Decision Firewall: what building an AI governance framework taught me

Описание и применение

Decision Firewall – это открытая платформа для авторизации действий ИИ, аудита и восстановления после неудачных запросов, применяемая к обработке возвратов подписок. Она обеспечивает долговременное состояние, проверку правил и аудит, но добавляет задержку и сложности интеграции.

Задача

Developers needed a framework to authorize AI actions, audit outcomes, and handle uncertain execution, especially for subscription refunds where duplicate charges and timing rules apply.

Результат

Decision Firewall achieved zero contract‑violating effects and zero unresolved attempts after recovery in synthetic tests, while a minimal model‑directed arm produced 95 violations across 170 trials. However, the framework added latency (p95 67.89 ms vs 26.52 ms) and required library‑contract workarounds, reducing its integration appeal.

Рабочий процесс

• Interpret the request and classify it as a duplicate‑charge claim
• Authorize the action by verifying ledger entries and policy eligibility
• Dispatch the refund and reserve the balance, recording the idempotency key
• Recover unresolved attempts by reconciling with the payment adapter and updating the audit record

Доказательства

A duplicate-charge claim needs two payment records. Cancellation eligibility depends on timing and usage. A prior refund reduces the amount still available.

The recorded Laya run used CUDA in FP32: 28.42 seconds to load and 47.45 ms median inference across 34 assessments, including the first.

Источник

The minimal arm produced 95 violating effects across 170 request trials.

Источник