# Readable silent-viewing caption baseline

Use this reference for TrueTraceShorts / AutoShorts proof shorts when the viewer must understand the video without sound.

## Session-derived lesson

The user rejected the first `erf-007` fake-support-popup cut because the text moved too quickly for silent viewing and some subtitle pages split sense units awkwardly. The accepted revision became the baseline:

- Do not optimize only for voice pacing. Optimize for a viewer reading on mute.
- It is acceptable, and preferred, to extend runtime when readability requires it.
- Keep caption pages as complete sense blocks: sentence-complete or clause-complete.
- Avoid orphan-word pages entirely. Never leave a single leftover word on the next caption page.
- If a short tail would occur, rebalance adjacent blocks, merge a short label with the next phrase, or extend timing.
- Prefer fewer, longer caption pages over rapid-fire word chunks when the content is explanatory.

## Accepted pattern from `erf-007` revision

- Previous rushed version: ~36s, 26 caption segments.
- Accepted readable version: ~46s, 15 caption segments.
- TTS generated sentence-by-sentence rather than as one continuous paragraph.
- Insert small inter-sentence pauses, around 650ms, before alignment/rendering.
- Keep active-word highlighting, but the *whole visible block* must be meaningful as text.
- Merge display artifacts such as `POP -UP` into readable words like `POP-UP`.

## Implementation notes

1. Draft script in short, complete sentences.
2. Synthesize voice sentence-by-sentence when normal TTS pacing is too fast.
3. Concatenate sentence chunks with short silence between them.
4. Align words after concatenation.
5. Build subtitle segments from source-script sentence boundaries, not only ASR punctuation.
6. Permit 8–11 word blocks when needed to keep an idea complete.
7. Split longer sentences at natural clause boundaries.
8. Extend video duration rather than forcing dense text into the old target.
9. QA real MP4 frames, not only contact sheets.
10. Inspect frames around the red flag and the safer-action sections specifically.

## QA checklist

- Can the video be understood on mute?
- Does each caption page express a complete or useful idea?
- Are there any one-word or two-word orphan pages? If yes, rerender.
- Are captions readable in the lower mobile safe zone?
- Is any caption too crowded or clipped?
- Does active highlighting help rather than distract?
- Did the final duration remain as short as reasonable, but not shorter than useful?

## Title/description/upload impact

This is a render-quality rule, not an upload rule. It does not change the private-upload approval workflow, hash binding, no-thumbnail default, or upload-only OAuth scope.