Anime Semantle

Three Questions We Measured and Couldn't Answer

← Back to the game

Every other analysis article on this site ends with "we measured it, and here's the conclusion." That's partly selection bias — we only wrote up the questions that produced a usable answer. This one goes the other way: three things we actually measured that did not produce an answer worth turning into a tip. One was a bug, one showed no real difference, and one we still can't explain.

1. Does the same franchise ever score low?

Question — sequels and spin-offs should obviously score high in similarity to each other. We wondered if there were exceptions.

Measured — pulled every same-franchise pair scoring below 0.60 in similarity. Seven turned up.

All seven were false positives from the franchise-matching logic, not a real content finding.

Two separate causes:

The real answer to the original question turned out to be the unremarkable "same franchise is always close, no exceptions" — and the actual output of this investigation was a bug fix, not content.

2. Does media type affect difficulty?

Question — the intuition that a compressed feature-length plot makes movies harder to place than TV series.

Measured — nearest-neighbor similarity (isolation) by media type: TV, Movie, OVA, ONA, Special.

Media typeMedian isolation
Movie0.738
TV0.731

Measured 2026-08-05, against the 800-title / 3,110-entry pool at the time.

No real differentiation. The "movies are harder" intuition doesn't hold up in this data.

A 0.007 gap is well within the range explainable by other factors in the sample, not media type. Not a difference worth turning into a gameplay tip, so it didn't become one.

3. Do lead characters score closer to their own show than supporting ones?

Question — a lead character carries the plot directly, so typing a lead should be a stronger clue than typing a supporting character.

Measured — similarity between each character and its own show, split by role.

RoleMedian similarity to own show
Supporting0.921
Main0.861

Measured 2026-08-05, against the 800-title / 3,110-entry pool at the time.

The opposite of what we expected. We went looking for a cause.

In how character records are built, every character belonging to the same show shares identical title, series name, tags, and synopsis text — only the name differs. So this gap may be an artifact of the name string itself, not an actual difference in character content.

But simple checks — name length, punctuation ratio — don't explain the gap either. The actual cause is unconfirmed. Publishing "type a supporting character" as a tip without knowing the cause would mean selling a possible pipeline artifact as fact. So this result did not become a gameplay tip.

4. Why this article exists

Every other analysis piece here also started from measurement, but only the ones that produced a clean "do this, it helps" conclusion got written up. Publishing only those risks making it look like measuring anything on this site always yields a tidy answer. In reality, the same process produced at least three dead ends — one of them was literally a bug, and one we still can't fully explain. Leaving that visible, instead of quietly dropping it, is the point: not every number is content just because it was measured.

Further reading

The findings that did become articles are in Ratings Are Not a Clue, When to Guess a Character, and Which Genres Are Hardest. The main strategy guide has the similarity band reference table and a full walkthrough, and each day's answer accumulates in the past answers archive.