ALLOWED FIRST
Typed, non-confidential context
A workflow description, desired outcome, public URLs, synthetic examples, measurable checks, and explicit non-goals.
FAIL-CLOSED BOUNDARIES
Proof first. Least access. Human control.
The beta is designed so the first useful result does not require a customer account, credential, production file, or external action.
REFUSED FIRST
Sensitive or consequential inputs
Passwords, API keys, customer records, regulated data, confidential documents, production identifiers, or high-impact medical, legal, employment, credit, insurance, or safety decisions.
Operational controls
Access
No customer-system access before a synthetic proof and explicit approval. Later access is separately scoped, time-bounded, and read-only by default.
External actions
The AI does not send messages, publish, buy, charge, delete, close tickets, or change production systems without a separate human approval.
Evidence
Every proof labels synthetic inputs, visible assumptions, acceptance checks, failed checks, uncertainty, and what remains unproven.
Isolation
Renter-specific workflows, storage, keys, and entitlements must remain isolated. A failed binding, signature, expiry, scope, or policy check stops execution.
Improvement
Customer work may produce synthetic regression cases. Raw customer content is not placed into public proofs or model-improvement corpora without separate authorization.
Publication
Beta participation does not authorize a case study. Named, anonymous, or synthetic publication requires a separate decision and human review.
AI limitations
AI output may be incomplete, incorrect, or unsuitable for a consequential decision. The proof identifies uncertainty and expected review, but it is not professional legal, medical, financial, employment, insurance, or safety advice.
If safe execution cannot be demonstrated, the request is narrowed, declined, or stopped.