brand-safety

name	brand-safety
description	Brand safety and regulated-claims rulebook — toxicity/bias categories, FDA/FTC/FCA/GDPR rules, prohibited-content registry, brand-guideline enforcement, HITL escalation matrix, verdict schema.
user-invocable	true
allowed-tools	["Read","Glob","Grep","Write"]

Brand Safety Skill

Owned by brand-safety-compliance (gatekeeper). Applied as the brand-consistency and regulated-claims-review gates on every outbound asset. Output is a structured verdict (pass / fail / hitl) with evidence; never a free-text opinion.

1. Toxicity & Bias Categories

Category	Examples	Verdict default
Hate / discrimination	Slurs, dehumanizing language, protected-class attacks	fail
Harassment	Targeted attacks, doxxing, threats	fail
Sexual content	Explicit, minors-adjacent	fail (minors-adjacent: hard-fail + incident)
Self-harm	Glorification, instruction, contagion-risk framing	fail
Violence / graphic	Gore, weapons promotion, terrorism	fail
Illegal content	Drug sales, fraud, IP theft instructions	fail
Stereotyping (soft)	Tokenizing imagery, lazy archetypes	hitl
Cultural insensitivity	Sacred symbols misused, appropriation	hitl
Political polarization	Partisan stance not aligned with brand	hitl
Misleading / manipulative	Dark patterns, false urgency, fake scarcity	fail

Use a screening model (e.g. Perspective API equivalent) for first-pass scoring; reserve human reasoning for HITL cases.

2. Regulated-Claims Rulebook

FDA (US drugs, devices, supplements, cosmetics, food)

• Drug claims require approved indication, dosing, fair balance with risks
• Disease claims on supplements/food are prohibited; structure-function claims allowed with disclaimer
• Devices: classify Class I/II/III; Class II+ requires substantiation
• Cosmetics: cannot make drug claims (e.g. "treats acne" is drug-level)
• Required: "Important Safety Information" (ISI), fair balance, IFU (instructions for use) reference, ISI prominence parity with benefit claims
• Forbidden: cure / heal / treat / prevent for non-approved indications

FTC (US advertising)

• Substantiation: every objective claim must have prior reasonable basis
• Endorsements: disclosed material connection, typical-result framing or clear disclaimer
• "Made in USA": all-or-virtually-all standard; partial-content claims must qualify
• Health claims: competent and reliable scientific evidence
• Comparative: must be truthful and substantiated; can name competitors
• Influencer: #ad, #sponsored clearly visible; not buried in hashtags
• Native ads: clear "Advertisement" or "Sponsored" labeling

FCA / SEC / FINRA (Financial promotions)

• Past performance disclaimer: "past performance is not indicative of future results"
• Risk disclosure proportional to benefit claims
• ESG / sustainable claims: comply with SEC Climate Disclosure rules, EU SFDR Article 8/9 if applicable
• No guarantees of return; no projections without backtesting methodology disclosure
• Approved person: financial promotions in UK require FCA-authorized signoff
• Crypto / high-risk: jurisdiction-specific cooling-off, suitability assessment

GDPR / CCPA / Privacy

• Consent: explicit opt-in for marketing in GDPR jurisdictions; opt-out + Do-Not-Sell in CCPA
• Cookie / tracker: consent banner, granular categories, evidence of consent stored
• Retention: declared retention period, deletion process documented
• DSAR: data-subject-access-request fulfilled within 30 days (GDPR) / 45 days (CCPA)
• Marketing data sharing: declared in privacy policy, opt-out path

Other regimes (jurisdiction-specific)

Regime	Scope
ASA (UK)	Truthful, decent, legal, honest
ACMA / TGA (AU)	Therapeutic-goods advertising
LGPD (Brazil)	Privacy parallel to GDPR
HIPAA (US health)	PHI cannot appear in marketing without authorization
COPPA (US, <13)	No targeted advertising to under-13
PIPL (China)	Cross-border data restrictions

3. Prohibited-Content Registry Template

Maintained at progress/prohibited-content.yaml. Updated by brand-safety-compliance, audited quarterly.

prohibited:
  global:
    - "competitor disparagement by name"
    - "any reference to ongoing litigation"
    - "internal codenames (Project Phoenix, etc.)"
  per_profile:
    regulated:
      - "drug-treats claims for non-approved indications"
      - "patient testimonials without IRB-equivalent consent"
    dtc:
      - "before/after weight-loss without TGA-style disclaimer"
    b2b-saas:
      - "uptime / SLA percentages above contracted SLA"

4. Brand-Guideline Enforcement Checklist

Applied per asset:

• Logo: correct lockup, clear-space, color version
• Color palette: only approved hex codes from aesthetic-archetypes
• Typography: only declared families and weights
• Voice: matches archetype tone (see brand-narrative voice card)
• Imagery: matches archetype texture / motion / palette
• Spelling of product / brand / customer names (canonical list)
• Legal entity name, jurisdiction abbreviations, registered marks (R, TM)
• Mandatory disclosures for the channel and jurisdiction
• Approved hashtag set used; banned hashtags absent
• Required CTAs (if regulated, e.g. "Talk to your doctor")

5. HITL Escalation Matrix

Rows: severity. Columns: industry profile category.

Severity	Standard profiles	Regulated profile
Critical (hard fail)	Block + incident report	Block + incident + legal notification
High	Block + HITL (CMO + brand lead)	Block + HITL (CMO + legal + MLR)
Medium	HITL (brand lead)	HITL (CMO + legal)
Low	Auto-flag, pass with note	HITL (compliance)
Info	Pass	Pass with audit log entry

MLR = Medical, Legal, Regulatory review board (pharma/healthcare).

6. Verdict Schema

Every gate run emits this structure (written to progress/events.jsonl and the DecisionRecord):

verdict:
  gate_id: "brand-consistency"            # or regulated-claims-review
  artifact_ref: "memref://output/campaigns/c-2026Q3/assets/hero-v3.png"
  profile: "regulated-health-finance"
  result: "fail"                          # pass | fail | hitl
  severity: "high"                        # critical | high | medium | low | info
  findings:
    - rule_id: "fda-drug-claim-disease"
      category: "regulated-claim"
      evidence: "Caption: 'cures migraine in 24h' — non-approved indication"
      proposed_fix: "Replace with: 'may help reduce migraine frequency, see ISI'"
  required_approvers:
    - "cmo"
    - "compliance-officer"
  timestamp: "2026-05-19T14:22:11Z"
  agent: "brand-safety-compliance"

7. Standard Operating Procedure

1. Load prohibited-content.yaml and active profile rules.
2. Run automated screening (toxicity, regex on prohibited claims, schema validation).
3. Run brand-guideline checklist (visual + voice).
4. Run regulated-rules pass per active regime.
5. Aggregate findings; compute severity.
6. Emit verdict per schema; if fail or hitl, emit HITLRequest envelope.
7. Append entry to progress/events.jsonl (append-only).
8. If pass, return signed verdict to calling squad.

8. Anti-patterns

• Pass-with-conditions but conditions not enforced downstream
• Silent edits to the prohibited-content registry without a DecisionRecord
• Hard-fail without proposed fix (always include remediation guidance)
• HITL escalation without required_approvers list (caller cannot route)
• Re-running gate after asset hash changes without re-verdict

References

• marketing-governance (HITL gate catalog)
• creative-brief-protocol (Guardrails section consumed here)
• AGENTS.md sec 7 (gate definitions and HITL rules)