Audio CAPTCHA answered correctly faster than its clip plays — ASR automation, not a listener
bh.arena_audio_superhuman · convicts
What it catches
Kitsune arena AUDIO-gate solve coherence — the ASR-benchmark twin of bh.arena_captcha_superhuman. The owned arena/audio gate mints a spoken-digit WAV and measures the SERVER-OBSERVED solve time (issue->verify, tracked in the audio store — unforgeable, unlike a client-asserted proof). A human must PLAY the clip in real time before answering; an ASR solver batch-transcribes it in milliseconds. A CORRECT transcription submitted faster than the clip's real-time listen floor (len x 450ms, under FSDD's ~600ms/digit playback) is automation, not a listener. The gate returns anomaly:solved_faster_than_audio; the detector relay (/arena/audio/verify) joins it to ks_sid so a solver that PASSES the gate is CONVICTED on coherence instead of cleared (the arena thesis: a challenge is not a bot/human discriminator — the SOLVE-BEHAVIOUR is). HUMAN-TIME-bound (listen+type), NOT compute-bound, so HARDWARE-INDEPENDENT. FP-SAFE BY CONSTRUCTION: the floor sits under the clip's real-time playback so no human answers before hearing it; precision 1.0 is load-bearing. GROUNDED LIVE 2026-07-08: the matched-filter evaders/arena-solver-audio transcribes an easy clip (solve_ms << floor) -> anomaly + the joined session convicts; a real-time listen is silent. Convicting (w0.85). See docs/research-radar.md (arena-gate-gaps loop).
Signals it reads
behavioral.arena_audio_superhuman
How it fires
present
Bypassed by 10
Frontier evaders that reach the detector uncaught (scored only suspicious, defeating every convicting tell) — this check is not one that stops them. The red-team frontier this detection still has to convict.