Perpetual R&D — The Lab: six live experiments

The symptoms are now in the business press. Entrepreneur’s account of software quality as a founder-level problem carries the measurements: AI-co-authored code with roughly 1.7× more issues and security vulnerabilities at up to 2.74× the human rate, per a December 2025 analysis of 470 open-source pull requests — and 25% of Y Combinator’s Winter 2025 startups shipping codebases roughly 95% AI-generated.

What the reporting never does is name the disease — and a failure with no name cannot be specified against or bought against; it gets absorbed and filed under bad luck. This lab names it. Counterfeit code: code that is not what it claims to be. It compiles. It passes review. It reads clean. Its inputs never reach its outputs.

Most of the market answers with a second model — have an AI check the AI. But a probabilistic failure cannot be caught by a probabilistic checker: a model that hallucinates code can hallucinate a passing verdict, which is precisely the defect it was installed to catch. The gate has to be math. That is what CodeRadar is — legitimacy reduced to four invariants, checked deterministically, at file-save speed.