# MOSS embodied-podcast engineering outcome

The Zhang Yi canary process returned successfully, but the complete-episode engineering check
failed. Its 632 parsed segments end at 1284.75 seconds while the decoded input lasts 2861.511125
seconds: 44.8976% timeline coverage and a 1576.761125-second trailing gap. The raw transcript ends
mid-marker (`[12`). Therefore this is a truncated output, not a completed podcast transcription.

The clean inference took 5576.416 seconds (RTF 1.94877), produced exactly 16,384 reported tokens,
and retained valid runtime measurements. That RTF describes this incomplete generation only. It is
not comparable as a completed-episode RTF and cannot be used for accuracy or ranking.

The frozen runner did pass the requested `max_new_tokens=65536`; the model default of 5,120 was
overridden, and prompt plus reported output remained below the 131,072 text context. The runner did
not persist generated token IDs, final token/EOS state, scores, a streamer count, or the stopping
criterion that fired, so the exact internal reason for stopping at 16,384 is indeterminate from the
retained artifacts. No rerun was performed.

The Shibo episode was not started because the canary failed. The Jia Peng episode remains excluded
because its 2:47:58.968 duration exceeds the official 90-minute input range; it was not chunked to
claim a native complete-episode run.

The successful retry has valid 8-artifact lineage and an exact reusable raw cache hit. All 113 tests
and Ruff pass. The initial m4a attempt is retained as a no-GPU pre-model failure; one later lineage
observation record has a parent sample-id mismatch and was deliberately not rewritten.