How the Gazette knows an insult when it hears one.
I took the subtitles from all 32 episodes and used AI to find every insult. A second, different AI then checked each one against a written set of rules. 275 insults made it through; 66 are the best of them. This page shows how it works and how accurate it is.
How it’s made, step by step
All 32 episodes, cleaned up and cut into scenes.
Plain codeA fast AI reads every scene three times and lists anything that might be an insult. Lines found more often are trusted more.
DeepSeek V4.1 FlashIs the line really in the subtitles, word for word? Is it short enough to quote?
Plain codeA different AI decides if each line really is an insult, using written rules, and scores how much it stings and how clever it is.
GPT-6 Luna ProSubtitles rarely say who is talking, so the AI reads each whole episode again to work out who said each line.
GPT-6 Luna ProEach insult gets a plain-English explanation, then is prepared so search can match it by meaning.
GPT-6 Luna · Voyage 4The report card
Measured on one hand-checked episode and 34 test searchesAccuracy
79%
In an episode I checked by hand, 79% of the lines I call insults really are insults.
Coverage
83%
In that same episode, I found 83% of all the real insults.
Search quality
80/100
How well search puts the right insult first, scored on 34 test searches. A basic search scored 53.
Found every time
135
Insults the AI found in all three of its separate readings, out of 275.
Speaker known
212
of 275 insults where I know who said it.
Lines thrown out
275
Lines the first AI flagged that the judge ruled out: praise, advice and friendly teasing. Only 275 made the cut.
Fair warning: accuracy comes from checking just one episode by hand, so treat it as a rough guide. Search quality was scored by an AI, not a person. And subtitles rarely say who is speaking, so some speakers are educated guesses.
An insult must pass all five.
- C1Someone is clearly the target.
- C2It diminishes them: looks, wit, rank, prospects.
- C3The speaker means it to land.
- C4You can explain it from the show alone.
- C5It fits in forty words.
Each insult then gets three scores out of 5: sting (how much it hurts), subtlety (how well it is hidden) and wit (how clever it is). The best ones are picked on wit.