The analogy slot machine

The week's most famous demonstration is word arithmetic: king − man + woman ≈ queen. Does the same trick work one level up — start from a character, subtract a word, add a word, and let the machine name the closest superhero? This is the week's play round: all the TF-IDF, PPMI and topic-mixing the pages built gets compressed into one toy with a lever. Pull the lever:

Same recipe as the Marvel map: a character's position is the TF-IDF-weighted average of their fingerprint words (names removed) in our Marvel-trained Word2Vec space. A fingerprint is just the handful of highest-TF-IDF words on a single page, and names are dropped so the machine can't shortcut through people's titles. Trade a word and the traded vectors have to come from the 3,791 words we shipped — anything outside the vocabulary would silently poison the arithmetic, so those inputs get bounced with an error instead. The space itself was trained on the full week's corpus, 727,440 tokens across all 303 pages, which is why words like magic or hammer carry real Marvel meaning inside it.

The new direction is scored against all 303 character positions by cosine — read it as a similarity dial, where 1 means "same direction" and 0 means "unrelated"; it is not a percentage. The podium is who's closest. If you already know the hero you expected, pick them too — the machine reports their exact rank, cosine, and who got in the way. Nothing is precomputed; every spin is live arithmetic.

Every spin reads the same way. The podium is the machine's top five, with its best guess highlighted. Beneath it, the moved-words line shows which of the starting fingerprint words pulled toward each pole — the trade's geometry in one sentence. If you named an expected hero, the verdict card reports their exact rank out of 302, their cosine, and up to three pages that got in the way; that ranking deliberately includes near-duplicate pages, so you can see the corpus's own copy problem in the numbers. Click any podium row to recenter the whole equation on that hero, and audit anything suspicious with the sentence drawer underneath. The map above the podium redraws once per spin: its x-axis is the traded direction, so the podium's dots herd toward the arrowhead (in roughly rank order — cosines and projections disagree a little), y is the widest spread left over. It is a projection, not a ruler — the podium above, computed in the full 100-D space, is the ground truth.

• The obvious answer, measured. Run the default Spider − man + woman round and set the expected hero to Spider-Woman: she lands at #16 of 302 (cosine 0.49) — Wild Child, Spider-Girl and Father Time got there first, and the blockers list puts the machine's excuse right in front of you.
• Family is not a ranking. Spin Hulk − man + woman and expect She-Hulk — the canonical answer in family terms — yet she comes in at #42, with Wild Child, Northstar and Jean Grey standing in the way. The traded pole is one of a dozen fingerprint words; family resemblance in this space has to be earned.
• Then build your own round. Start from your favorite character, trade a word for its conceptual opposite, and use the sentence drawer to decide whether the machine was strange-right or strange-wrong. The preset chips under the lever are safe starts; every other combination is yours.

• The biology round works. The default preset, Spider − man + woman, puts Spider-Girl right at the top of the family (with Wild Child, the machine's own punchline, at #1) — and the moved-words row shows why: man-carrying fingerprints pushed out, woman-adjacent ones pulled in.
• Some rounds land in the right family. Luke Cage − street + magic crowns Doctor Druid — not the Doctor you ordered, but a real sorcerer from the same neighborhood of the space, with Doctor Strange himself close behind. The analogies rarely name the exact hero; they reliably name the right kind of hero.
• And some rounds are honest failures. The echo trap: Venom − aliens + technology returns Eddie Brock — the direction barely escapes the symbiote and re-lands on its host, because Venom's fingerprint is about being a symbiote wearing a man. Storm − weather + mind drifts to Satana, a mystic, instead of a telepath. The evidence drawer under each spin is there exactly for rounds like these.

A character is reduced to a dozen weighted words, the space to 100 dimensions learned from 727k tokens, and every round inherits the corpus's biases — the war↔home lean that the map post caught. Treat the machine as a comedy writer with a statistics degree: sometimes strangely right, always worth auditing.