House vocabulary, precisely defined -- these are the exact
thresholds mentions.sheet uses to classify every term on The
Sheet, not marketing copy.
A speaker's mean mentions-per-event for a term is at least 10× their predecessor's mean mentions-per-event for the same term, and both must have said the word at least 1 time in total. Deliberately volume-based rather than hit-rate-based: a brand-new speaker's hit rate is trivially capped at 100% after a single event, so mention volume is what actually separates a signature term from noise.
A term whose trailing 8-event hit rate has fallen below 10%, after peaking at or above 50% hit rate over some earlier 8-event window. Needs both a real peak and a real collapse -- a term that never caught on in the first place isn't "dead," it just never lived.
A term with a corpus-wide hit rate of at least 95%, over at least n=20 events. The n floor matters as much as the rate: a 100% hit rate over three events is a coincidence, not a lock.
When a word has been said in every single call we have -- say 4 out of 4 -- the raw base rate is 100%. We never publish 100% (or 0%): with only a handful of calls, "always so far" is not "always". Shrinkage is the standard statistical fix: before dividing, add half a hit and half a miss to the tally. 4-for-4 becomes 4.5 out of 5, i.e. 90%; 0-for-4 becomes 0.5 out of 5, i.e. 10%. The more history a term has, the less this matters (11-for-11 shrinks only to 96%). The name on the technique is the "Jeffreys prior" -- when a sheet row's note says Jeffreys shrinkage, this is all it means. Rates that aren't at the 0%/100% extremes are published as-is.
"Kalshi mode" implements Kalshi's own published NEWMENTION Payout
Criterion as closely as a text-only engine can: plural/possessive forms of
a listed term count, other grammatical derivations (tense, adjectival
forms, and the like) don't, open/hyphenated compounds count as separate
words, closed compounds don't. It is the only matching mode this project
ever actually publishes or trades a number against -- see
src/mentions/match.py for the full rules citation and worked
examples.
Yes, like the score. Born and raised in the brier patch.