Before a writer transports thoughts to page, he must first construct an understanding of a corpus: determine which sources are actually useful, which claims genuinely differ, where evidence converges or conflicts, which causal explanations remain plausible, what is still missing, and which structure might make all of this intelligible to a reader. As the corpus grows, that model quickly exceeds what any person can reliably hold in working memory. And thus we writers and thinkers require tools that actually acknowledge these realities rather than forcing us to mutate our actual optimal behavior around the tools.
Most writing software begins too late. It assumes the writer has already done the difficult epistemic work and now needs somewhere to arrange sentences. It preserves that assumption while offering to produce those sentences faster. Research assistants often stop too early: they retrieve sources, answer questions, or generate a report, but the understanding assembled during that work remains transient - scattered across documents and conversations or hidden inside model context. The next question, draft, or writing session begins by reconstructing much of it again.
Caro is my attempt to create a full-stack platform that understands and enhances the research + writing process; a co-writing environment in which humans and AI can construct, inspect, challenge, and continually refine that understanding together - and then transform it into beautiful prose.
Caro treats the evolving model of a corpus - not the prompt or even the draft - as the central object of the writing process. Sources, annotations, entities, events, claims, evidence, arguments, open questions, authorial positions, and possible narrative structures become durable, connected objects that can be revisited and revised. This is where Caro connects directly to OKGC: no single graph can make a corpus fully intelligible.
Together, these representations allow both human and machine to work over an explicit model of the material rather than repeatedly reducing it to summaries and prompts.
Writing from that model is not a final, linear stage. Drafting reveals gaps that research did not expose, weak arguments that an outline made plausible, and forms that could not have been discovered in advance. A failed draft may therefore be a productive attempt: it teaches the writer what the piece is not and changes the conditions of the next attempt. Caro preserves that history and allows the process to move recursively among research, interpretation, structure, composition, critique, and revision. Multiple arguments or narrative structures can be explored without pretending that the first coherent version is necessarily the best one.
The human+AI relationship is equally important. Caro rejects the idea that the ideal writing system is an autonomous model that accepts instructions and returns finished prose. Its better analogy is pair programming - a continuous exchange in which each participant proposes, inspects, questions, corrects, and takes the lead when their capabilities are most useful. The human supplies purpose, judgment, taste, experience, and responsibility. Caro's durable substrate preserves continuity across a large body of material, surfaces connections, proposes alternatives, and allows a developing piece to be inspected through specialized lenses such as structure, evidence, logic, and voice. Caro supports both live co-writing, in which human and AI alternate at a small cadence, and slower asynchronous work, while making the actions and contributions of every participant visible.
The project takes its name not from an ambition to imitate Robert Caro’s prose, but from the research ideal his work represents: exhaustive engagement with evidence, attention to power and causality, and the transformation of immense source material into writing that changes how its subject can be understood. Caro asks how much further that practice might extend when human judgment is paired with computational memory, explicit representations, and systematic exploration. The goal is not a machine that writes on our behalf, but a system that helps us understand more than we could hold alone - and write from that understanding with greater originality, rigor, and consequence.