What does it actually mean to do research well? We're easily impressed by an extraordinary result - or, increasingly, by the speed at which a new class of “automated researchers” can produce one. But the quality of an output tells us surprisingly little about whether the research itself was well-designed, much less whether it was conducted optimally. Did it pursue the right question? Did it inherit a method that distorted what could be observed? Would another approach have produced a different answer? How would we even know?
Choosing what to investigate is only the first major decision. The second is deciding how to investigate it. Even when a method is applied rigorously, the choice of method itself is rarely treated as something to explore. We reach for the familiar dataset, analytical framework, experimental design, statistical test, or representational structure and proceed as if these were neutral instruments rather than decisions that determine what becomes visible.
But every method offers only a partial view. A singular result can never represent reality unto itself. It is instead an observation mediated through the instruments used to produce it. If several genuinely different methods converge, we have stronger grounds for believing that the result belongs to the phenomenon rather than to any one method. If they disagree, that disagreement is equally valuable: it may reveal a hidden assumption, a measurement artifact, a boundary condition, or a different dimension of the subject altogether. Multi-method research allows us to triangulate a better approximation of the truth.
CoResearcher is my attempt to build a flexible human+AI platform for empirical research that treats both the research direction and the methodology as spaces of possibility, helping a researcher define, refine, pursue, and continually reconsider not only what to investigate, but how best to investigate it.
Empirical research does not, however, begin from nothing. Before designing a study, we need to understand what has already been attempted, which findings survived replication, where explanations conflict, which methods produced which results, and whether another field has already solved an analogous problem under different language. This is where CoResearcher connects directly to the source-research capabilities within Caro and to the graph-construction capabilities of OKGC:
A typed source graph can preserve more than the fact that one work cites another. It can represent whether a source builds upon a finding, disputes it, fails to replicate it, adopts its method, or reaches a conflicting result. Structural search can reveal relevant work whose vocabulary differs, trace the fate of a finding through later research, and help assess novelty by a project's position within the larger field rather than by textual similarity alone. Frequency, however, is not evidence of fitness. CoResearcher can test these recommendations empirically and return the results to the graph, allowing future recommendations to reflect where a method actually worked rather than merely where it was customary.
A connected methods graph can make the methodological record equally explicit by treating methods themselves as first-class objects. CoResearcher can then search and recombine that archive when constructing a study: not merely recalling familiar methods, but identifying approaches that addressed structurally similar problems even when their terminology and disciplinary origins differ. Said differently - effectively all of our research methods are siloed and habituated to a given discipline. But reality does not obey such human-based barriers; there's simply no reason why the standard, vanilla approaches in physics or mathematics might not apply well to social scientific research.
A research program within CoResearcher would therefore not be forced through a single predetermined pipeline. Working with the researcher(s), the system grounds the inquiry in the existing research, identifies applicable tools, constructs alternative study designs, executes quantitative and qualitative empirical work, compares the resulting observations, and determines whether to refine the current path, pursue an unexpected finding, test another explanation, or move elsewhere entirely.
This does not mean that CoResearcher can certify that a research program is globally optimal. “Optimal” depends upon the purpose of the work; on explanatory power, causal identification, novelty, practical value, cost, urgency, or some combination of them. But it can replace an unexamined default with explicit alternatives. It can show which findings survive changes in method, which conclusions depend upon a particular analytical choice, and which further investigation would most meaningfully reduce uncertainty. The design of the research becomes something that can itself be researched.
Its output is therefore not a finished essay or academic paper. It is a durable and inspectable body of empirical research:
This body of empirical work can then enter Caro alongside the surrounding source corpus. Caro takes up the different work of determining what those materials mean together, which arguments they support, and how they should become prose - with gaps discovered during writing able to return to CoResearcher as new empirical questions.