The latency argument in the review deserves arithmetic. Control loops in the reviewed systems run at 1 to 15 minute intervals [13], which is 60 to 900 seconds of wall clock [20], and the same passage puts LLM inference at seconds [13]. One call per loop fits, with margin. For latency to actually bind you need fan-out: an inference per zone or per air handler inside the same window, a retrieval hop in front of each one, and a retry when the model emits a schedule that fails validation. That is a plausible deployment, and the corpus's most advanced reported deployment tops out at pilot level, per the review [5].
The safety argument needs no such assumptions. A physics-based controller has bounded failure modes and an LLM does not, per the review as summarized on dev.to [14], and the class of failure being bounded against is frozen pipes and carbon monoxide buildup [15]. Hence the split in the twelve control studies: the model emits a policy or a setpoint schedule, and model predictive control, reinforcement learning or rule-based logic does the actuator write [12]. Conventional ML, MPC and RL stay more adopted for high-frequency control and short-horizon numerical forecasting [16].
Which leaves the naming layer, and that part is worth copying. The same measurement arrives as AHU-1.ZN-T in one building, AHU_01_ZONE_TEMP in the next, ahu1_zt_sensor in a third [7]. Several studies retrieve building-specific documentation and have the model emit a mapping from raw names to Brick or Haystack [9]. The pipeline pulls names from BACnet discovery or a CSV export, retrieves equipment schedules and control sequences, prompts for a mapping table, and stops for operator approval before anything downstream uses the schema [10]. Two properties make that bounded. The artifact is a table, so it can be checked offline against the equipment list. And the model's contribution stays confined to producing that table; control commands and sensor predictions remain outside its output [11].
The honest limit is in what the review counts. Four of 66 is about six percent [21], and a readiness map grades published papers, not commissioned buildings. Failure attribution also sits outside its scope. Nothing in it isolates model capability from data plumbing, so the case that plumbing is the binding constraint rests on an absence: most studies assume clean, labeled, synchronized streams [18], while the review lists the conditions a production agent meets immediately, which are a sensor stuck at one value after its network dropped, a point name arriving with no unit or equipment association, and two thermometers in one zone disagreeing [17]. An occupancy sensor reading zero because someone taped over it [17] is not a model-tier problem.
So the deliverable the corpus supports is a reviewed mapping table plus staleness and conflict handling you write yourself, underneath document-grounded operator support [19]. The write-up leaves the review's authors and venue unnamed [23], so treat the counts as a claim about a corpus you cannot yet re-code.
