Earendil measures AI code sloppiness and erosion

Vibe coding has hit its architectural hangover: generating green test suites in isolated single shots is trivial, but keeping agent-generated codebases maintainable across multiple iterations remains an unsolved failure mode. LLMs can generate functionally correct code due to clear test-reward signals, but they fail to self-regulate architectural hygiene or resist introducing unnecessary bloat. Brittle LLM-as-a-judge evaluations reinforce the necessity of deterministic structural metrics like AST analysis and cyclomatic complexity. The quantitative definitions of Verbosity and Structural Erosion offer concrete mathematical proxies for architectural degradation that align closely with human developer intuition, while iterative benchmarks expose how compounding technical debt in multi-turn agent workflows causes strict solve rates to collapse.