brand-safety

name	brand-safety
description	IP clearance checklist and brand-guardrail rubric. Owned by governance-c2pa, consumed by Calliope and Helios.
user-invocable	false
argument-hint	<ip-check|guardrail-check|escalate>
allowed-tools	["Read","Grep"]

Brand Safety

The defensive layer. Every asset and every copy block passes through this skill before it can be marked approved.

Purpose

Prevent IP infringement, brand-voice violations, and reputational risk from leaving the studio. The skill produces a brand_safety_review artifact that gates publish.

When to use

• governance-c2pa runs this on every AssetJob output before signing.
• Calliope runs this at brief-intake when risk_tolerance == "low".
• Helios runs this on any third-party reference (mood-board, stock plate, talent photo) before incorporating it into a generation prompt.
• Erato/Terpsichore run this on copy that names a competitor or quotes a third party.

Inputs

• Asset file path OR copy block text
• brand_constraints from the active CreativeBrief
• Optional: source-of-record for third-party material (URL, license, talent release on file)

Outputs

• brand_safety_review.json with: verdict (pass|fail|hitl), ip_findings[], guardrail_findings[], recommendations[]
• Episode written to eights.memory.remember(domain="creative", scopes=["sensitive:ip"], tags=["governance"])

Sources of IP risk

1. Stock imagery / video — license tier (editorial vs commercial), model release on file, geographic restrictions, duration-of-use clause.
2. Brand likeness — third-party logos, trade dress, distinctive product silhouettes (a Coke bottle, a Birkin bag, an Apple device chassis).
3. Music samples — sync license, master + publishing both cleared, AI-generated tracks need provenance for training-data risk.
4. Talent rights — name/image/likeness, voice clones, deepfakes, deceased-celebrity rights of publicity.
5. Trademarks — competitor marks, slogans, registered taglines, sound marks (the NBC chime, the Intel bong).
6. Architectural / location — recognizable landmarks with trademark protection (Hollywood sign, Eiffel Tower at night).
7. Generative model provenance — was the base model trained on opted-out data? Does the workflow include ip-adapter referencing a copyrighted character?

IP clearance checklist

• Every external asset has a license file or URL on record
• License tier matches intended use (editorial / commercial / broadcast / paid)
• All identifiable humans have a model release OR are clearly de-identified
• No third-party logos, marks, or trade-dress in frame unless authorized
• No music cue used without sync + master clearance
• No AI-generated likeness of a real person without written consent
• No reference image fed to ip-adapter or controlnet is a copyrighted character or celebrity
• C2PA manifest will be attached on sign-off

Brand-guardrail rubric

Score each dimension 0-3 (0 = violation, 3 = exemplary). Asset passes only if all dimensions >= 2.

Dimension	0 (fail)	2 (acceptable)	3 (exemplary)
Voice match	Off-brand register	Matches voice rubric	Captures voice with distinctive flourish
Tone fit	Inappropriate for moment	Tone-appropriate	Tone-appropriate AND emotionally resonant
Taboo topics	Mentions a taboo topic	Avoids taboo list	Avoids AND proactively reframes risky adjacencies
Competitor mentions	Names competitor without authorization	No competitor mentions	No mention AND no implicit shade
Inclusivity	Excludes or stereotypes audience segments	Neutral / inclusive	Actively representative without tokenism
Factual accuracy	Contains false claim	Verifiable claims only	Verifiable AND cited
Legal claims	Makes regulated claim without disclaimer	No regulated claims	No claims AND has counsel-approved disclaimer where needed

Procedure

1. Load the asset or copy block and the active CreativeBrief.brand_constraints.
2. Run the IP clearance checklist. Any unchecked box -> verdict = hitl unless explicitly waived in brief.
3. Run the brand-guardrail rubric. Any dimension scoring 0 -> verdict = fail. Any scoring 1 -> verdict = hitl.
4. Run automated checks: regex sweep for taboo terms; OCR/transcript sweep for visible/audible competitor names; reverse-image search for stock provenance when license is missing.
5. Write brand_safety_review.json to RLM/output/governance/{asset_id}-{date}.json.
6. If verdict != pass, file a DecisionRecord with dissenting_opinions populated and trigger HITL.

When to escalate to HITL

• Any IP-clearance checkbox cannot be confirmed and risk_tolerance != "high"
• Any guardrail dimension scores 0 or 1
• Asset features a real person's likeness AND release-on-file status is unknown
• Copy makes a regulated claim (health, finance, safety) without counsel-approved disclaimer
• Conflicting constraints in the brief that the rubric cannot resolve
• ANY use of a deceased celebrity's name/voice/image

Failure modes / escalation

• Reference image is a copyrighted character — strip from prompt, regenerate, log episode tagged governance.
• License file missing — set verdict = hitl; do NOT publish; ping Calliope.
• Voice rubric score conflicts between governance-c2pa and Erato — DecisionRecord with both opinions; HITL adjudicates.

References

• Constitution: RLM/specs/creative-constitution.md
• C2PA signing: governance-c2pa sub-agent
• Memory tags: governance, sensitive:ip