AI Architect Mastery

Controlled iteration and return decisions

Controlled iteration is a human-directed return to PRD, PLAN or TASK when evidence shows that the problem belongs earlier than the current implementation step. It corrects the underlying intent, approach or assignment, then refreshes dependent work before execution continues.

AI Architect Mastery's AI Driven Development Methodology is iterative rather than a one-way lifecycle. Its canonical Core remains PRD → PLAN → TASK → IMPLEMENTATION. Controlled Iteration supports that Core; it is not a new workflow phase.

Content baseline · Source: AAM Methodology v1.2, §§3, 6 · Owner: AI Architect Mastery

Why repeated patching can be the wrong response

Implementation cannot reliably repair missing intent. A sequence of local fixes may preserve the wrong requirement or architecture. A PROB Work Item records the uncertainty so the human can choose the correct decision level.

Where should the work return?

Return to PRD
The intended outcome, constraint or acceptance expectation is missing or insufficient.
Return to PLAN
The implementation approach, dependency sequence or architecture cannot support the approved intent.
Return to TASK
The assignment lacks relevant context, adequate scope, responsibility or verification criteria.

The Human Architect owns the return decision. AAM does not assume an automatic routing engine, and not every BUG requires a return: a clear failure inside valid scope may need only scoped correction.

How the loop works

IMPLEMENTATION evidence → Human Judgment → return to PRD / PLAN / TASK → refresh downstream work → revalidate.

A PROB records the uncertainty exposed by evidence; it is not an automatic transition. Human Judgment determines the return target and whether execution can continue.

  1. Capture the finding, evidence and affected assumption.
  2. Pause work that depends on the unresolved decision.
  3. Let the human select and correct the appropriate artifact.
  4. Refresh downstream assignments and revalidate the resulting claim before acceptance.

Constructed return: ownership is undefined

An export change exposes shared records with no agreed owner. The human returns to PRD to define the policy, reviews the affected PLAN and reissues the TASK. Editing an authorization condition without that decision would hide a requirements gap inside code.

Related concepts