We have spent four parts on a single, uncomfortable idea: that a literature is not a body of settled knowledge but a network of arguments, most of them unread, some of them unfounded, all of them measured by a proxy — citation count — that stopped tracking quality the moment it became a target. If you have followed this far, the natural response is defensive: how do I avoid being fooled? That is the wrong question, or at least the small one. The large question is generative. Where an argument stops is where your next experiment starts.
A gap is a claim nobody has tested yet
When you decompose a paper into premises and conclusions and rate each claim honestly, the interesting output is not the strong claims — it is the weak joints. The premise asserted with confidence but supported by nothing. The result shown under one condition and cited as if it were universal. The citation chain that bottomed out in a review citing a review. Each of these is a place the field is standing on ground it never poured. And each can be rewritten, with almost no effort, from a complaint into a hypothesis: not "they never tested whether X holds under condition Y," but "X does not hold under condition Y" — a claim with a truth value, a required method, and a sample size.
That rewrite is the whole discipline. A gap phrased as a grievance is a dead end. A gap phrased as a falsifiable claim is a pre-registration waiting to happen.
The literature's weakest joints are not warnings to step around. They are the coordinates of every experiment that has not been run yet. A field's gaps are its to-do list, written in the negative.
Why the falsifiable framing is the one that pays
This is not a stylistic preference; it is the lesson the reproducibility crisis taught at enormous cost. When Brian Nosek and colleagues laid out the case for pre-registration in PNAS, the core argument was that the line between generating a hypothesis and testing it has to be drawn before the data are seen, or the test quietly becomes a story fitted to noise. The Open Science Collaboration's mass replication in Science found that a minority of high-profile psychology results reproduced at the original effect size. Monya Baker's Nature survey of 1,576 researchers found more than 70% had failed to reproduce another scientist's experiment, and more than half had failed to reproduce their own. Ioannidis had predicted the shape of all of this a decade earlier, from first principles, in the paper whose title still stops people cold: most published findings may be false.
The through-line is that a claim is only worth building on if it was framed so it could have failed. So the discipline for mining gaps has to inherit that same rigour. It is not enough to notice that a premise is unsupported. You state the gap as a specific claim, specify the method that would decide it, specify the statistical test and the sample size that would give the test power, and commit to that specification before you run it. The gap, treated this way, is not a vague "future work" sentence — it is a protocol.
The mindset shift, concretely
What this does to how you read is worth naming, because it inverts the anxiety. Reading defensively, you scan for reasons a paper might be wrong so it can't embarrass you. Reading generatively, you scan for exactly the same weaknesses — but each one is now an opportunity with your name on it. The badly-supported premise is not a landmine; it is an open question you are unusually well-placed to close. The field's citation chains that dead-end in nothing are not evidence of decay; they are unclaimed territory. You stop being a consumer of the literature defending yourself against its errors and become a producer mining its edges for the next result.
Practically, the loop is:
- Walk the field from its founding paper, so you see how the arguments were built rather than inheriting the current summary of them.
- Decompose the key papers into rated claims, so the weak joints are visible instead of averaged away.
- Audit the citations — grounded, dead-ended, or circular — so you know which "established" facts are actually established.
- Read each paper's intrinsic quality without the citation count, so popularity stops standing in for soundness.
- Convert every real gap into a falsifiable hypothesis with a method and a power calculation — and treat that as the pre-registration it already is.
ReSach, and where to watch it run
This is the entire premise of ReSach. It walks a field from its founding paper, decomposes the arguments into individually rated claims, and audits every citation with OnCite — labelling each one grounded, black-hole (the chain ends in nothing), or circular. It scores each paper's intrinsic quality with a citation-free Paper Weight, so soundness is read from the work rather than borrowed from the crowd. And it surfaces the field's gaps not as prose but as testable hypotheses, each with a suggested methodology and the statistics that would give it power.
You do not have to take any of that on the strength of an argument — which would be a poor way to end a series about not taking claims on faith. There is a live demonstration that runs the whole loop on a field you know: the origins of AI, walked from its founding papers, its arguments decomposed, its citations audited, its gaps written as hypotheses you could pre-register tomorrow. Open it, follow one citation chain to ground or to nothing, and read one gap as the experiment it already is. That is the fastest way to find out whether reading a literature for its arguments, rather than its popularity, changes what you would do next.