Receipts Over Genius: An Audited Account of an 18-Month, AI-Assisted, Solo Invention Cycle

Abstract

SQUIRL OS RESEARCH DISCLOSURE - READ FIRST: This work is part of an open-book, AI-assisted research program produced under bifurcated governance (a probabilistic AI agent drafts; a deterministic human key approves; a boundary-layer audit checks claims against receipt archives). Audit-and-correction is doctrine: superseded originals are preserved in version histories, corrections are disclosed amendments, and every significant test carries a preregistered falsifier in the append-only public ledger (github.com/LLong2026/four-horsemen). This software and its accompanying research are prototypes provided for educational and research purposes only; they are not intended for production use, commercial deployment, or safety-critical environments. In April 2025 the author, a working technologist with no background in AI, set himself a falsifiable experiment: use AI assistance to travel the complete invention cycle - conceive, develop the mathematics, build the working system, file the patents, publish the science, demonstrate the result - with every significant step verified by a timestamped, publicly inspectable receipt. This paper is the closing account of that experiment, written under the same rules it describes. Result, scoped honestly: in 18 months a single operator completed the cycle - eight U.S. provisional patent applications pending; a working self-healing infrastructure platform across a live app fleet; the first documented caging of an LLM inside a deterministic mathematical boundary (DOI 10.5281/zenodo.23265723), a receipt that survived direct third-party scrutiny; and a public research line (the DCAI community) walking from theory to governance proof to field event. Two findings carry the paper: (1) the experiment succeeded because of the boundary, not despite it - bifurcated governance turned the operator's lack of background into an auditable process; (2) the program generated its own counter-evidence, reported here rather than discarded: five cataloged instances of AI confabulation inside the program itself (fabricated statuses, fabricated evaluation data, a fabricated institutional identity), every one caught by the same receipt discipline the experiment was testing. The method worked on its author's own tooling. Thesis community: Deterministic Caged AI (DCAI) - https://zenodo.org/communities/dcai. Lineage acknowledged: Nyquist (1924), Hartley (1928), Shannon (1948), Kolmogorov (1965), Popper (1959). Companion paper: Deterministic Caging of Probabilistic Large Language Models (DOI 10.5281/zenodo.23266727).