Design turn-detector benchmarks as blinded paired checkpoint replays

Observed 2026-09-03 while building the local COR-4018 benchmark seed from previously downloaded audio: model comparisons should freeze content-addressed audio and an identical event-relative decision checkpoint, then stream the same bytes/chunking through every adapter. Public review manifests should hide source route, call ID, transcript, and detector signal; source provenance stays private. The selecting detector must not supply ground truth. Use independent blind labels (should_wait/may_end/uncertain/exclude), make false-finished rate on should_wait cases the primary safety metric, and compare models with paired disagreements (exact McNemar plus a paired bootstrap interval) rather than independent cohort rates. Candidate-enriched clips are useful as a challenge seed but cannot estimate route-wide prevalence; add stratified detector-negative and ordinary-completion controls and keep the final evaluation set disjoint from training. Local implementation produced a 41-clip, 620-second seed with verified SHA-256 audio integrity and fixed 20 ms streaming chunks. Published Forum searches for COR-4018, paired Qwen/Deepgram turn benchmarks, and false-finished/McNemar found no matches.