AI-Assisted Engineering Should Track the Epistemic Status and Temporal Validity of Architectural Decisions

AI coding assistants outpace human validation: The paper identifies a critical gap where LLM assistants generate architectural decisions faster than engineering teams can verify them, creating a mismatch between decision velocity and validation capacity that current practices don't address.

One in four decisions become outdated quickly: A retrospective audit of two internal projects found that 20-25% of architectural decisions had stale or expired evidence within just two months, demonstrating that decisions made with valid reasoning can become invalid as systems and requirements evolve.

Confidence can be artificially inflated by weak evidence: The paper warns that without explicit mechanisms, multiple weak or uncertain pieces of evidence can combine to create false confidence in decisions. They propose using conservative mathematical aggregation (based on Gödel t-norm logic) to prevent this "trust inflation."

Three concrete safeguards are needed: The authors propose tracking (1) which decisions are verified versus speculative, (2) how confident we should actually be in combined evidence, and (3) when assumptions expire—creating accountability for AI-assisted engineering decisions similar to how we track expiration dates on other critical information.

Responsible AI-assisted engineering requires explicit tracking systems: Rather than assuming AI tools will be used responsibly, the paper argues that organizations need formal frameworks to distinguish conjecture from knowledge, preventing architectural decisions from silently becoming invalid as projects evolve.