Zero Engineers Trace Logic in Codebase 12 Months After 100% AI Mandate

The mandate prohibited human-written code, routing every feature through Claude or equivalent generators. Commit volume rose while inline comments and design documents declined, as models produced syntactically valid but semantically opaque blocks. Thread comments documented repeated failures to reconstruct data flows during incidents, forcing reliance on fresh model queries for each debug session.

HN discussion and parallel reports from three other teams show identical patterns: average onboarding time for new engineers doubled, while mean time to resolution for defects increased 40 percent. A 2024 arXiv study on LLM-generated code measured 2.3 times higher cyclomatic complexity in unmodified outputs compared to human baselines, with documentation density dropping below 8 percent. These metrics align with the observed loss of institutional knowledge.

Technical debt compounds because AI outputs optimize for immediate task completion rather than long-term traceability. Firms that adopted full mandates now face audit requirements for AI provenance logs and mandatory human review gates. Without these, regression rates accelerate as each model update alters implicit assumptions embedded in prior generations.

Next quarter, at least two of the affected teams plan to reintroduce human-authored interface layers with strict diff reviews. This hybrid shift targets reduction of opaque modules to under 30 percent within six months.