Yeah, you can see the AI code output for the gen_profanity_list:

#
***
provenance (keep in sync with third_party/profanity-list/PROVENANCE.md)
***
SOURCE_REPO = "dsojevic/profanity-list"
SOURCE_URL = "https://github.com/dsojevic/profanity-list"
SOURCE_COMMIT = "c27924319aa9bd6f917e3782b4f4b6604a50b652"
RETRIEVED = "2026-07-18"

#
***
curation dials (edit + regenerate to retune)
***
SEVERITY_FLOOR = 2  # drop severity 1 (mild/clinical, e.g. "anus")
EXCLUDE_TAGS = {"lgbtq"}  # identity terms -> handled contextually by the ML identity_attack pass
MIN_LEN = 3  # never ship 1-2 char tokens (whole-token over-block risk)

# Terms with heavy innocent usage that would over-block a public room as whole-token matches
# ("the opposite sex", "nude" the colour, "collared shirt", "white-collar"). Genuinely sexual usage is
# still caught contextually by the ML sexual_explicit pass. Edit + regenerate to adjust.
CURATED_DROP = {"sex", "nude", "nudity", "topless", "collared", "collaring"}

# The original hand-curated starter terms, union'd in so a regeneration never regresses the set the
# project shipped before adopting the corpus. All are already single [a-z]+ tokens.
ALWAYS_INCLUDE = [
    "fuck", "fucker", "motherfucker", "shit", "bullshit", "asshole", "bitch",
    "bastard", "dickhead", "cunt", "slut", "whore", "piss", "prick", "wanker", "twat",
]