Lindsey Provenance Discipline

Open-source · Python · MIT licensed

An LLM will happily write the code and the documentation that says the code works. This is the discipline that keeps those two honest.

Install

PyPIlindsey-provenance

Dependencies

Minimalstdlib + numpy

License

MITuse it freely

What it is

The rule is simple: a claim can't outrun its evidence. Easy to say, hard to hold when an LLM will write the code and, in the same breath, write the documentation describing code that doesn't quite exist yet. This is the set of practices that holds the line — for one person working at a pace that used to take a team.

It doesn't limit what you build. It gives the gap between what an artifact actually is and what you wish it were a place to show up early — early enough to close by doing the missing work, instead of absorbing it into a claim.

The reference implementation is a Python package. A companion preprint is published, open access (the arXiv version of record is in progress). The same discipline runs underneath my trade-side and industrial products — AENORIS, GIZUIZ, and BYGYZE.

Install

pip install lindsey-provenance

Standard library + numpy only. No deep-ML dependencies. Python 3.10+.

The four practices

Tap a practice to see how it works.

Phase-chain freeze+

Each phase produces a SHA-256 manifest that inherits the previous phase's hash. The chain runs unbroken from the sealed baseline forward. If a file drifts silently, the next freeze fails loudly instead of letting the drift through.

Six-state proof-state ledger+

Every artifact sits at exactly one state — idea → planned → implemented → simulated → artifact-generated → physically-validated — and the machine is monotonic. You can't describe something in artifact-generated as physically validated. The words have to match the state.

Closed-form re-route at intake+

When an exchange introduces an algebraic claim, it's checked against numerical truth before anything is committed — and rerouted to a stdlib + numpy substitute when a heavy dependency isn't earning its place. Default gate: Pearson r ≥ 0.95, RMSE ≤ 10⁻⁶.

Multi-modal brief assimilation+

Briefs arrive as .docx, .eml, whiteboard photos, handwritten notes. A seven-phase pipeline ingests, classifies, and binds them to the project's evidence surface before any code is written — so intent doesn't get lost in a long thread.

Resources

→ github.com/bradmlindsey/lindsey-provenance — source, license, docs
→ Verification procedure — validate the Ed25519 retro-signed ledger yourself
→ Preprint (methodology) — Phase-Chain Freeze and Closed-Form Re-Route. Zenodo DOI · arXiv version of record forthcoming
→ Companion papers — One Operator, Nine Trunks, Seven Weeks · Zero of Forty-Nine
→ Writing — notes on the discipline, each pointing at a real artifact

Author

Built by Brad M. Lindsey — Master Electrician (DoD/DHA), Master HVAC Tech, PMP, independent engineer. He wrote his first line of code on April 4, 2026; this discipline is the part he built to keep AI-collaborative work from drifting out from under its own claims. ORCID 0009-0004-6392-2720.