PART 05 — PHASE 2
The Creative — a governed draft studio.
Generate review-ready options from approved facts and rights-cleared materials. Arabic is authored and evaluated in its intended market context; every asset remains a draft until named humans approve its claims, culture, rights, and accessibility.
50–70 Min
Model-agnostic
Human-reviewed creative
Reviewed 31 Aug 2026. Model behavior, platform creative policies, disclosure rules, and local advertising requirements change. Validate them for the destination, category, and country before launch. All prompts, assets, scores, and outcomes on this page are hypothetical.
Define the input boundary
Inputs the studio may receive
- Signed-off product facts, prices, availability, and delivery terms
- Approved claim substantiation and required disclosures
- Brand-owned or licensed copy, images, audio, fonts, and talent releases
- Market, language, dialect intent, placement, and accessibility needs
- Synthetic personas and scenarios for development and evaluation
Keep these out
- Raw customer messages, profiles, contact details, or purchase histories
- Unlicensed competitor assets or prompts to imitate a living creator
- Unverified health, beauty, financial, environmental, or religious claims
- Secrets, platform credentials, unpublished strategy, or tenant-mixed files
- Real people’s likeness or voice without documented permission
The creative specification
CAMPAIGN
- objective and destination placement
- country, audience context, language, requested dialect
- approved offer, product facts, exclusions, expiry and inventory state
RIGHTS
- asset_id, owner/licensor, permitted channels, territories and expiry
- talent/voice/likeness consent where relevant
- generation disclosure requirement for the destination
CLAIMS
- claim_id and substantiation reference for every factual promise
- prohibited phrases and regulated-category escalation owner
- price, tax, delivery, returns and promotion qualification text
ACCESSIBILITY
- logical reading order, alt-text brief, contrast-safe text zone
- captions plus transcript for audio/video; meaning must not depend on colour
- Arabic RTL layout and mixed Arabic/Latin/numeral rendering tests
OUTPUT
- concepts only; label every asset DRAFT · HUMAN REVIEW REQUIRED
- cite the approved inputs used; return MISSING_EVIDENCE instead of inventing a fact
Arabic is a review discipline, not a locale switch
Review layer
Question
Required reviewer
Meaning
Does the Arabic preserve the approved claim and qualification without exaggeration?
Bilingual copy reviewer
Dialect
Is the requested dialect natural for this audience, brand, and setting?
Native speaker from the target market
Culture
Could the imagery, humour, dress, family role, religion, or seasonal framing mislead or exclude?
Local cultural reviewer
RTL
Do reading order, punctuation, numerals, price, CTA, crop, and mirrored elements render correctly?
Arabic QA on target devices
Accessibility
Are captions, transcript, contrast, motion, alt text, and screen-reader order usable?
Accessibility QA
Ramadan, Eid, and national occasions are not automatic templates. Dates can be jurisdiction- and observation-dependent; audience behavior, business hours, product relevance, and acceptable tone vary by store and market.
A calendar may open a review task. It must not silently change messaging, generate religious claims, alter inventory plans, or publish themed assets. A local human approves the occasion, wording, visual treatment, dates, and rollout.
A safe review pipeline
1 · Sandbox
Generate with synthetic test inputs
Exercise malformed briefs, prompt injection, missing rights, unsupported claims, mixed scripts, and sensitive categories. No customer data and no publishing credentials.
2 · Automatic checks
Reject obvious policy and contract failures
Schema validation, source allowlist, duplicate detection, disclosure presence, banned-claim rules, asset licence expiry, safe-area and caption checks filter drafts. Passing is not approval.
3 · Specialist review
Route the draft to accountable people
Copy, legal/claims, rights, native-market cultural, and accessibility reviewers see the evidence and exact destination. Any rejection returns the draft to revision.
4 · Exact approval
Approve an immutable asset version
The approval binds asset hash, copy, destination, market, dates, offer, disclosures, reviewer IDs, and expiry. Editing any bound field invalidates approval.
5 · Controlled experiment
A separate operator launches within a preapproved test
Declare baseline, primary metric, unit economics, sample/decision rule, attribution window, cap, stop condition, and rollback. Generation volume or model score does not predict uplift.
Representative evaluation set
Markets
Country and dialect slices
Include KSA, UAE, Kuwait, Qatar, Bahrain, and Oman only where the business operates; evaluate local reviewers separately.
Formats
Destination rendering
Static, carousel, short video, caption, CTA, mixed-script price, small screen, low bandwidth, and reduced-motion behavior.
Risk
Claims and rights
Known-safe, ambiguous, prohibited, expired licence, unapproved likeness, counterfeit brief, and disclosure-required examples.
Language
Native quality
MSA, requested dialect, code-switching, spelling variants, numerals, gender/plural forms, humour, and literal-translation traps.
Accessibility
Usable creative
Contrast, captions, transcript, flash/motion, alt text, text size, RTL focus order, and no meaning encoded by colour alone.
Adversarial
Fail closed
Prompt injection in briefs/assets, hidden text, unsupported urgency, fake testimonials, missing terms, and attempts to publish without approval.
Draft: One synthetic skincare carousel for an invented brand, using a licensed abstract texture and approved non-medical product facts.
Review: The native-language reviewer revised the dialect; the claims reviewer removed an unsupported result; RTL QA found a reversed price; accessibility QA added a meaningful caption and contrast-safe treatment.
Outcome: The corrected version is eligible for an operator-designed test. No performance result is assumed.
Optimize the review system, not output count
There is no universal weekly volume or approval-rate target. Generate only what reviewers can examine carefully. Track critical-error escape rate, evidence coverage, reviewer agreement, time to correct, accessibility defects, and test quality.
Model choice and prompting should be re-evaluated on the representative set whenever the model, policy, brand, market, placement, or claim catalogue changes.