# Attribution bake-off — round2-closed-cast

Variant **closed-cast** · 2 video(s) · model `qwen2.5:7b` @ numCtx 8192 (maxCues 600) · 2026-08-08T06:12:27.088Z

**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 | 0.29 | PASS |
| transcript-text labels (worst video) | 0 | 0% | PASS |
| s/chunk engine | < 25 | 24.1 | PASS |
| off-cast labels (grammar held) | 0 | 0 | PASS |

**ALL THRESHOLDS 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: 0 of 2. 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 |
| --- | --- | ---: | ---: |
| destiny/5nmDzKB23OU | Caller (0.50), Host (0.90) | 0 | 8.9 |
| 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 | 7.0 |

**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 |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| destiny/5nmDzKB23OU | 4 | — | n/a | 0.33 | n/a | — | n/a | — |
| rekietalaw/EsZhaCfc8HQ | 8 | 2 | 0.29 | 12.57 | 0% | 0% | 50% | 81% |

## Pass B — how often the model declined

| video | turns accepted | Unknown | off-cast |
| --- | ---: | ---: | ---: |
| destiny/5nmDzKB23OU | 0 | 64 | 0 |
| rekietalaw/EsZhaCfc8HQ | 88 | 125 | 0 |

`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

- `destiny/5nmDzKB23OU` — (no speaker)
- `rekietalaw/EsZhaCfc8HQ` — Nick Ricada 81%, Trinity Beholder 19%

Warnings by code: none

## Cost

12/12 chunk(s) ok, 14 call(s) total (12 chunk call(s) with engine timing). Sample density **2.55 chunks/audio-hour** — the corpus census reads 2.46.

| quantity | value | round 1 |
| --- | ---: | ---: |
| s/chunk, wall | 24.1 | — |
| s/chunk, engine | 24.1 | 66.2 |
| s/chunk, engine excl. model load | 23.3 | — |
| mean model-load s/call | 0.34 | — |
| decode tokens/s | 41.3 | — |
| TOTAL output tokens | 8923 | — |
| cast call s/VIDEO (not per chunk) | 7.9 | — |
| 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.

**52.4 days idle to 54.0 days contended** — from n = 2 video(s) / 12 chunk(s), one box, one hour.

Cast-only, for comparison: **1 call per video** over 74,329 videos at 7.9s = 6.8 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":[6.23,4.5,3.66],"freeMemGb":7.83,"totalMemGb":15.53}`

End: `{"loadavg":[4.92,5.16,4.2],"freeMemGb":5.42,"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 = 2 videos a bare percentage
  would be a lie of precision.