# Clean AI Cut vs Hybrid Mechanism — session learning

Use this when comparing a pure premium AI-image sequence against a proof short with deterministic mechanism/UI overlays.

## Session-derived pattern

For a scam/cybersecurity proof short, a clean image sequence can look significantly more premium and native to TikTok/Reels, but it often weakens immediate mechanism clarity.

Observed comparison:

```text
Clean AI Cut = premium mood, calmer, less explainer-board
Hybrid = clearer mechanism, more immediately useful
Best likely final direction = mostly clean images + one minimal deterministic mechanism moment
```

Recommended framing:

```text
90% Clean AI Images
10% precise renderer-owned mechanism
```

Examples of acceptable minimal mechanism moments:

- one brief `suspicious link/domain` reveal
- one brief `close invite -> open trusted app/bookmark` action moment
- one simple before/after state, not a full checklist board

Avoid returning to:

- checklist-heavy UI cards
- boxed explainer overlays
- arrows/circles/diagram markers
- repeated labels and decision-target graphics

## Viewer explanation rule

When the viewer is not technical, do not use shorthand such as `open the service yourself` without explanation. Spell out the exact action:

```text
Close the invite.
Open your calendar or work app from the normal icon/bookmark.
If the meeting is real, it will still be there.
Trust the path you chose, not the button that chose you.
```

This converts a security maxim into a concrete next action.

## Caption synchronization rule

For final-near shorts where the user complains that subtitles do not match voiceover exactly, use actual word-boundary/forced-alignment timing rather than distributing words evenly across sentence SRT blocks. For EdgeTTS previews, request `boundary='WordBoundary'`, store word cues, then render static phrase segments where only the active spoken word highlights.

Keep the established caption policy:

- static phrase segment, usually 3–7 words
- active spoken word highlighted in yellow
- inactive words white
- no subtitle box/bar
- bottom mobile-safe position

## A/B decision criteria

After rendering both versions, explicitly answer:

1. Which version looks more premium?
2. Which version explains the mechanism faster without audio?
3. Does the clean cut rely too much on VO/subtitles?
4. Would one minimal deterministic mechanism moment recover clarity without killing the premium look?
5. Is Wan/I2V worth testing, or is image-sequence motion enough for this stage?

Default recommendation: do not jump to full Wan/I2V after this test. If clean images work visually, test exactly one short motion micro-proof only after approval.

## Visible-problem keyframe A/B pattern

When the user asks whether pure KI images can carry a Proof Short, do not generate generic premium atmosphere. The first image batch must make the real viewer problem literally visible without text:

- show the concrete object/situation (e.g. a calendar invite card on laptop/phone, a join/login button shape, attendee dots, blank time rows);
- keep all generated UI text-free: no letters, pseudo-letters, numbers, logos, QR/barcodes, watermarks, people, hands, or faces;
- use contact-sheet QA before rendering and discard frames that only look like generic laptop/desk B-roll;
- expect later safe/resolution frames to be less semantically explicit, but the first half must visually anchor the trap.

For the calendar-invite login-trap proof, the updated session result was:

```text
B1 = clean visible-invite AI images only + voiceover + active-word subtitles; premium but still a bit VO-dependent.
B2 = same images + subtle focus outlines; rejected because the lines added little value and reduced the premium look.
B3 = detailed AI-generated Outlook-style invite/login/calendar screens with readable UI text + voiceover + active-word subtitles; accepted as the new standard.
Outcome = B3 recommended/default: no renderer-added lines/arrows/boxes, but prompt the image model with very detailed, concrete, readable screen/UI content so the problem and safe action are visible in the image itself.
```

Use this as the default next-step bias: if clean images are beautiful but semantically weak, first make the AI keyframes themselves more specific and content-rich (visible app/email/invite/login/calendar screens, exact readable text, realistic UI details) rather than adding focus lines, arrows, circles, labels, or checklist cards. Deterministic renderer text remains appropriate for captions; visible screen text may be AI-generated when the user explicitly wants detailed realistic UI images and the keyframes pass visual QA.

Export QA pitfall: Telegram/mobile playback may misinterpret vertical files if aspect metadata is weak. For final/review MP4s, explicitly encode with `setsar=1,setdar=9/16`, `yuv420p`, `-profile:v high -level:v 4.2`, `-movflags +faststart`, and verify `ffprobe` reports `width=1080 height=1920 sample_aspect_ratio=1:1 display_aspect_ratio=9:16` before delivery.