Refresh dependencies and revalidate the claim
After an earlier decision changes, dependent plans, assignments, context and verification must be refreshed. Revalidation then checks the resulting claim: the original intent, clarified requirement, relevant invariants and affected regression behavior, not merely the most recent local fix.
This closes the Controlled Iteration loop in AI Architect Mastery's AI Driven Development Methodology. Refresh and revalidation follow a human return decision; they do not add phases to the Core.
Why downstream artifacts matter
A corrected PRD does not automatically update a TASK or an agent's context. A test can still assert the old requirement. Without an explicit refresh, implementation may faithfully execute superseded instructions and produce misleading evidence.
How to close the loop
- Identify the changed decision and which previous assumptions no longer apply.
- Review affected PLAN steps, TASK scope, acceptance criteria and verification.
- Mark superseded instructions clearly and provide current context at handoff.
- Execute the revised bounded work under renewed human authorization.
- Check clarified behavior and relevant regressions against the current implementation state.
- Present new evidence and remaining limitations at the human checkpoint.
Constructed example: clarified ownership
The PRD now states how shared-record ownership works. The PLAN must include the affected authorization path, the TASK must reference the clarified policy, and tests must cover both shared and ordinary records. A local shared-record test does not replace the original owner and non-owner checks.
Revalidate the claim, not everything indiscriminately
Choose verification based on the changed decision and its dependencies. A narrow assignment error may not invalidate the whole project, while a shared authorization change may require broader regression checks. State what was covered and what remains unverified rather than equating test volume with confidence.