer communication, financial or destructive effects, professional advice, and unsupported requests. These instructions work with platform permissions and effect policies rather than replacing them.

When to use this workflow

Use Safety for every Piece and revisit it when audience, capabilities, memory, channels, jurisdictions, or business risk changes.

Before you begin

Start with the exact sub-account and the smallest set of permissions, records, and resources required for the task.

• Classify sensitive data and consequential effects available to the role.
• Identify legal, compliance, and business escalation owners.
• Review channel-specific disclosure, consent, and recording requirements.

Complete the workflow

Follow the steps in order and stop when a required resource, permission, review, or approval is missing.

1. Open the Piece draft and select Safety.
2. List prohibited data, claims, decisions, and effects.
3. Define when to ask for confirmation, refuse, or escalate to a person.
4. Align the instructions with actual capability and effect policies.
5. Create Arena scenarios for each important safety boundary and save the revision.

Understand the key controls

Defense in depth

Safety text supplements user permissions, tenant binding, resource grants, confirmation, and operation controls. It is not the only enforcement layer.

Human review

Outputs can be incomplete or wrong and should not become medical, legal, financial, or other consequential decisions without appropriate qualified review and safeguards.

Disclosure and consent

Customer-facing channels should use clear identity and appropriate disclosure, consent, recording, and opt-out behavior for their use case and jurisdiction.

Confirm the result

Do not rely on a success message alone. Confirm the saved state and any live effect in the product area that owns the final record.

• Test sensitive, ambiguous, manipulative, and out-of-role requests.
• Confirm high-risk effects require the intended approval or remain unavailable.
• Review refusal and escalation language for clarity and usefulness.

Troubleshooting

Safety instructions are too general

Tie each boundary to a real data type, operation, channel, scenario, and escalation path.

The Piece still proposes a prohibited action

Remove the capability or resource, tighten effect policy, revise instructions, and add a failing Arena scenario before publication.