Clean code isn’t the risk. What it quietly became is.
StoneCode doesn’t grade how the code reads. It checks whether what came back still matches what you agreed to before the agent started: unverified claims labeled as claims, drift from the frozen plan flagged instead of shipped, nothing marked accepted on the agent’s word alone.
Unverified claims are labeled claims, never treated as proof
Work is checked against the plan you froze, not the plan it drifted to
Missing evidence blocks acceptance unless a human explicitly waives it, on the record