Skip to content

Corpus comparison — v2-cloud vs v3-dgx

Shared episodes (joined by GUID): 9

Deterministic metrics only — no LLM judge. Summary/insight text quality needs a cross-vendor judge panel (autoresearch/JUDGING.md) and is scored separately.

metric v2-cloud n v3-dgx n delta
has_summary 1.0 9 1.0 9 +0.00
insights 12.0 9 9.78 9 -2.22
kg_edges 26.11 9 23.78 9 -2.33
kg_nodes 27.11 9 22.11 9 -5.00
kg_person_placeholders 0.0 9 0.0 9 +0.00
kg_persons 0.0 9 8.44 9 +8.44
n_segments 1067.56 9 1052.78 9 -14.78
quote_attributed_pct 100.0 9 100.0 9 +0.00
quote_ts_valid_pct 100.0 9 100.0 9 +0.00
quotes 24.67 9 2.33 9 -22.34
summary_chars 1488.89 9 1284.33 9 -204.56
timeline_error_pct 1.92 9 0.19 9 -1.73
transcript_span_s 4017.69 9 4100.49 9 +82.80
voices 8.11 9 7.44 9 -0.67
voices_named 2.67 9 6.11 9 +3.44

How to read the load-bearing rows

  • timeline_error_pct — how far short of the audio the transcript ends. This is the #1173 signature: a corpus transcribed from silence-stripped audio ends early, and every timestamp derived from it is wrong. Near 0 is correct.
  • quote_ts_valid_pct — a range check only: does the timestamp fall inside [0, duration]? A quote drifted by a minute still does, so this cannot see drift. It scored 99.6% on the corpus whose timestamps were broken. Never read it as accuracy.
  • voices_named vs voices — voices resolved to a real person rather than left as SPEAKER_NN. Unresolved placeholders must never reach an entity surface (#1167).
  • kg_person_placeholders — should be 0. Anything above it is a leak.