# Attribution pilot — round1-bakeoff

Lane **text-only** · sample `bakeoff` · 8 video(s) · model `qwen2.5:7b` @ numCtx 8192 (maxCues 600) · 2026-08-08T04:42:41.909Z

Outcomes: attributed 8

## Cross-chunk identity — the property the lane exists for

| video | chunks | labels | new labels/chunk | singleton rate | top share | near-dup | generic |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| the-quartering-rumble/v6ve4n0 | 1 | 4 | n/a | n/a | 35% | 0% | 0% |
| chibi-reviews/VQykVuHd9xQ | 1 | 1 | n/a | n/a | 100% | 0% | 100% |
| the-quartering/J0ySGwzP4Nw | 1 | 2 | n/a | n/a | 54% | 0% | 100% |
| destiny/5nmDzKB23OU | 4 | 16 | 0.33 | 100% | 64% | 81% | 6% |
| angryjoeshow/RPJqkewZP5I | 3 | 3 | 1.50 | 100% | 43% | 0% | 0% |
| rekietalaw/EsZhaCfc8HQ | 8 | 90 | 12.57 | 100% | 17% | 79% | 0% |
| chrissie-mayr/cfLF2o2-0BA | 9 | 242 | 30.25 | 100% | 22% | 98% | 0% |
| HasanAbiVODs/GjPX_ueTdfc | 13 | 346 | 28.83 | 100% | 26% | 52% | 0% |

**Headline** — new labels per chunk after the first, over 5 multi-chunk video(s): mean 14.70, max 30.25.

Read it WITH the top-share column: one label at ~100% scores a perfect 0 here
and is degenerate, not good.

Single-label (degenerate-risk) videos: 1 of 8. Videos whose chunk reconstruction disagreed with the record: 0.

## Coverage and label quality

| video | segments | mark-free chunks | coverage | max gap | median gap | non-ascii labels |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| the-quartering-rumble/v6ve4n0 | 4 | 0 | 100% | 00:00:00 | 00:00:00 | 0% |
| chibi-reviews/VQykVuHd9xQ | 1 | 0 | 99% | 00:00:05 | 00:00:05 | 0% |
| the-quartering/J0ySGwzP4Nw | 8 | 0 | 99% | 00:00:05 | 00:00:04 | 0% |
| destiny/5nmDzKB23OU | 16 | 0 | 82% | 00:13:15 | 00:06:40 | 63% |
| angryjoeshow/RPJqkewZP5I | 5 | 2 | 24% | 00:42:09 | 00:21:22 | 0% |
| rekietalaw/EsZhaCfc8HQ | 251 | 0 | 90% | 00:19:34 | 00:10:08 | 1% |
| chrissie-mayr/cfLF2o2-0BA | 313 | 2 | 67% | 01:08:05 | 01:08:05 | 0% |
| HasanAbiVODs/GjPX_ueTdfc | 509 | 3 | 83% | 01:20:13 | 01:20:13 | 0% |

Coverage is ~1.0 BY CONSTRUCTION for this lane — marks are tiled, each running
to the next change — so it is a structural check, not a quality signal. The
honest coverage metric here is mark-free chunks.

Warnings by code: chunk-failed 3, bad-timestamp 11

## Cost

Measured over the 8 video(s) this run actually attributed (37/40 chunk(s) ok, 37 call(s), 37 with engine timing). Sample density **2.26 chunks/audio-hour** — the corpus census reads 2.46.

| quantity | value |
| --- | ---: |
| s/chunk, wall (video wall ÷ chunks) | 70.8 |
| s/chunk, per-call wall | 66.2 |
| s/chunk, engine | 66.2 |
| s/chunk, engine excl. model load | 65.4 |
| mean model-load s/call | 0.42 |
| decode tokens/s | 41.1 |
| digest chapters baseline, s/chunk engine | 11.2 |

## Corpus projection

Census: 74,321 transcribed video(s), **194,053 chunks** at maxCues 600.

**146.8 days idle to 159.0 days contended** — from n = 8 video(s) / 40 chunk(s), one box, one hour.

This is a SECOND sweep on the same 8 GB card, additive to a digest sweep that
has completed 0.17% of its own ~194k calls, with nothing arbitrating between
them. One lane, no parallelism.

## Box

Start: `{"loadavg":[10.67,6.69,3.89],"freeMemGb":7.75,"totalMemGb":15.53}`

End: `{"loadavg":[2.04,1.55,1.65],"freeMemGb":11.65,"totalMemGb":15.53}`

## Units

- The unit of text-only work is the **chunk** (one model call). Seconds-per-
  audio-hour is not a unit here: chunk density varies 4× across this corpus, and
  pricing in it is the error behind a retracted throughput headline. This report
  does not print one.
- The unit of diarized work is **calls per video**.
- Wall and engine are different quantities and are never substituted for each
  other. Wall is the realistic ceiling on this shared box; engine excl. load is
  the floor.
- Every rate carries its denominator. At n ≈ 8 videos a bare percentage would be
  a lie of precision.