Sep 21 04:01 shared_audio seeds 43/44, widened split 04:01 ihm vs sdm cross-domain eval of the audio model, 5 test meetings 04:01 group-vs-individual breakdown: text-only vs audio-augmented, widened test split 03:35 sdm (distant array mic) log-mel features for the 5 widened-split test meetings (mixed-audio ablation) 03:08 ihm log-mel prosody features, all 22 meetings accumulated (fixes last-meeting-only bug) 02:38 ihm log-mel prosody features for all labeled dacts (22 meetings, RMS-validated channel mapping) 02:06 seed stability: baseline+shared, seeds 43/44, widened split 01:51 tau dial (val+test) for baseline and shared INT8 models on widened split 00:30 consolidated improvement-phase metrics: widened split baseline + shared-context encoder + INT8 latency Sep 20 22:32 widened leave-meetings-out split (test=ES2002a,ES2008a,IS1003b,IS1008b,TS3005a) 21:28 Name-regex baseline on test split (ES2002a/ES2008a/TS3005a), computed in sandbox, reconstructed verbatim from run output after sandbox access loss Baseline: candidate in proposed set iff first name in dact text (word-boundary, case-insensitive); 1 match -> singleton, >1 -> all matched, 0 -> fallback. TS3005a names unrecoverable from AMI manual 1.6.2 (participantSummaries missing; introductions redacted) so it runs under unknown-name rule (degenerate all-4). 19:19 Add data/stats.json 19:19 Add data/records_all.jsonl 19:19 Add data/records.jsonl 19:11 Add data/stats.json (stage-1 AMI addressee extraction) 19:11 Add data/participants.json (stage-1 AMI addressee extraction) 19:11 Add data/records.jsonl (stage-1 AMI addressee extraction)