Developer Forward · answer surface
How to Verify AI-Generated Code
Start with the promise the code must keep. Verify the consequential paths against that promise with deterministic evidence where possible, then inspect the parts where correctness still depends on architecture, context, security, or judgment. Do not treat an AI claim of completion as proof that the required state occurred.
Evidence
Cases 33 and 39 — Verification Scales With Consequence; Probably Complete Is Not Complete
- Case 33 shows why verification depth should follow consequence rather than implementation effort.
- Case 39 records three live AI-workflow failures: required sections silently disappeared, an expected workflow transition was skipped, and a later artifact preserved meaning while drifting from the required structure.
- Those failures led to a practical split: people judge substance; deterministic checks should carry mechanical completion and conformance whenever those states matter.
What changes now
A useful verification stack therefore begins with compilation, tests, static checks, and observable runtime behavior, but it does not end there. Confirm that required workflow steps happened, required fields exist, consequential assumptions are true, and the person shipping the change can still explain what they are standing behind.
Boundary
Mechanical validation does not prove good judgment. A perfectly conformant artifact can still embody a bad decision. Verification should remove avoidable checking from human attention so that human review can concentrate on substance and consequence.
Grounded in Ben Chan's canonical Cases 33 and 39. Case 39 is live human/AI workflow evidence from September 2026; hidden model causes remain unknown.