CAISI

Independent research and operating notes on AI agent governance.

Home / Field Notes / Gait Methods Archive

CAISI Methods Archive / Gait

Policy Before Action

This four-part methods archive uses Gait repo patterns as implementation context for a problem many teams still avoid naming clearly: policy only becomes real when it can change what an agent is allowed to do before the action executes. YAML policy, boundary verdicts, signed traces, and CI regressions matter because they turn AI governance from advisory language into operating discipline.

The reader output is a concrete boundary model: allowed, approval-required, and blocked actions with portable evidence and CI/CD regression checks.

Boundary before side effect Signed traces, not screenshots Fail-closed runtime control
Start the series Repo source See the Wrkr series

Why keep Gait in the methods archive

The CAISI research and operating-model posts already argue that execution boundaries matter. This collection narrows the lens to the enforcement layer itself: the moment a tool call is allowed, blocked, or held for approval, and the artifact trail that proves what happened.

That is worth separating because teams routinely collapse policy into prompts, style guides, or after-the-fact observability. Gait is useful implementation context precisely because the repo is explicit about a different model: tool-boundary verdicts before side effects, signed evidence, and CI regressions that turn incidents into durable tests.

The 4 posts