The Lindsey Provenance Discipline now has a reference paper: Phase-Chain Freeze and Closed-Form Re-Route: A Discipline for LLM-Collaborative Engineering with Cryptographic Provenance.
The problem it addresses is not that LLMs write bad code. It's the rate — output exceeds the operator's ability to verify it, and three failure modes follow predictably: architectural drift, claim inflation, and loss of provenance. The paper names four practices that hold the line against them:
- Closed-form re-route at intake — resolve ambiguity before it propagates downstream.
- Phase-chain freeze — a SHA-256-inheritance pattern that turns silent drift into a runtime error at every phase boundary.
- A six-state proof-state ledger — every claim carries exactly one state from idea to physically-validated, and the ledger is monotonic, so nothing skips ahead of its evidence.
- Multi-modal brief assimilation — a seven-phase intake so a project starts from an accurately-stated input.
The reference implementation is the open-source lindsey-provenance package (MIT, Python). The discipline's falsifiability instrument is committed in advance: a phononic-bandgap case study whose physical measurement is named before the result is known. Measurement pending.
The paper is open access. arXiv version of record forthcoming.