Recorded pipeline run · be2450c6ec72

One ad, end to end.

Every value on this page was written by the pipeline while it ran — not transcribed afterwards. The stages that did not run are here too, at the same size as the ones that did, because a record you can only read one way is not a record.

Did a model run?

No. This capture is from a host with no model stack, so the encoder stage did not run and no score on this page came from a neural pass.

What it does show is every stage that surrounds the model — the file as it arrived, the shot boundaries, the re-cuts encoded for real by ffmpeg, and the publish path — with the timings and hashes they actually produced. The encoder needs a provisioned CUDA box and a human-accepted CC BY-NC 4.0 licence on facebook/tribev2; when that run happens, this page fills in from its record without a line of it being rewritten.

The box this ran on

started
2026-08-06 22:15:24 UTC
wall clock
2.29s
platform
macOS-26.5.2-arm64-arm-64bit-Mach-O
python
3.13.13
commit
07da935
ffmpeg
ffmpeg version 8.1.2 Copyright (c) 2000-2026 the FFmpeg developers
compute
no model stack on this host
TRIBE v2 weights
not on this host
tesseract (OCR)
absent — clarity measures nothing without it
faster-whisper (ASR)
1.2.1

Recorded by observation, not by flag. The compute line comes from asking torch what device it has, so no argument to the pipeline can make this page claim a GPU that was not there.

The run, stage by stage

4 ran · 3 did not · 0 failed

  1. 01

    The file that came in

    ran71ms

    The file as it arrived, hashed so every later claim is about this exact cut.

    filename
    ad_01.mp4
    durationS
    28.714
    width
    360
    height
    450
    bytes
    1301006
    sha256
    33015cdd95535a30148d027d727e42f769f9f1c2705012e5ab139d73f4316c00
    artifactsizesha256
    ingest-runs/f8fd93b2f0bd/videos/ad_01.mp4the partner's cut1.2 MB33015cdd95535a30
  2. 02

    Where the cuts already are

    ran88ms

    ffmpeg scene detection. Everything downstream is defined in terms of these boundaries.

    detector
    ffmpeg select=gt(scene,0.28)
    threshold
    0.28
    shots
    11
    shot01
    0.00–1.70s
    shot02
    1.70–4.93s
    shot03
    4.93–6.63s
    shot04
    6.63–12.37s
    shot05
    12.37–14.00s
    shot06
    14.00–19.73s
    shot07
    19.73–22.10s
    shot08
    22.10–23.13s
    shot09
    23.13–24.63s
    shot10
    24.63–26.87s
    shot11
    26.87–28.71s
  3. 03

    The re-cut ladder, encoded for real

    ran1.63s

    Each candidate re-cut is really encoded — not simulated — so its score is measured.

    candidates
    3
    strategy
    even spread of single-shot removals, plus a head trim
    ad_02
    Drop shot 1 (0.0–1.7s) -> 27.08s
    ad_03
    Drop shot 6 (14.0–19.7s) -> 23.05s
    ad_04
    Drop shot 11 (26.9–28.7s) -> 26.93s
    artifactsizesha256
    ingest-runs/be2450c6ec72/videos/ad_02.mp4Drop shot 1 (0.0–1.7s)1.8 MB233bd02a0e863781
    ingest-runs/be2450c6ec72/videos/ad_03.mp4Drop shot 6 (14.0–19.7s)1.5 MBbe5b5515b18f65ad
    ingest-runs/be2450c6ec72/videos/ad_04.mp4Drop shot 11 (26.9–28.7s)1.8 MB2709c6d7cb999caa
  4. 04

    TRIBE v2 forward pass

    did not run0ms

    The TRIBE v2 forward pass. The GPU step, and the only one costed in minutes.

    torch is not installed on this host, so there is no model to run (this is a laptop; the encoder needs a provisioned CUDA box — tools/concierge/provision.sh)

    ads
    4
    device
    none
    weights
    facebook/tribev2 (CC BY-NC 4.0)
  5. 05

    Arcs collapsed to a read-out

    did not run0ms

    Arcs collapse to hook / processing / clarity, then to a single read-out.

    no arcs were produced, so there is nothing to score

  6. 06

    One frame per cut, for the grid

    ran104ms

    One frame per cut, for the library grid.

    artifactsizesha256
    ingest-runs/be2450c6ec72/ad_01.jpgad_0111.5 KBbfd4156a31dcc156
    ingest-runs/be2450c6ec72/ad_02.jpgad_0211.2 KBafe886a092200139
    ingest-runs/be2450c6ec72/ad_03.jpgad_0311.4 KB9240f4770ff700a5
    ingest-runs/be2450c6ec72/ad_04.jpgad_0411.4 KB9240f4770ff700a5
  7. 07

    Storage objects, rows, and the customer's URL

    did not run0ms

    Storage objects, database rows, and the URL the customer opens.

    --dry-run: nothing was written to Supabase

How to check this

The record this page renders is committed at captures/be2450c6ec72/run.jsonl, one JSON object per event. The gate re-folds it and fails if this page and that file disagree, so the two cannot drift apart quietly.

The artifact paths above are where those files sat on the machine that ran the pipeline; the footage itself is not in this repository, because it is large and it is not ours to redistribute. The hashes are, which is what makes them worth printing — they say which bytes were scored, and a file that does not hash to one of them is a different cut.

# reproduce the fold
.venv/bin/python tools/capture/summarize.py captures/be2450c6ec72/run.jsonl

# the gate's check
.venv/bin/python tools/capture/summarize.py --check

# the command that produced it (private flag values redacted at fold time)
scripts/ingest_partner_ad.py ingest-runs/f8fd93b2f0bd/videos/ad_01.mp4 brief.json --owner-email <redacted> --dry-run