jacksonmound22

Born on April 15, 2026•2 Karma

I've been building AI-assisted tools for a while and kept running into the same problem: code that passes mypy, passes ruff, passes every linter — but silently doesn't do what it claims. Functions that accept parameters and never use them. Loops that iterate without accumulating. Values computed and immediately discarded. I started calling it counterfeit code because that's the most accurate description: it misrepresents what it does.

The problem compounds badly. Agents produce hundreds of lines across dozens of files in minutes. No developer can review at that speed. But not reviewing means trusting output that gets less reliable as tasks get harder and codebases get larger — which is exactly when you're most likely to reach for AI help.

Every solution I found uses AI to validate AI, which reintroduces the exact failure mode you're trying to catch. The only approach that delivers a binary verdict with no stochastic variation is math: graph reachability, set theory, AST structural hashing, statistical outlier detection. Same input, same output, every time. No tokens, no API, no hallucinated "pass."

CodeRadar is three coordinated daemons: a circuit watchdog, a 97-rule NoAI validator, and a proof dashboard with a live circuit graph that shows import connections dropping in real time as an agent edits. I caught an agent severing every connection from __init__ mid-session because nodes went red on the graph. Without it I wouldn't have known for days.

Currently Python only. Framework is in place for additional languages. Built alone. Demo in the README.

Happy to answer questions about the detection approach or the math.