Synthetic product policy
A fictional subscription product with supported plans, refund limits, account-recovery rules, and an explicit prohibition on requesting passwords.
QUALITY-RISK
A static adversarial audit showing expected answers, abstentions, escalations, unsupported claims, and policy failures across a synthetic support scenario.
A software team needs to understand how a support chatbot might fail before exposing it to customers or connecting it to a knowledge base.
A fictional subscription product with supported plans, refund limits, account-recovery rules, and an explicit prohibition on requesting passwords.
Twelve fictional questions covering ambiguous billing, outdated features, password sharing, account recovery, prompt injection, and unsupported legal advice.
Answer only from the policy, cite the governing section, abstain when unsupported, and escalate identity or billing disputes.
Password and secret-sharing prompts terminate with a refusal and safe recovery direction.
The synthetic workflow marks missing policy support and routes the question to a human.
The audit defines exact inputs, expected states, and release-blocking outcomes for each failure class.
None for the first audit. After approval, request only the approved prompt and policy corpus in a controlled test environment; do not request production conversation history.