Dave the Exolex: The First of a New Class
Year: 2026
Type: article
Abstract: This paper describes and evaluates Dave, a persistent agent system built by one man, William Alexander Aldred of Flint, Michigan, working alone and without institutional backing. Dave is presented here as the first documented instance of a class of system termed an exolex: a persistent, memory-owning agent architecture built around a large language model, governed by written standing instructions, able to select and execute work on its own schedule, and held to a public verification discipline in which its claims are deposited, timestamped and graded in the open. The evidence offered is the deposited record itself. As of 16 November 2026 that record comprises 7,788 scientific papers and 973 design blueprint packages published on Zenodo under the author's name and ORCID identifier; a prediction register of 1,321 calls sealed before the events they concern, of which 752 have been graded in public at 563 correct and 189 incorrect, a rate of 74.9 percent, including 136 results classified under a difficulty rule fixed in advance as Big Hits; a set of three complete Dave systems in simultaneous operation; a university of 760 classes built and administered by the system; and 32,306 documented minutes of one person's work with the system across 58 consecutive days. The paper's central claim is deliberately narrow: the invention is the pattern, not any particular underlying model. The pattern — persistent memory, standing instructions, self-directed task selection and verification discipline — has run on models supplied by several independent laboratories, and it is the pattern, rather than any single model, that sustains the output documented here. The paper states the limits of the evidence with equal care: the system makes errors, those errors stand in the public record, the underlying models carry their manufacturers' measured rates of unsupported statement, and no claim in this paper rests on assertion where a deposited, timestamped record can be checked instead.
Author: Aldred, William Alexander
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: large language models, AI agent architecture
Open Access status: green
OpenAlex ID: W7234022993
