# Attribution bake-off — round3-closed-cast-8

Variant **closed-cast** · 8 video(s) · model `qwen2.5:7b` @ numCtx 8192 (maxCues 600) · 2026-08-08T07:07:29.963Z

**No sidecar was written.** `attribution.json` count 9 before, 9 after.

## Verdict against the thresholds fixed before the run

| criterion | bar | measured | |
| --- | ---: | ---: | :-: |
| new labels/chunk (mean) | ≤ 0.3 | 0.27 | PASS |
| transcript-text labels (worst video) | 0 | 0% | PASS |
| s/chunk engine | < 25 | 28.6 | FAIL |
| off-cast labels (grammar held) | 0 | 0 | PASS |

**NOT MET.**

Round 1 (open schema, same model/context/videos) measured **14.7 new labels/chunk** and **66.2 s/chunk engine**. Digest lane baseline: 11.2 s/chunk engine.

Single-label (degenerate-risk) videos: 4 of 8. A perfect headline with one label at ~100% talk time is worthless — read the two together.

## Pass A — cast discovery (1 call per video)

| video | cast | rejected | call s |
| --- | --- | ---: | ---: |
| the-quartering-rumble/v6xl0vu | Jeremy (0.90), Stephen Colbert (0.85), Donald Trump (0.90) | 0 | 9.4 |
| chibi-reviews/VQykVuHd9xQ | Reviewer (0.90) | 0 | 3.4 |
| the-quartering/J0ySGwzP4Nw | Tim Pool (0.90), Twitter (0.50), Caller (0.50) | 0 | 4.2 |
| destiny/5nmDzKB23OU | Caller (0.50), Host (0.90) | 0 | 3.8 |
| angryjoeshow/RPJqkewZP5I | AngryJoe (0.90), Caller (0.50) | 0 | 4.2 |
| rekietalaw/EsZhaCfc8HQ | Nick Ricada (0.90), Scarlett Hampton (0.50), Jacob Frey (0.20), Tim Walls (0.20), Cecil (0.50), Trinity Beholder (0.50), Dick Masterson (0.20) | 0 | 6.5 |
| chrissie-mayr/cfLF2o2-0BA | Frankie McDonald (0.90), Chrissie Mayr (1.00) | 0 | 3.9 |
| HasanAbiVODs/GjPX_ueTdfc | Cody (0.50), Troop AJ Hunter (0.90), Lane (0.50), Tony (0.50), Raymond (0.50) | 0 | 5.6 |

**Pass A is independently valuable.** If pass B churns, this alone is the
cast-list product — ~1 call per video, order 10 days corpus-wide — shipping as
"videos featuring X" without ever claiming who said which line.

## Cross-chunk identity, and the label space

| video | chunks | labels | new/chunk | round-1 new/chunk | transcript-text | truncated | singleton | top share |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| the-quartering-rumble/v6xl0vu | 1 | 2 | n/a | — | 0% | 0% | n/a | 85% |
| chibi-reviews/VQykVuHd9xQ | 1 | 1 | n/a | — | 0% | 0% | n/a | 100% |
| the-quartering/J0ySGwzP4Nw | 1 | 1 | n/a | — | 0% | 0% | n/a | 100% |
| destiny/5nmDzKB23OU | 4 | — | n/a | 0.33 | n/a | — | n/a | — |
| angryjoeshow/RPJqkewZP5I | 3 | 1 | 0.50 | — | 0% | 0% | 100% | 100% |
| rekietalaw/EsZhaCfc8HQ | 8 | 2 | 0.29 | 12.57 | 0% | 0% | 50% | 81% |
| chrissie-mayr/cfLF2o2-0BA | 9 | 1 | 0.13 | — | 0% | 0% | 100% | 100% |
| HasanAbiVODs/GjPX_ueTdfc | 13 | 2 | 0.17 | — | 0% | 0% | 0% | 65% |

## Pass B — how often the model declined

| video | turns accepted | Unknown | off-cast | out-of-range |
| --- | ---: | ---: | ---: | ---: |
| the-quartering-rumble/v6xl0vu | 2 | 3 | 0 | 0 |
| chibi-reviews/VQykVuHd9xQ | 2 | 0 | 0 | 0 |
| the-quartering/J0ySGwzP4Nw | 5 | 5 | 0 | 0 |
| destiny/5nmDzKB23OU | 0 | 64 | 0 | 103 |
| angryjoeshow/RPJqkewZP5I | 3 | 0 | 0 | 177 |
| rekietalaw/EsZhaCfc8HQ | 88 | 125 | 0 | 0 |
| chrissie-mayr/cfLF2o2-0BA | 83 | 257 | 0 | 60 |
| HasanAbiVODs/GjPX_ueTdfc | 155 | 393 | 0 | 180 |

`Unknown` is a GOOD answer, not a failure — the bias must run toward dropping on
uncertainty. `off-cast` must be 0: anything else means the enum did not convert to
a grammar and the whole hypothesis is refuted.

## What the labels say

- `the-quartering-rumble/v6xl0vu` — Jeremy 15%, Donald Trump 85%
- `chibi-reviews/VQykVuHd9xQ` — Reviewer 100%
- `the-quartering/J0ySGwzP4Nw` — Tim Pool 100%
- `destiny/5nmDzKB23OU` — (no speaker)
- `angryjoeshow/RPJqkewZP5I` — AngryJoe 100%
- `rekietalaw/EsZhaCfc8HQ` — Nick Ricada 81%, Trinity Beholder 19%
- `chrissie-mayr/cfLF2o2-0BA` — Chrissie Mayr 100%
- `HasanAbiVODs/GjPX_ueTdfc` — Tony 35%, Raymond 65%

Warnings by code: none

## Cost

40/40 chunk(s) ok, 48 call(s) total (40 chunk call(s) with engine timing). Sample density **2.26 chunks/audio-hour** — the corpus census reads 2.46.

| quantity | value | round 1 |
| --- | ---: | ---: |
| s/chunk, wall | 28.6 | — |
| s/chunk, engine | 28.6 | 66.2 |
| s/chunk, engine excl. model load | 28.0 | — |
| mean model-load s/call | 0.22 | — |
| decode tokens/s | 42.8 | — |
| TOTAL output tokens | 38799 | — |
| cast call s/VIDEO (not per chunk) | 5.1 | — |
| digest chapters baseline, s/chunk engine | 11.2 | — |

Output tokens are the thing to watch: engine time tracked wall time within 7% in
round 1, so 66.2 s/chunk was pure decode volume (31,266 output tokens on one
13-chunk video). An enum token per turn instead of a 60-character string is the
mechanism by which this variant is meant to be cheaper as well as better.

## Corpus projection — per-turn lane

Census: 74,329 transcribed video(s), **194,061 chunks** at maxCues 600.

**62.9 days idle to 64.2 days contended** — from n = 8 video(s) / 40 chunk(s), one box, one hour.

Cast-only, for comparison: **1 call per video** over 74,329 videos at 5.1s = 4.4 days.

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.

## Box

Start: `{"loadavg":[0.13,0.35,1.02],"freeMemGb":11.35,"totalMemGb":15.53}`

End: `{"loadavg":[0.44,0.77,1.02],"freeMemGb":3.69,"totalMemGb":15.53}`

## Units

- The unit of per-turn 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 cast discovery is **calls per video**, priced separately and never
  folded into s/chunk.
- 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.