# A transcript is not an edit list

Two timelines, explicit decisions and an actual neutral-video editing test.

L00P.AI · 1.0 · 2026-09-26

https://www.l00p.ai/en/resources/series/transkrypcja-nie-jest-lista-ciec/

## 01 / An idea needs boundaries

A transcript suggests where a speaker changes topic. A production plan proposes a caption, chart or quotation. Neither document alone determines which frames enter the finished film. Between an idea and an export, an edit decision list, or EDL, is needed.

Our starting point is an archived videocast plan: chapters, proposed excerpts, graphics cues and editorial questions. Different parts sometimes give conflicting statuses to the same information — confirmed in one place, awaiting checks elsewhere. We develop a working method without repeating unverified statements from the discussion.

## 02 / Four layers of work

The recording preserves what was captured. A transcript is an interpretation, with a version number and possible errors. A plan proposes excerpts and presentation. An EDL records approved operations on specific files. Correcting a name in a transcript must not silently replace an already accepted graphic.

For each decision, record the source, in and out boundaries, output position and reason. A graphic also needs its version, permitted use and content review. “Show a map” is a task, not evidence that the completed map contains correct data.

## 03 / A proposal does not approve itself

Use explicit states: proposed, needs checking, approved, rejected. A contradiction between notes holds that element for resolution. High model confidence cannot replace listening or the source of a number. Specify who resolves the issue and what evidence closes it.

A quotation needs fidelity to the speech and correct attribution. A chart needs units, a period and a source. A photograph needs verified provenance and permitted use. One “AI checked” label does not answer these different questions. Until required checks are closed, the element stays out of the broadcast version.

## 04 / Two timelines

After removing an excerpt, a second in the film no longer corresponds to the same second in the recording. Preserve a source-to-output map for captions and graphics too. A card planned using source time must reach the correct position after editing.

Our neutral example retains source seconds 1–3 and 4–5. The first excerpt occupies output seconds 0–2, the second 2–3. Interval ends are exclusive: a frame at the end boundary is not included in the selection.

![Source 1–3 s maps to output 0–2 s; source 4–5 s maps to output 2–3 s. 75 frames at 25 fps.](https://www.l00p.ai/wydawnictwo/transkrypcja-nie-jest-lista-ciec/dwie-osie-en-v1.svg)

Original map of the actual synthetic-video trial, Codex / L00P.AI. Interval ends exclusive. No audio.

At a constant 25 frames per second, the notation is [25,75) and [100,125). Together, 50 + 25 = 75 frames, or three seconds. This calculation applies to this constant-frame-rate trial, not universally to every video file.

## 05 / A decision template

Below is the demonstration structure, not a universal import format for editing applications. Before production, add the source hash, transcript version, reasons and editorial acceptance.

```json
{
  "source": "source.mkv",
  "fps": "25/1",
  "synthetic": true,
  "audio": false,
  "segments": [
    {"in_frame": 25, "out_frame_exclusive": 75, "output_start_frame": 0},
    {"in_frame": 100, "out_frame_exclusive": 125, "output_start_frame": 50}
  ]
}
```

Each overlay should state whether its timing refers to the source or output. Include a start and end, not just an appearance time. If its associated excerpt is removed, reconsider the card's meaning rather than automatically moving it to the nearest available position.

## 06 / An actual editing test

On 26 September 2026 we created six seconds of our own synthetic test image: 160 × 90 pixels, 25 frames per second, without sound. FFmpeg extracted and joined the specified ranges into an FFV1 file in a Matroska container. The source contained 150 frames and the output 75; the duration reading confirmed three seconds.

We compared every decoded output frame with its corresponding source frame in the same YUV420P pixel format. The data matched exactly. We retained the files, EDL, parameters and checksums. This tests that cutting and joining workflow, not equality for every codec.

We did not test speech recognition, quotation meaning or camera synchronization. The trial contains no human voice and does not establish that the archived plan is ready for broadcast. Editing a conversation still requires listening to boundaries, questions and answers.

## 07 / Cameras need separate evidence

Similar recording times on several files do not prove synchronization. A future trial needs a common observable reference and offset checks at the beginning, middle and end. A result for one point is insufficient to assess the entire recording.

Camera synchronization remains a development direction here. A UTC clock records the operation time; it does not itself prove that picture and sound agree. We claim no one-second accuracy without actual measurement. We do not add fictional audio times to this demonstration's frame map.

## 08 / Acceptance beyond export

Check sources, duration, order, boundaries, graphics completeness and readability on a phone. Record results and explicit gaps. The [article on accepting a render](/en/resources/series/dobry-render-to-jeszcze-nie-odebrany-material/) develops this checklist. To apply it to your work, start with a short excerpt and its edit plan.

The archived production plan was read. Assertions about people, numbers, rights and earlier model assessments were not adopted as established facts. Codex produced the PL/EN text, diagram and local test; there was no independent second-model review. Changes to the source, edit, codec or a discovered error require appropriate renewed acceptance checks.