Mention-detection ships because its trigger is observable. Microsoft's Facilitator marks a topic covered once the discussion for that topic has started [1], which is a claim about the agenda rather than about the record, and Balto ticks when an item is mentioned [2]. Both verdicts are decidable from the transcript alone, and both fail without noise. A wrongly ticked item raises no error. It produces a blank field the practitioner finds at her desk that evening, when closing it costs a phone call or a second visit [6].
Adjudication is the harder problem because the verdict is a judgement: does "a couple of wobbles" satisfy a form item that asks for the number of falls and the circumstances of the most recent [7]. Junwei Lai's build, called intake and written for an August 2026 hackathon [8], makes that judgement the gate and then spends the engineering on making it cheap.
Two mechanisms carry that. The first is that the slot state is the state, not the transcript. The obvious design accumulates the conversation and re-asks the model what is still missing each turn, which turns a forty-five minute interview into tens of thousands of growing tokens that degrade as they grow [18]. Instead, each call receives the open items, a fixed-size struct of current values, and the new audio, so a three-hour interview costs the same per chunk as a ten-minute one [19]. Flat per-chunk cost is what makes a long home visit affordable at all.
The second is fan-out behind a router. Each open item gets an isolated adjudication call, so a wrong verdict cannot corrupt a neighbour, each item is separately scoreable by the eval, and k concurrent calls cost about one call's wall time [20]. The bill is tokens, not wall time. Without the router, every open item was adjudicated against every chunk, and each independently decided a vague remark was relevant to it, so "a couple of wobbles" attached itself to mobility, memory and low mood as well as falls [21]. One cheap classification call cut items touched from seven to two [22]: a 71 percent reduction in adjudicator calls per chunk [23], and five fewer chances per chunk to be wrong about the same sentence [27].
Then there is the line that proves the intent. The coach's output schema has a field for a question and a field for quotes, and no field an answer could go into, so it structurally cannot suggest one [25]. The agent tracks coverage against a human-authored form and quotes the span it relied on, saying that item M14 has no recorded answer rather than that something may indicate falls risk [24]. That is the right place for the rule, because a prompt instruction is advisory and a schema is enforced.
The numbers transfer only under conditions the writeup names. The 47 labelled cases are the product, and they were authored by the same person who wrote the prompt [9]; the harness exits non-zero when an answer labelled insufficient is marked sufficient [10], which is the operating point that matters, because a false sufficient is indistinguishable from the mention-level behaviour being replaced [11]. Twelve of the cases are adversarial [13], and Lai's own standard is worth keeping: a set that scores 100 percent on its first run cannot tell you anything [16]. The text available to us stops mid-sentence before any behavioural evaluation figure [26], so the accuracy is unreported rather than good. Per-item isolation also assumes items are independent; a form whose item is derived from three other answers breaks the fan-out.
