# Digest bake-off — chapters-baseline

Sample: 12 video(s) from `plans/bakeoff/sample.json` (buckets: short, medium, long, verylong), 39.61 audio-hours.
Sweep days are projected as measured seconds-per-audio-hour x 79151 corpus audio-hours, one lane, no parallelism.

| Candidate | Zero-yield chunks | Chapters/h | Rejection rate | Max gap | Generic | Dup | tok/s | s per audio-h | **Sweep days** |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| `qwen2.5:7b@8192/chunk-local` | 7/87 (8.1%) | 15.96 | 14.8% | 00:51:52 | 16.5% | 4.4% | 211.6 | 47 | **43.1** |

## Boundary accuracy vs the uploader's own chapters

The oracle is `metadata.info.json.chapters` — marks a human authored while
watching, who never saw our prompt. Boilerplate marks ("Intro", "Sponsor")
and the boundary at 00:00 are dropped before scoring: neither carries
segmentation information, and the origin would be a free hit for every
candidate. Matching is one-to-one and closest-pair-first, so a cluster of
boundaries around one uploader mark scores one match, not many.

| Candidate | Videos scored | Uploader marks | Median offset | Within 30s | P@30 | R@30 | **F1@30** | F1@60 | Name-in-title |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| `qwen2.5:7b@8192/chunk-local` | 12 | 163 | 109s | 22.7% | 5.6% | 21.5% | **8.8%** | 14.1% | 0% |

## Rejections by guard

| Candidate | language-drift | non-monotonic | out-of-range | seam-duplicate |
| --- | --- | --- | --- | --- |
| `qwen2.5:7b@8192/chunk-local` | 22 | 23 | 65 | 1 |

## Per-video

| Candidate | Video | Bucket | Chunks | Zero-yield | Chapters | Max gap | Engine s | Marks | F1@30 | Speakers |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| `qwen2.5:7b@8192/chunk-local` | `nuxanor/9bgVoJY0jEg` | short | 1 | 0 | 4 | 00:05:39 | 16 | 9 | 16.7% | — |
| `qwen2.5:7b@8192/chunk-local` | `chibi-reviews/XYyejb64QpU` | short | 1 | 0 | 2 | 00:07:28 | 3 | 8 | 20% | — |
| `qwen2.5:7b@8192/chunk-local` | `the-quartering/br9koSf6D0Q` | short | 1 | 0 | 2 | 00:05:37 | 3 | 13 | 14.3% | — |
| `qwen2.5:7b@8192/chunk-local` | `destiny/5nmDzKB23OU` | medium | 4 | 1 | 15 | 00:20:01 | 81 | 6 | 0% | — |
| `qwen2.5:7b@8192/chunk-local` | `angryjoeshow/8AJsKyh0x7w` | medium | 2 | 1 | 3 | 00:34:34 | 48 | 5 | 50% | — |
| `qwen2.5:7b@8192/chunk-local` | `kirsche/GchuqKQsnHk` | medium | 4 | 0 | 28 | 00:15:25 | 78 | 17 | 17.8% | — |
| `qwen2.5:7b@8192/chunk-local` | `HasanAbiVODs/p7-GXvM3eXg` | long | 7 | 0 | 51 | 00:39:26 | 156 | 4 | 7.3% | — |
| `qwen2.5:7b@8192/chunk-local` | `omnibased/S-9F_Jr3jmA` | long | 7 | 0 | 66 | 00:22:31 | 169 | 35 | 17.8% | — |
| `qwen2.5:7b@8192/chunk-local` | `destiny/WL9htojKbGY` | long | 14 | 1 | 67 | 00:19:28 | 263 | 9 | 2.6% | — |
| `qwen2.5:7b@8192/chunk-local` | `HasanAbiVODs/GjPX_ueTdfc` | verylong | 13 | 0 | 149 | 00:35:23 | 304 | 9 | 1.3% | — |
| `qwen2.5:7b@8192/chunk-local` | `HasanAbiVODs3/bW2lKpHlmMQ` | verylong | 16 | 1 | 122 | 00:51:52 | 351 | 15 | 4.4% | — |
| `qwen2.5:7b@8192/chunk-local` | `HasanAbiVODsbackup/5IufSgJSI0E` | verylong | 17 | 3 | 123 | 00:49:34 | 369 | 33 | 12.8% | — |

## Sample output (first chapters per video)

### `qwen2.5:7b@8192/chunk-local`

**nuxanor/9bgVoJY0jEg** (short, 00:13:37)

- 00:00:04 Introduction to the controversy
- 00:05:02 Discussion on Alistair's design and Voodoo symbols
- 00:10:40 Harassment and death threats
- 00:12:51 Personal reflections on the issue

**chibi-reviews/XYyejb64QpU** (short, 00:10:19)

- 00:00:07 Villain Character Designs
- 00:02:52 Media and News Impact on Villains

**the-quartering/br9koSf6D0Q** (short, 00:10:06)

- 00:00:01 Introduction and Media Criticism
- 00:04:29 Prank Call Details and Backstory

**destiny/5nmDzKB23OU** (medium, 01:13:39)

- 00:00:06 Introduction and Initial Agreement
- 00:01:42 Defining Religion and Atheism
- 00:04:31 The Role of Reason in Belief
- 00:07:44 Atheist vs. Agnostic Perspective
- 00:10:35 Reasoning Processes and Their Differences
- 00:13:51 Discussion on God of the Gaps Argument
- 00:16:14 Impact of Organized Religion on Society
- 00:19:29 Introduction and Skepticism

**angryjoeshow/8AJsKyh0x7w** (medium, 00:50:41)

- 00:34:34 Gameplay Issues
- 00:43:21 Endgame and Crafting Issues
- 00:47:11 Final Verdict and Future Prospects

**kirsche/GchuqKQsnHk** (medium, 01:09:15)

- 00:15:16 Shooting at Wazada High School Graduation Ceremony
- 00:15:35 University of Minnesota Police Response
- 00:16:30 Surveillance Video and Suspect Description
- 00:17:46 Suspect's Arrest and Charges
- 00:19:46 University of Minnesota Decision to Stop Hosting Graduations
- 00:21:06 Impact on Local School Districts
- 00:21:52 Anoka-Henipin School District's Response
- 00:37:17 Introduction of the incident

**HasanAbiVODs/p7-GXvM3eXg** (long, 04:18:43)

- 00:23:08 Initial frustration and request to stop
- 00:23:11 Discussion of RP frogs
- 00:23:14 Request for an ad break
- 00:23:18 Realistic portrayal of money laundering in movies
- 00:24:01 Money laundering basics
- 00:25:13 Realistic portrayal of money laundering in 'Gotcha'
- 00:27:02 Drug money laundering through gems and emeralds
- 00:28:02 Realistic portrayal of drug dealer's cash stashes

**omnibased/S-9F_Jr3jmA** (long, 03:31:45)

- 00:00:48 Introduction and Opening Remarks
- 00:02:23 Discussion on the Situation in Minneapolis
- 00:05:16 Response to Questions about ICE Operations
- 00:08:17 Explanation of the Shooting Incident
- 00:14:44 Discussion on the Storm and Emergency Preparedness
- 00:19:33 Response to Questions about the Shooting Incident
- 00:25:45 Conclusion and Closing Remarks
- 00:29:02 Introduction and Setup

**destiny/WL9htojKbGY** (long, 03:57:39)

- 00:19:29 Introduction and Context
- 00:22:03 Discussion on Asmin Gold's Views
- 00:24:46 Criticism of the Video Content
- 00:27:58 Investigations and Prosecutions under Biden Administration
- 00:32:52 Fraud Allegations in Daycares and Charities
- 00:36:14 Investigation Outcome Uncertainty
- 00:37:21 Introduction and Context
- 00:37:58 Entertainment vs. Journalism Debate

**HasanAbiVODs/GjPX_ueTdfc** (verylong, 08:02:45)

- 00:35:24 Stream Snipers and Stream Saviors
- 00:36:13 Conservative Floridian Mom Lobster
- 00:37:50 Bondo Simulator 2021
- 00:39:05 Boomer of the Year
- 00:41:15 Compliments and Reactions
- 00:45:28 Chat Union Demands
- 00:49:12 Food Cravings
- 00:53:25 Sweat Dripping Down

**HasanAbiVODs3/bW2lKpHlmMQ** (verylong, 07:59:00)

- 00:07:55 Introduction and Podcast Update
- 00:09:20 Kyrie Irving's Performance in Game 7
- 00:12:31 LeBron James' Jab Step and Its Impact on the Game
- 00:16:24 Discussion on Kyrie Irving's Shot Selection
- 00:23:22 Podcast Episode Rankings and Pod Save America
- 00:31:41 Support for Raes Texas
- 00:35:56 TikTok Video with Twin Peaks Music
- 00:39:44 Thanksgiving and Business Continuation

**HasanAbiVODsbackup/5IufSgJSI0E** (verylong, 07:59:07)

- 00:30:09 Fear of Russia and NATO's Protection Racket
- 00:30:28 Mafia Analogy for NATO Membership
- 00:31:16 Finland's Consideration to Join NATO
- 00:32:21 Anti-NATO Stance and Protection Racket Argument
- 00:34:02 Putin's Actions in Ukraine
- 00:35:05 Criticism of Putin's Military Strategy
- 00:36:17 Humanitarian Impact and Social Media Response
- 00:37:32 NATO Expansion Concerns