Answers are grounded in this paper and cite their sources.
Green-Route is a general routing framework instantiated for digital tourism that jointly considers inference-related operational carbon, modelled service adequacy, queue capacity, response-time targets and mandatory safety escalation. A benchmark of three local models, 30 English-language tourism prompts and 5 repetitions on one Apple M1 Pro calibrated energy and latency. A 24-h queueing simulation evaluated four synthetic demand scenarios using 200 parameter draws and 10 paired arrival replications. The hosted-tier roles are hypothetical; same-device calibration does not validate hosted performance. Among the evaluated deployable policies at their fixed settings, Green-Route had the lowest operational carbon when comparison was restricted to policies reaching its own median modelled service adequacy. Only large-model-only routing cleared that threshold, at 18.5–38.5% higher carbon. Mid-default and fixed-threshold policies emitted 27.0–55.4% less but achieved lower modelled adequacy. Green-Route’s median modelled service adequacy remained between 83.4 and 87.1%. Under festival and disruption surges, response-time compliance fell to 68.2% and 77.3%, with timeout rates of 13.8% and 13.3%. All flagged safety-critical requests were escalated, verifying the routing rule rather than response safety. Three alternative adequacy functions preserved the deployable carbon-adequacy frontier in 11 of 12 cases; the flattest function made Green-Route dominated off-peak. These results demonstrate conditional carbon-service trade-offs within a defined inference boundary. They support site-specific calibration and testing, not field effectiveness or superiority to untested routing policies.
No takes yet. Share an insight, caveat, or question.
SharifAli et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: