# Proof Short AI Background Rendering + QA

Use this after a Proof Short has validated the real viewer problem and the user wants the video to look more premium without losing factual/control safety.

## Pattern

```text
AI image layer = text-free atmosphere/background/keyframes
Deterministic renderer = invoice, email, UI, payment route, labels, facts, captions, claims
```

This is especially useful for scam-awareness, finance, invoice, email, cybersecurity, health, or other high-trust topics where generated text or UI could become misleading.

## Generation scope

Generate only a small set of text-free keyframes, usually 2–3:

- provider/model: highest available reasonable GPT Image tier, e.g. `openai-codex` + `gpt-image-2-high`
- aspect: vertical 9:16 / portrait
- purpose: premium atmosphere only
- no readable text, logos, brands, QR/barcodes, people/hands, company names, bank data, account numbers, invoice IDs, addresses, watermarks, pseudo-writing
- ask for clean negative space for renderer-owned captions/UI overlays

Do not generate the invoice/email/payment-route itself. Those remain renderer-owned.

## QA and selection

Do not blindly use every generated keyframe. Create a quick contact sheet and inspect before render.

Reject/discard a keyframe if it contains:

- pseudo-readable document lines or fake text that might be interpreted as content
- brand-like marks/logos
- QR/barcodes
- people/hands if not explicitly desired
- bank/payment/account-like details
- overly busy composition that competes with captions or deterministic UI

It is acceptable, and often correct, to generate 3 and use only 2. Record discarded images and why.

## Rendering treatment

Keep AI backgrounds subordinate to the truth layer:

- cover-crop to 9:16
- add subtle deterministic push-in if needed
- blur very slightly if background details compete with foreground
- darken/desaturate or add translucent navy overlay so UI/captions remain readable
- place deterministic invoice/email/payment-route overlays above the background
- keep active-word captions and safe-zone bounds unchanged from the approved proof short

The upgrade should look more premium, but the first-second visible fail must remain obvious.

## Review package fields

Include:

- `visual_upgrade_mode: ai_background_plus_deterministic_mockups`
- AI provider/model/quality
- generated keyframe count
- used keyframe count
- discarded keyframe count + discard reasons
- per-image path/SHA256/resolution
- statement that all UI/invoice/payment-route/factual/caption layers are deterministic renderer-owned
- `wan_used: false`, `i2v_used: false`, `t2v_used: false` unless separately approved
- visual QA result
- AI disclosure note

## AI disclosure stance

If only text-free atmospheric backgrounds are AI-generated and there are no realistic synthetic people or real-event depictions, platform synthetic-media disclosure is often not mandatory. Still include a conservative note: if the platform UI asks whether AI-assisted media was used, disclose according to that prompt.

## Pitfall

Do not let AI polish reintroduce the old failure mode: pretty but semantically weak. The deterministic foreground must still answer in the first second: what is broken, where is the fail, what is at stake, and what rule fixes it.