What should a research scaffold decide?

A new software project rarely needs to invent its directory structure. A scaffold supplies a working starting point: files in predictable places, commands that run and conventions the next contributor can learn.

I wanted that convenience for academic research. It led me to create-academic-research, a scaffold for sources, analysis and manuscripts, with instructions an AI agent can follow.

The analogy has a limit. A research project may discover that its original question was badly posed, that the evidence is insufficient or that a promising approach does not work. A scaffold that makes those outcomes awkward could organize the project beautifully while making the research worse.

Give uncertainty somewhere to go

An empty section labelled “Research gap” looks like a request for text. Add an agent instructed to complete every section and the template starts exerting pressure on the conclusion. The sources may not support a new contribution, but the workflow still asks for a contribution-shaped paragraph.

A literature review ought to be able to end with a narrower question, a contradiction or a reason not to proceed. An experiment that fails ought to remain findable after the manuscript takes another direction. The workspace should make these ordinary records, rather than exceptions someone must remember to explain outside the process.

This changes what it means to finish a task. “Complete the survey” could mean recording which sources were examined, what they support and what remains unresolved. It need not mean finding a novel claim before the agent is allowed to stop.

The problem is more subtle than fabricated citations. A document can cite real papers and still give a false impression of agreement by removing everything that did not fit its outline.

Standardize what helps someone check

A source can have a stable record containing the paper, its metadata and notes. An analysis can keep its inputs, code, report and reproduction instructions together. Manuscripts can share a bibliography rather than gradually developing incompatible copies.

These conventions are useful because they reduce the effort of returning to a question. If a reference is wrong, where was it used? If an analysis changes, which result needs to be regenerated? A folder structure cannot answer every such question automatically, but it can make the necessary material easier to find.

Cookiecutter Data Science provides structure for data-science projects, while Quarto projects organize related computational documents. My particular interest is the relationship between a researcher and an agent: what each can find, which records they can change, and how a written claim remains connected to its source.

A scholarly lookup can recover a paper’s identity and metadata. Reading the paper is still necessary to decide whether it supports a sentence. A summary can help with that reading, but later work needs a route back to the relevant passage. Otherwise an interpretation can become an authority simply through being easier to retrieve than the original.

This is also a limit on what structural checks can certify. A valid bibliography and a complete set of folders show that certain conventions were followed. They cannot settle whether the question was worthwhile or the inference sound.

I want the scaffold to make the clerical work less demanding and the reasoning more open to inspection. A useful template would leave a project easier to continue, even when the next researcher decides its main conclusion is wrong.