Decision Firewall: what building an AI governance framework taught me

说明与用途

Decision Firewall 是一个开源平台，用于 AI 行动授权、审计和失败请求恢复，应用于订阅退款处理。它提供持久状态、规则验证和审计，但会增加延迟和集成复杂性。

问题

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.

结果

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.

工作流

• 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

证据

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.

来源

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

来源