An availability word from the clerk now revokes Oathra's earlier booking confirmation

The interesting case is where the negation lives. In "We can hold September 25 at 7:30 pm for now, but it is not confirmed yet", the clause carrying the time is the clause being negated ^[7]. Extract per utterance and the time lands in one variable while "not confirmed yet" lands in another, and the negation is lost on the way ^[8]. So the extractor has to work at clause level.

The third reply is worse, because the agent finishes the trap itself. The clerk says "The only thing left that evening is 9 pm. Would that do?" A number was said, and it was offered as a question ^[9]. The agent's next turn is usually "9 pm works, thank you", so if the extractor took 21:00 from the clerk's turn you get a fully populated booking that nobody agreed to ^[10]. Oathra holds the clerk's turn as a proposal until the caller accepts it and the clerk acknowledges the acceptance ^[11].

Two of the five shapes were mislabelled by the checker itself until this week, and they failed in opposite directions ^[26]. The fully-booked reversal failed open, because the old retraction pattern required the refusal to name the booking, in the manner of "we cannot take the reservation" ^[13]. The plain confirmation failed closed: "You are all set for September 25 at 7:30 pm, party of two" returned nothing at all, not "unconfirmed" but empty ^[18]. The engine kept two lists, one for the callee agreeing to a value the caller proposed and one for the callee confirming the booking, and "you are all set" was in the second but not the first ^[19]. A confirmation is bound to the values settled when it was spoken, so a confirmation with no settled values behind it is stale and gets dropped ^[20].

That bug surfaced by accident. "I only found it because I pasted an English log into my own public checker while writing a comment on someone else's thread," the maintainer wrote ^[21]. Three of the five natural English confirmations worked; the Japanese side, used daily, was fine ^[22]. "The lesson is not about regexes; it is that the language you don't test in is the language that's broken," the maintainer wrote ^[23].

The seeded corpus behind the release gate is the author's own: the five shapes plus voicemail, transfers, hold-then-reply, dialect confirmations and wrong restatements ^[25]. A confirmation followed by a bare availability refusal cannot have been in that set while the runtime was still scoring it as completed ^[28]. For the zero to say anything about your callees, your callees' phrasings have to be in the set.

PR #37 is narrow on purpose. An availability word from the callee (full, private hire, closed, "fully booked") revokes a confirmation spoken strictly earlier ^[14]. "We can't take cards" after a booking stays a payment remark, and the ordinary "7 pm is full but 7:30 is free" before a booking is untouched ^[15]. The tests run in both directions ^[16]. It is still a vocabulary, so a clerk who kills the booking in words outside it leaves a confirmed result standing.

The case for taking the verdict away from the model rests on those five replies as the maintainer reads them; the post does not report a measured comparison between a model's verdicts and the checker's ^[27]. I would take the separate check anyway, on the strength of the first reply: a check that counts filled fields passes it, and the date, the time and the party size are all correct, while the clerk has only promised to call back ^[5][6].