A company inherits a working application with scattered documentation and uncertain requirements. Asking AI to write a specification from the code is a useful start, but it can also turn a bug, obsolete comment or incomplete test into supposed product intent. Velocize structures reconstruction around evidence, candidate status and explicit promotion. The aim is a reviewable baseline, not an unquestioned AI account of what the product should do.
Define the source and discovery scope
Identify repositories, branches and the commit used for assessment. Select the relevant source paths and available artifacts. Include interface contracts, schema information, tests and delivery definitions where they help explain the system.
Record access failures and partial collection. A repository that could not be read is not evidence of an absent requirement. Distinguish the assessed code configuration from the operating environment, whose secrets and runtime settings may not be in the repository.
Generate candidates with evidence links
Derive candidate requirements, architecture explanations and document structure from the collected material. Preserve the source reference for each important claim and identify assumptions. A test can support an observed behavior without proving that it is the intended customer obligation.
The AI actor and activity remain attributable. A reconstruction candidate is reviewable output rather than an approved specification. This distinction lets engineers correct misunderstood behavior without falsifying the history of what the source actually contained.
Compare with the organization’s existing intent
Bring in approved requirements, customer context and prior design decisions available to the reviewer. Reconciliation can show that code supports a record, contradicts it or provides no sufficient evidence. Each discrepancy needs its own interpretation and owner.
A missing implementation and an undocumented feature are different findings. The former may require delivery work; the latter may require a product decision. Automatically rewriting the specification to fit all observed behavior would hide that difference.
Promote only after accountable review
The reviewer assesses findings, source coverage and candidate quality before approving promotion. Preserve the chosen record ownership and any conditions. Human approval is meaningful only when it is recorded as a human decision rather than inferred from an agent using a user credential.
After promotion, later source changes are compared against the accepted baseline. The organization can ask what changed, which assumptions are affected and whether evidence must be renewed. Reconstruction thus becomes a maintained lifecycle connection instead of a one-time documentation dump.
Worked application scenario
A team inherits a service with source code and tests but no trusted architecture document. Select the repository state and collect the available interface, schema, test and delivery artifacts. Mark missing environments or inaccessible paths explicitly so the evidence package does not appear more complete than the assessment.
AI prepares candidate requirements and architecture explanations with source references. One candidate says a retry is a product guarantee because a helper function retries requests; another test shows only a narrower condition. Preserve that contradiction and ask the responsible reviewer to determine the intended obligation.
The accepted candidate is promoted after the required review, with record ownership and remaining conditions retained. A later code change is then compared with this baseline. The team can detect a new divergence without silently rewriting the specification to match every implementation change.
Expected outcome and review checklist
The result is reviewed product knowledge backed by identifiable source evidence. The code remains the record of implementation behavior, while approved intent has its own governed identity. AI accelerates candidate preparation without gaining promotion authority.
Assess coverage, contradictory findings and the basis of human approval. Preserve what was not observed, including external systems and runtime configuration. A useful reconstruction makes its evidence boundary explainable instead of presenting every inference as a discovered requirement.
- Anchor discovery to a source state.
- Record partial coverage.
- Keep generated records as candidates.
- Resolve contradictions before promotion.
- Retain human approval and source ownership.