The Audit
Matrix.

Every change traces to something you actually asked for — no silent scope

The set of touched files matches the approved outline

Every identifier the agent wrote exists in the real system — none invented

Every write to pre-existing data captured the original in the same operation

Resets restore reality, not defaults

Destructive paths are reversible or explicitly signed off

Duplicate submissions cannot double-apply side effects

The full user workflow was executed end-to-end in a real environment

The error paths were exercised — timeout, missing row, double submit

Edge states (empty, first-run, expired) render and behave correctly

Success was verified against observed results, not the agent’s summary

Scheduled jobs ran on schedule in the deployed environment — verified, not assumed

Job failures alert somebody; silent failure is impossible

Time-dependent logic was tested at its boundaries, not just “now”

Agent and integration credentials are least-privilege and scoped

Secrets live in config, never in code, prompts, or logs

Untrusted input cannot steer the agent or the app off-course

Repository, platform dashboard, and runtime all agree on what shipped

The deployed build is the audited build — verified at the runtime surface

Platform-side config (env, integrations, webhooks) matches the documented state

A written record exists of what changed, what did not, and what was deferred

Real elapsed time was tracked from timestamps — never estimated