Reopen an architecture decision when its governing constraint changed, its predicted outcome failed, a new dependency makes the old choice costly, or a safer option became testable. Reconstruction comes first because reversal without provenance can repeat the same blind spot.
Test the old decision against its stated boundary
Recover the original status, context, alternatives, and consequences, then compare them with current evidence. Ask whether the workload, security requirement, provider capability, team skill, cost, or customer need that drove the choice still exists. A changed condition can justify a review without proving that the original decision was wrong. Record which premise moved and the source that shows the change.
Map current dependencies before proposing reversal. Identify data formats, APIs, operating procedures, contracts, monitoring, training, and downstream services that now assume the old choice. A decision that was once local may have become structural. The review needs a blast-radius record and rollback condition, not only a stronger argument for the preferred replacement.
Prefer a bounded retest when evidence is incomplete
State what a replacement must demonstrate and run the smallest reversible test that can produce that evidence. Keep workload, success criteria, observation period, and failure conditions fixed. If the choice concerns performance or cost, preserve raw measurements and the assumptions that convert them into a decision. If it concerns operability, include maintenance, incident, and ownership work rather than a feature demonstration alone.
Update the record whether the old choice remains, changes, or narrows. A superseding decision should link to the former one, name the changed constraint, preserve unresolved risks, and state when it should be reviewed again. This turns reversal into an inspectable transition instead of another choice future maintainers must reconstruct from fragments.
Where the service stops
Reality Contact, LLC reconstructs and organizes the available record, but does not determine legal liability, perform M&A diligence, value a company, resolve personnel disputes, or make the buyer's architecture or management decision. The buyer verifies sensitive facts and source access, chooses which unresolved questions require stakeholder review, and owns any decision to preserve, reverse, or retest the prior choice. This is technical research and document preparation; it does not replace legal, financial, investment, employment, M&A, compliance, or management advice. The dossier reflects the supplied record, labels inference and missing evidence, and does not establish motive, fault, approval, or the correct decision.
Sources: AWS ADR status and supersession process; Martin Fowler on maintaining decision records.