Understand AI-generated code — before you merge it
Claude, Cursor or Copilot just produced two thousand lines, and every line looks plausible. That is exactly the problem: AI-generated code fails at the structure level, not the line level. The bugs live between files — a cycle here, a bypassed layer there, an old code path nobody deleted — and a diff view shows you none of it.
prism0x2A scans the repository on your machine and answers the questions a reviewer actually has. Local-first is the point: your code never leaves your computer, which also means you can run it on the codebase your employer would never let near a cloud tool.
What did it actually build?
An AI edit rarely stays where you pointed it. The Blueprint Explorer draws the modules and dependencies as they are now — so a change that quietly grew a new layer, or wired the UI straight into the database, is visible as a shape, not a surprise in review.
Did it create cycles?
Import cycles are the classic AI-code failure: each file looks fine, the loop only exists between them. prism0x2A finds every cycle in the graph and names the files in it — the kind of defect a diff view structurally cannot show you.
Is anything now dead?
Generated rewrites love to leave the old path behind. Dead exports and unreachable modules show up in the scan, so the cleanup happens while you still remember why the code existed.
Does the documentation still tell the truth?
The AMBER layer reads code and docs side by side and flags where they disagree. After a week of AI-assisted edits, this is usually the first place drift shows.
The honest part
prism0x2A does not review your logic — no tool that has never seen your requirements can. What it does is make the architecture visible and the drift measurable, with a published methodology you can recompute yourself. The judgement stays yours; the map is ours.