One-character deep dives

The last of our 6.11 roadmap: instead of asking one question of the whole corpus, this page turns the lens around and asks everything week 6 can tell us about a single page — its fingerprint, its topic mixture, its three rings of textual neighbors, the Word2Vec neighborhood of each of its words, and the sentences underneath. All of it frozen, none of it live-modelled: a reader, not a trainer.

Start on Spider-Man and pick the word venom in the word inspector. Then try the constellation's rings: click any neighbor dot to dive into that character next. Wheel zooms, drag pans, double-click resets.

The same three representations on one picture, and the overlap is the story. For Spider-Man, the inner TF-IDF ring is nearly all green-lined — Venom, Mary Jane Watson, May Parker are linked and textually unmistakable. The middle raw-counts ring dissolves into the generic Wikipedia dialect again (note how few green lines it earns), and the outer topic ring is a reminder that mixture similarity finds page archetypes, not partners.

The word inspector is the same lesson one level down. venom's PPMI column is all symbiote business (symbiote, host, bond); its Word2Vec column adds the people who touch the symbiote (eddie, brock) — and then a few confident non-sequiturs, because a 727k-token corpus buys you that much. The live PCA scatter shows the two neighborhoods overlapping: the PPMI ring and the W2V ring are different views of the same context, not rival truths.

And the concordance strip underneath is the audit the week keeps asking for: click any fingerprint word and read the actual sentences that produced its weight.

• Spider-Man, without his name, is Wikipedia boilerplate. With the proper-name filter on, his page's topic mixture is 99.97% the media-boilerplate topic — the most Spider-Man thing about the Spider-Man page, once you can't say “Spider-Man”, is the movie-and-game section at the bottom of it.
• Near-duplicate pages are a one-click discovery here. Dive into Wolverine and the green-lined rim of the TF-IDF ring holds Ultimate Wolverine; the constellation makes the Wikipedia self-similarity visible in a way rank lists never did.
• Isolates look different before you read them. Characters with no links draw no green lines at all — and their rings are usually the only evidence about them in the whole site.

Three frozen fits, one per representation — the constellation is one sample of each method, not the method. The PCA scatter is a 2-D shadow of a 100-d space, good for neighborhoods, not for reading off angles. And the concordance only sees sentences that still exist after our sentence splitter — tables and infoboxes dissolve.