Originally written in French. Translated by AI — the meaning has been preserved, not the prose.
Angle
Desk research done with an AI is better steered by the state of what it knows than by the number of sources it gathers: that state — what is established, fragile, contradictory or absent within the mission scope — points to the next source to look for, and it is also what says when to stop, once searches designed to topple it fail to change it.
Synthesis
The ideas assembled here cover the whole life of a piece of research — its framing, its start, its unit of measurement, its iterations, its end — and they converge on a single shift: the quantity in charge is no longer the volume collected but the representation built.
That shift has a consequence no single note shows: deciding what comes next and deciding when to stop become the same operation. The diagnosis that says which type of evidence is missing also says, when it finds nothing left to ask for and the break attempts come back empty, that the mission is done. Volume-driven research needs two rules — how much to search, when to stop — that don't talk to each other; state-driven research needs only one.
This is only possible under two conditions. You need a scope, otherwise the state has nothing to refer to and convergence has no meaning. And you need a layer of extracted knowledge, otherwise there is no state to read, only documents. These two conditions, not the power of the tool, are what separate research that knows where it stands from research that is merely deep.
Tensions / contradictions
The method relies on the diagnosis to see the gaps, but a diagnosis only sees what the corpus lets you suspect. An angle no source has left a trace of escapes the state just as it escaped volume. That is the limit that prior-art searches conducted outside your own vocabulary try to lift: saturation within a frame says nothing about what lies outside it.
And convergence is still established by whoever built the model. The break searches they design target the threats they can imagine.
Questions
- How do you bring into the loop a source the diagnosis wouldn't have asked for, without falling back into volume-driven collecting?