Multimodal Ingestion

The branch was forced for evaluation; product admission remains critical-only. This qualifies one English historical printed page and the real single-process worker path, not handwriting, multilingual or arbitrary-layout OCR, representative resources, production service levels, independent assurance or certification. Open the byte-mirrored result.

One 90-second English four-speaker meeting only. The reference serializes overlapping speakers; this is not diarization, speaker attribution, overlap-aware scoring or a production service level. Open the checked result.

Two English recordings with controlled zero-valued silence envelopes. This qualifies deterministic gating and absolute-time preservation, not frame-labelled VAD recall, real noisy voice notes, enhancement or transcription accuracy. Open the checked result.

One 8.308-second English source and one controlled -20 dB case. This qualifies the named build and decisions, not arbitrary codecs, corrupted media, field noise, denoising, ASR accuracy, patents or a production service level. Open the checked result.

Eight English read-speech recordings selected after screening one 113-file VOiCES shard. This is a regression characterization, not a prospective or representative voice-note/conversation, generic denoising, diarization, language/accent, production or certification claim. Open the checked result.

One five-minute English four-speaker conversation selected before model execution. No speaker labels were emitted: this qualifies the conditional background fallback for the exact profile, not diarization, representative field voice notes, language/accent breadth, production or certification. Open the checked result.

The real S21 loopback passed with 79 bounded spans and 4 anonymous clusters, but it cannot override the failed DER/JER gates. Transcript spans missed by an otherwise valid timeline remain unattributed instead of receiving invented labels. Each derived receipt is capped at 200,000 spans and 32 MiB, with 100,000 identities per directory. The profile remains experimental, outside the retrieval chunk index, production-unqualified and uncertified. Open the checked decision.

Timeline readback, overlap preservation and raw-source deletion all passed. This qualifies one real single-process lazy lifecycle; PostgreSQL writing was replaced by canonical-record validation/counting. It does not qualify the diarization model, unknown-speaker counting, representative meetings or voice notes, distributed operation, production service levels or certification. Open the byte-mirrored result.

The first request-count prediction was rejected, transparently corrected from the deterministic VAD plan, and re-frozen before this accepted result. This qualifies one real forced production-worker transition, not semantic accuracy, representative devices/languages/meetings, distributed operation, production service levels, independent assurance or certification. Open the byte-mirrored result.