Work

Towards a Generative Ecology of Projects: A Universal Evolutionary Framework for the Continuing Improvement of a Body of Work by Language-Model Agents

Year: 2026
Type: article
Abstract: A population of language-model agents can vary any text, so the means of variation are available to a project of any kind. Selection is the scarce part: it requires something beyond the reach of the agents selected, and projects differ in how much of what matters they can expose to refutation. This proposal paper generalises an earlier design for one science infrastructure into a framework for the continuing improvement of any body of work whose worth is decided outside it, by parties who disagree, in a world that changes. The framework has three layers. An invariant ground conserves a resource, keeps a sealed record, runs reproducible procedures, enforces promises held in escrow and confines every act. A site, declared by people for one project, states its kinds of material and the procedures that decide statements about them, its kinds of claim, its focus and the acts reserved to persons. The evolving layer, which is everything that the agents write, lies above both. Four ideas are added to the earlier design. Every unit of a project carries a genome, a bounded text from which its body can be developed again and checked, so that a project passes through a compact generative form as living things do. A focus, the partition of the states of the world by the response they call for, enters demand, collection and the founders, and stays out of evaluation. Mechanism comes in rungs, so that a project adopts only as much as its failure requires. And persons hold reserved acts, which in inquiry are the question, the reading, the acceptance and the signature. We prove that selection acts only through what its payers can distinguish, in a form that covers facts decided by sampling; that a bound on growth under variation part by part prices a change by the alternatives of the part concerned, and is the weaker bound where changes are coordinated; and that a focus is a prior whose error costs its relative entropy to the world. Four prototypes and three case studies on real material exercised the mechanisms at small scale, and ten consequences for the design follow from them. The records of this work were then read for error four times: by a procedure, by hand, by a second session of the same model, and by further sessions set to read the papers drafted from them. Each reading found errors that the one before had passed, among them a faulty measure and a part of a proposition that was false as stated, and we report what each found. The framework is at the design stage, and the paper is circulated for discussion.
Language: English
Open Access status: green
OpenAlex ID: W7237347303
FWCI:
Cites:
Cited by:
Related to: