Put the two headline rates into one stream of traffic and they stop meaning what they appear to mean. The 92 and the 76 are per-class rates [6][7], and a badge does not appear inside a class. It appears in a chat where most messages are pleasantries. The source never says how often a message deserves an action, so here is a round assumption of mine: one in twenty. Across 1,000 messages that gives 50 real requests, 46 of them caught, and 950 harmless ones, 228 of which get a badge anyway. Of the 274 badges shown, 228 are wrong, or 83 percent [1]. The author's line about a badge being wrong one time out of four [12] describes the harmless-message rate. What a user experiences is the badge's own hit rate, and in a quiet channel that is the worse number.
That arithmetic also settles whether more anchors were ever going to help. An anchor is one more point in the same space, and the rule is a comparison between two nearest neighbours [4]. "The outage is fixed, thanks" keeps the vocabulary, and therefore the topic, of the tool it matches [8]. Adding "it's fixed" to the anchor list only wins if that phrase pulls on the sentence harder than the outage words do, and the author's own two-minute check says it does not: append "it's fixed, thanks" to positive sentences and the scores barely move [9]. So the remaining lever is the bar, and the bar works inside a band 0.12 wide [2], where "Hello" and "create a ticket" already score almost the same [3]. Twelve anchors and a higher bar looked good on the test set; on unseen messages one real case in two went missing [10][11], roughly six times the miss rate the test set reported [3].
Half the bounding is already in the architecture, and that is the part worth copying. Stage 2 runs only when a user clicks the badge [2], so a false positive costs nothing in tokens and nothing in latency. It is charged entirely against the user's willingness to look at the next badge [12]. That moves the design question from how high specificity can be pushed to what a wrong badge is permitted to do. A passive suggestion survives 24 misfires per 100 harmless messages [7]. Anything that files the ticket itself does not.
Two details make the published numbers optimistic rather than pessimistic. The evaluation ran on 300 generated messages with no users on the system, and the author states plainly that reality will do worse [5], so 76 is a ceiling and 24 is a floor. And the detector splits long messages on punctuation to lift a request out of surrounding politeness, which quietly does nothing when there is no punctuation, as in voice dictation [13]. Generated test text punctuates. Dictated text does not, so the blind spot cannot show up in this score at all.
The deliverable here is not a tuned threshold but a written limit: this stage does not read negation, and nothing downstream may behave as though it does [14].
