Protocols and papers
From Model Output to Accepted State
The working paper that proposes the State Transition Protocol as a typed boundary between model proposals and accepted state.
Working paper ยท owner-review draft v4
- Printed subtitle
- A typed state boundary for AI-assisted operations
- Format
- PDF, 36 pages
- Date
- 15 August 2026
- License
- Intended for release under CC BY 4.0, as stated in the draft
What it means
A language model's output is a proposal. It is not an authorization, a completed action, an observation, or an accepted record of state.
The paper proposes keeping those records separate through a conceptual ten-stage lifecycle in which only a valid acceptance record advances governed state. It reports a reference implementation, contracts, reducers, and test harnesses that keep the records separate within the evaluated boundary, and that, used on one public software project, found real defects, including defects in the system itself.
A proposal is not an authorization, a completed action, an observation, or an accepted record of state.
Page 1
This is cross-language replay parity, not independent reproduction.
Page 1
How it functions
Each result-bearing claim carries one of four markers. TESTED means exact behavior over a named finite corpus, reproducible by a stated command. OBSERVED means a bounded inspection at a recorded time. PROPOSED means specified beyond the tested boundary, not a result. OPEN marks where the evidence stops.
Public case-study claims are bound to commit 275d0b3e7474 of this repository, and Appendix C lists full digests, exact locators, and the verification boundary for each artifact: public rows can be rehashed at that commit, while retained-source rows are not public inputs and cannot be checked from the repository. Section 11 states in one place what the work does not establish. The PDF here is the 15 August build of draft v4, 36 pages. A later 37-page build of the same draft, with an added related-work section, is under review in pull request 15.
Status and claim boundary
Working paper, owner-review draft v4, dated 15 August 2026. The draft states that it has no completed publication commitment. The ten-stage lifecycle is conceptual and PROPOSED; the public packet implements a narrower seven-event profile over synthetic fixtures and is not a complete implementation of that lifecycle.
Cross-language replay parity between JavaScript and Python ports is not independent reproduction. The 34 runner-reported numbered holds across twelve suites are a diagnostic inventory, not a coverage measure. The paper does not establish causation, lawful authority, regulatory compliance, statistical reliability, general safety, or independent validation.