Which characters live between topics?

One of the 6.11 openers asks: find the pages with unusually mixed LDA topic proportions — do those characters really bridge themes, or is the model confused? We asked, we drew one figure, and we dragged the actual sentences in underneath as evidence.

Representation: the week-6 names-removed LDA fit (8 topics, seed 0) — in that fit every page is a mixture of themes: mutant life, symbiotes & hosts, media boilerplate, and friends. A page that is 98% one theme is an easy label. A page at 47%/28% is the interesting one. We measured each page's mixture entropy (how evenly its topic shares are spread) and sorted all 303 pages by it.

One row per character: the 8 topic shares as a stacked bar (topic colors below the panel), sorted most-ambiguous first; the number at the right is the entropy, of a maximum of 3.00. Pure pages are a single solid block. The colored dot is the week-4 community — watch what the straddlers' dots do. Click a row for the page's top topics and the actual sentences that carry its minority topic.

• Almost half the corpus refuses a single label. 146 of the 303 pages put at least a fifth of their mass on a second topic. The purest pages are the satisfying ones — Eddie Brock, Morbius, Spider-Man, Venom, Black Cat — franchises the fit recognises instantly.
• The straddlers sit at the borders of the graph, too. Of the 20 most ambiguous pages, 14 have at least 5 of their top-10 TF-IDF neighbors outside their own Louvain community. Ricochet, the Crippler, the Caretaker and friends are textually ambiguous and structurally on the fence — two independent lenses agreeing on who belongs nowhere cleanly.
• But the model has a bias worth knowing. In 12 of those top 20, one of the two top topics is media boilerplate — the voiced, playable, film, list mixture that popular pages end with. Some crossings are real (Ghost Rider — screen-and-sword mysticism, check the drawer); some are just Wikipedia's infoboxes in topic clothing.

One frozen fit. Change the seed and the top-20 straddler list reshuffles; the boilerplate topic's grip on popular pages is the fit's signature, not the corpus's law. The ranking tells you where to look — the drawer's sentences are as far as it can take you, and the verdict stays human.