Main idea

In desk research carried out in loops — preparing for a conference, a literature review, an investigation file — the next search doesn't have to come out of a plan fixed at the start. It can be derived from the state of the corpus built so far: what seems established, what rests on a single source, what remains uncertain, the contradictions, the ambiguous terms, the poorly understood mechanisms, the over-represented areas and those that are almost absent.

The usual logic collects first and understands later. This one reverses the order: you understand enough to decide what to look for. The corpus stops being only the result of the research; it becomes its engine.

The question that guides each new search changes accordingly. It is no longer "what are the best articles on this subject?" but "given what we already know, what information would be most likely to change our understanding?".

Why it matters

It gives every new source a reason to exist: it is sought because it answers an identified gap, not because it is relevant in general.

And it brings out the blind spots that bulk collecting leaves in the dark: a gap becomes a search target instead of remaining a hole nobody sees.

Nuances and limits

The diagnosis only sees what the corpus lets you suspect. A dimension no source has ever left a trace of appears in no state, and the loop can run for a long time without coming across it.

And the method assumes a layer that describes the state of the corpus — uncertainties, contradictions, gaps. Without it, there is nothing to diagnose, only a pile of documents.

Open questions

• Who produces the diagnosis when the loop is handed to an agent: the same model that read the sources, at the risk of carrying over its own biases, or a second pair of eyes?