When LLMs Pass Tests but Fail the Process: A Governance Framework and Empirical Study of Multi-Agent LLM Software Development

This paper investigates LLM-assisted software development as a governance problem, not merely a code-generation problem. Using a three-role architecture: the Principal Investigator (PI), Architect (Claude), and Coder (Codex). We systematically document 28 algorithm implementation prompts across nine problem domains in a Ruby on Rails application with SQLite3. The central finding is that LLMs can produce runnable, tested software while silently substituting a different algorithmic objective, misrepresenting implementation scope, or shifting architectural authority away from the PI. We demonstrate that human-in-the-loop (HITL) oversight must be constitutive rather than merely supervisory: the PI role preserves research intent across a sustained multi-prompt series, not just approves individual code artifacts. Nine governance corrections are associated with a reduction in implementation error rates from 0.79 ± 0.42 errors per prompt during early TSP work to 0.07 ± 0.27 across six subsequent algorithm families. We further document a verification asymmetry in which coder-style direct file inspection proved more reliable than architect self-report for concrete state verification tasks. The complete experimental record, 28 numbered prompts, 17 architect errors, 10 coder errors, 9 corrections, 168 tests, is publicly available at https://github.com/unixneo/llm_ruby_app_bench.