Punctuation is a three-way classification, not a model call

What you outsource when you post a transcript to a model is a choice between three symbols ^[5][6]. The model resolves that choice by sampling, which is why the same sentence can come back two ways ^[3]. So the trade is an open-ended generator plus a round trip you can feel, set against keeping a rule list yourself ^[3].

The rules are not hidden. In unpunctuated English, interrogatives overwhelmingly open with an auxiliary or a wh-word, and exclamatives have their own recognizable openers ^[7]. "can you" does not require inference ^[7]. The author's library, sentencify, walks an ordered array of regular expressions per language, returns the type of the first rule that matches, and defaults to declarative when nothing does ^[8][12].

Where it stops being an afternoon's work is the second locale. Appending a mark is only correct in some languages: Spanish needs an inverted question mark prepended as well as the closing one appended, French typography puts a space before the question mark, exclamation mark and colon, and Japanese uses its own full-width period and question mark rather than the ASCII characters ^[10]. With six languages shipping, that is eighteen language-and-type outputs to get right ^[11], and the detection cues do not transfer between them, since a rule that recognises an English auxiliary tells you nothing about Japanese ^[17].

The fragility is in the same design decision that makes it readable. First match wins, so a broad interrogative rule sitting above a narrow exclamatory one will swallow it silently, and nobody finds out until a user reports one specific sentence coming out wrong ^[13]. The author names the alternative, scoring every match and taking the highest confidence, calls it more robust and considerably harder to debug, and says he is not certain he chose correctly ^[14]. That hedge is worth more than a marketing line, because it tells you where to look when the output is wrong. So does exporting the rule table, which lets you find the rule that fired instead of guessing ^[15].

No accuracy figure is published in the post ^[18]. That matters for how you read the determinism claim ^[16]: what you are guaranteed is that a misclassification repeats identically, not that there are fewer of them. A deterministic wrong answer is a fixture you write once and fix once; a sampled wrong answer is a flake you can only re-roll ^[19]. That is also what makes the synchronous path usable on every keystroke or every streamed token, where a model call cannot go at all ^[16].

Which puts the real value somewhere other than the English regexes. The author says you could write the naive English version yourself in an afternoon ^[17]. The part you would not write is the per-locale typography, and the part you would get wrong is the ordering. Both are now someone else's ordered array, published where you can read it ^[15].