Phase-Chain Freeze and Closed-Form Re-Route: A Discipline for LLM-Collaborative Engineering with Cryptographic Provenance

Large language models have lowered the cost of producing software so far that a single operator with no prior programming experience can generate a codebase that would have taken a small team months. The interesting problem is not the output but the rate at which output exceeds the operator's ability to verify it. This paper identifies three failure modes that emerge predictably under that condition — architectural drift, claim inflation, and loss of provenance — and presents a discipline of four practices that mitigate them: closed-form re-route at intake, phase-chain freeze (a SHA-256-inheritance pattern that detects silent drift at every phase boundary), a six-state proof-state ledger that enforces honesty about what is claimed versus demonstrated, and a seven-phase multi-modal brief assimilation pipeline. The paper documents an open-source reference implementation (lindsey-provenance) and a phononic-bandgap case study whose physical measurement is committed in advance as the discipline's falsifiability instrument (measurement pending). The claim is not that LLM-collaborative engineering is novel, but that without discipline of approximately this shape it generates the appearance of progress while losing the substance.