Earshot

Earshot listens to every fireground radio channel so no mayday goes unheard. Problem: 35.6% of firefighter maydays are missed on the first call (Project Mayday, >9,000 maydays). The incident commander (IC) monitors several channels on handhelds while talking on one of them, and the crew inside hears either everything or nothing. In Bucheon, Korea (16 Oct 2026) a firefighter's isolation was recognised at 09:09; he was found at 09:35. What it does — three pillars, all driven by live Meta Model API calls: 1) Voice triage for command. muse-voice-transcribe-1.0 realtime streams on every channel (command + tactical) with SOP keyword biasing; muse-spark-1.3 turns each transmission into a strict-JSON event (MAYDAY / DANGER / VICTIM_FOUND / HELP_NEEDED / ROUTINE) with a verbatim evidence span and the Korean SOP 109 fields (who / what / where / needs). The IC gets one must-acknowledge queue: max 5 open cards, a 10 s ack-escalate loop, missing fields highlighted. 2) Hands-free crew side. A clip-on display + earpiece (simulated as a phone-sized web view). A muse-spark-1.3 Responses-API agent with 7 tools decides, per firefighter, push / hold / ask: a two-sentence brief only to the crews it affects, and it asks the missing SOP field back over the earpiece. Four hold-to-confirm buttons (mayday, victim here, search complete, need support) plus push-to-talk transcribed by Muse. 3) World-model harness. Units, zones, strategy, cards, 7 time-bounded rules (ack escalation, three unanswered pings = SOP 109 emergency, danger-after-offensive, RIT not yet proposed, ...) and an invariant checker that rejects agent briefs aimed at the wrong sector. The agent decides who hears what; the commander decides who does what — every order is a proposal the IC approves. Evaluation (n = 36 synthetic Korean/English lines, clean and AMR-NB radio-degraded): MAYDAY recall 13/13, codeword-free distress 5/5, keyword biasing lifts critical-term recovery from 11/14 to 14/14 on radio audio; live card latency p50 4.3 s after end of transmission. Limits (stated in the README and video): simulated hardware, synthetic Gemini-TTS audio (the Meta API has no TTS; disclosed), reported zones not tracking, decision support only.