Question 1

What does the forbidden phrase library actually deliver, and where does it sit in the brand-consistency workflow?

Accepted Answer

An orchestration layer that sits above the operator brand-consistency + LLM-observability + AI-content-generation + trademark-monitoring + legal-research + watermarking + DAM + policy-as-code + WORM-storage stack and maintains the canonical list of phrases the operator brand cannot use anywhere in AI-generated content. The skill is a three-skill bundle on the brand-spec-authoring agent. Skill 1 — Curate: maintain the per-brand per-phrase library across 25+ phrase categories including competitor trademarks + competitor taglines + competitor product names + competitor mascots + celebrity likenesses (the right-of-publicity surface) + regulated claims that require substantiation + superlative claims that lack basis + guarantee claims that expose the operator + free-claim language that triggers FTC scrutiny + comparative claims that lack support + health/efficacy/safety claims regulated by FDA + environmental/organic/fair-trade claims regulated by FTC Green Guides + Made-in-USA claims regulated under the 2024 Labeling Rule + off-brand tone/register/POV/formality patterns the brand-team rejects + deprecated brand phrases (old positioning that has been retired) + retired brand phrases + legal-hold phrases (under active litigation) + compliance-hold phrases (under active regulatory inquiry). Curation sources include operator counsel, the operator brand-team, regulatory agency notices, industry-association guidance, Completions-curated additions, and operator-marketing-team contributions. Curation cadence runs quarterly plus on-update plus on-incident plus on-competitor-launch plus on-regulatory-action triggers. Skill 2 — Propagate: when a phrase enters or leaves the library, the change fans out to every downstream subscriber agent that needs to enforce the new state. Subscribers include the brand-voice-gate sibling skill (which enforces the library on every AI-generated artifact pre-publish), the content-generator, the copy-hook-variant-generation agent, the local-SEM agent, the social-content-orchestration agent, the communication-broadcast agent, the page-generator, the creative-asset-generator, the compliance-overlay-manager, and the governance-oversight agent. Propagation routing applies operator-counsel-approved policy with per-subscriber acknowledgment tracking, per-subscriber SLA tracking, and dead-letter-queue handling for subscribers that fail to acknowledge within the operator-counsel-set window. Skill 3 — Block: pre-publish blocking is the enforcement step. Block triggers run as a layered match cascade — exact match (verbatim phrase), fuzzy match (Levenshtein-tolerant variations), regex match (operator-counsel-authored patterns), semantic match (LLM-as-judge against the phrase intent), taxonomy match (the phrase falls in a category the operator-counsel marked auto-block). Block severity routes to warn (logged with attestor + can pass through), reject (refuses to commit), escalate to counsel, escalate to brand team, escalate to officer per operator policy. Block remediation suggests substitutions from the operator-counsel-approved alternative library, suggests rewrites via LLM, requires attestation from a human approver, or requires the addition of a disclosure clause. Every Curate, Propagate, and Block decision writes to the WORM audit trail with rule_id, policy_version, decision, attestor, timestamp, and evidence pointer. The brand-consistency + LLM-observability + AI-content + trademark-monitoring + legal-research vendors below ship strong primitives. The orchestration above them — phrase taxonomy curation, cross-skill propagation, layered match cascade blocking, compliance gate, audit trail — is operator-side architecture.

Question 2

Where does single-vendor forbidden-word filtering stop compounding for multi-brand AI-content operators?

Accepted Answer

Single-vendor forbidden-word filtering is solved. Acrolinx ships strong brand-term lists with category-tagged violations. Writer ships strong brand-language enforcement with style + tone violations. Grammarly Business ships strong brand-tone + brand-term enforcement. The compound case the brand-spec-authoring agent has to handle is the one where a multi-brand operator runs 1-50 brand portfolios, each brand has its own forbidden-phrase library (with overlap and divergence), AI content generates at scale across the content-generator + copy-hook-variant-generation + creative-asset-generator + page-generator + local-SEM + social-content-orchestration + communication-broadcast surfaces, the per-vertical compliance overlay (covered in the per-vertical compliance overlay sibling skill) determines which claims need substantiation pointers attached, the per-platform editorial policy library (covered in the brand-voice-gate sibling skill) determines what is auto-blocked per platform, the trademark-monitoring vendors (BrandVerity, Markify, Corsearch, TrademarkNow, CompuMark, Tradesparq) surface new competitor trademark filings the operator counsel needs to add to the forbidden library, the right-of-publicity surface (state-by-state including Tennessee ELVIS Act, California Civil Code 3344, post-mortem rights extensions) determines which celebrity likenesses cannot be used without authorization, and the operator counsel maintains versioned policy that has to propagate to every downstream subscriber agent with acknowledgment tracking. Without an orchestration layer above the brand-consistency + LLM-observability + AI-content + trademark-monitoring vendors, the forbidden library lives in one vendor’s console, propagation to downstream subscribers is manual, the audit trail of "who tried to publish what forbidden phrase, when, against what library version" fragments across vendor logs, and the compliance gate fires inconsistently across surfaces. The orchestration above the vendors is what holds the cross-brand + cross-skill + cross-vertical + cross-jurisdiction invariants.

Question 3

How does the layered match cascade work in Skill 3 Block, and how does it balance false positives vs false negatives?

Accepted Answer

The layered match cascade runs in priority order. Exact match runs first because it is the fastest and lowest-false-positive layer — verbatim phrase comparison against the active library. Fuzzy match runs second with operator-counsel-set Levenshtein tolerance (typically 1-2 character substitutions, depending on the phrase length and the operator-counsel-approved sensitivity). Regex match runs third against operator-counsel-authored patterns for phrase families (e.g., "best [vertical] in [city]" is a superlative-claim pattern; "guaranteed [outcome]" is a guarantee-claim pattern). Semantic match runs fourth via an operator-chosen LLM-as-judge (Claude, GPT, Gemini, Mistral — operator chooses) against the phrase intent — this layer catches paraphrases the regex layer missed, but at higher latency and higher false-positive risk. Taxonomy match runs fifth against the operator-counsel-approved category-blocked taxonomy (any phrase the LLM-as-judge classifies into a category the operator-counsel marked auto-block routes to the block decision regardless of whether the phrase appears in the library). The cascade is designed so high-confidence layers fire first and high-coverage layers backstop them. Operator counsel + the operator brand-team set per-layer sensitivity tuning quarterly against the operator’s realized false-positive (block what should have passed) + false-negative (let through what should have blocked) rates. The cascade results route through the block severity layer — warn (logged + attestor + can pass through with attestor sign-off), reject (refuses to commit, surfaces remediation suggestions), escalate to counsel (legal-hold or compliance-hold or active-litigation phrase categories), escalate to brand team (off-brand tone/register/POV/formality categories), escalate to officer (operator-counsel-defined high-stakes categories). The block remediation layer suggests substitutions from the operator-counsel-approved alternative library (a typical pattern: "best dentist" → "skilled dentist" or "experienced dentist" depending on the state licensing-board allowed claim), suggests rewrites via LLM (with operator-counsel-set safety bounds), requires attestation when the operator-counsel-approved disclosure path applies, or requires the addition of a disclosure clause (e.g., when a substantiation-required claim is detected, add the substantiation source).

Question 4

How does the cross-skill propagation in Skill 2 work, and how does it integrate with the customer-change-event-emission sibling skill?

Accepted Answer

When operator counsel adds, removes, or modifies a phrase in the forbidden library, the change becomes a propagation event. The Skill 2 Propagate stage routes the event across the downstream-subscriber agents (brand-voice-gate, content-generator, copy-hook-variant-generation, local-SEM, social-content-orchestration, communication-broadcast, page-generator, creative-asset-generator, compliance-overlay-manager, governance-oversight) through the operator-chosen event broker (Apache Kafka, AWS Kinesis, Google Pub/Sub, Azure Event Hubs, AWS EventBridge — the same broker the customer-change-event-emission sibling skill on the customer-graph agent uses). Each subscriber acknowledges receipt; the propagation layer tracks per-subscriber acknowledgment + per-subscriber SLA. Subscribers that fail to acknowledge within the operator-counsel-set window route to the dead-letter queue for triage. The operator-counsel-set SLA differs per subscriber: high-velocity surfaces (content-generator, copy-hook-variant-generation) need sub-minute propagation so AI-generated content stops emitting the deprecated phrase fast; lower-velocity surfaces (page-generator, creative-asset-generator) get longer windows because they batch produce; legal-hold and compliance-hold phrases get fastest propagation because they typically respond to active legal exposure. Propagation events log to the WORM audit trail with phrase_id + change_type + subscriber_set + per-subscriber-ack + propagation_timestamp + policy_version. When operator counsel needs to demonstrate, in a TCPA, FTC, Lanham Act, or state-AG investigation, that a specific phrase was added to the forbidden library on a specific date and propagated to all downstream subscribers within an operator-defined SLA, the audit trail returns the exact answer with cryptographic chain-of-custody.

Question 5

What compliance does the per-event gate enforce, and how does it map to Lanham Act, FTC Section 5 + Endorsement Guides + Made-in-USA + substantiation, per-vertical claims allowlist, right of publicity, and per-platform editorial policy?

Accepted Answer

Five anchors. Anchor 1: Lanham Act (15 USC 1114 trademark infringement + 1125 false advertising and trademark dilution and tarnishment) + Federal Trademark Dilution Act + state trademark dilution laws. Forbidden phrases that include competitor trademarks, taglines, product names, slogans, or mascots are gated against Lanham Act exposure. Comparative claims that reference competitor marks must be substantiated. Trademark dilution by blurring (using a famous mark in a non-comparative way that weakens its distinctiveness) and dilution by tarnishment (using a famous mark in a way that harms its reputation) are separate exposures the gate enforces. The operator counsel maintains the per-state trademark dilution policy library. Anchor 2: FTC Section 5 + FTC Endorsement Guides (16 CFR Part 255 with 2024 AI-disclosure amendments) + FTC Made-in-USA Labeling Rule (16 CFR Part 323) + claims-substantiation doctrine (Pfizer 1972 Reasonable-Basis). Forbidden phrases include superlative claims that lack substantiation (best, top, leading, #1, world-class, premier), guarantee claims that the operator cannot back, free-claim language that misrepresents qualifying conditions, comparative claims that lack support, health/efficacy/safety claims that lack the required substantiation level, environmental + organic + fair-trade claims that lack FTC Green Guides compliance, and Made-in-USA claims that lack the 2024 Labeling Rule substantiation. The gate composes with the per-vertical compliance overlay sibling skill and refuses to commit a phrase that fails the substantiation pre-check. Anchor 3: Per-vertical claims allowlist + state licensing-board advertising rules + per-vertical specialized regulators. For healthcare-vertical operators, forbidden phrases include unsubstantiated health-outcome claims + FDA-restricted disease-prevention language + state medical/dental/optometry/pharmacy/PT licensing-board prohibited superlative language. For legal-services operators, per-state bar advertising rules govern which phrases are forbidden (most state bars prohibit "best lawyer", "winner of [award]" without specific qualifications, "guaranteed results" entirely). For financial-services operators, FINRA Rule 2210 prohibits certain performance-claim phrases without specific substantiation. For prescription operators, FDA OPDP rules govern forbidden phrase categories. For  operators, per--regulator rules govern (e.g., several states prohibit any health-benefit claim). For alcohol operators, DISCUS Code of Responsible Practices + state liquor-board rules. For tobacco operators, FDA CTP rules. Anchor 4: Right of publicity (state-by-state including New York Civil Rights Law Sections 50-51 + California Civil Code 3344 living-individual rights + California Civil Code 3344.1 post-mortem rights 70-year extension + Tennessee ELVIS Act (2024) explicitly criminalizing unauthorized AI-generated voice or likeness uses + Indiana + Nevada + Texas + Florida + Pennsylvania + Illinois + Ohio + other state regimes). Forbidden phrases include celebrity names + recognizable celebrity descriptions + celebrity nicknames + celebrity catchphrases that the operator has not licensed. SAG-AFTRA AI provisions add union-performer informed-consent obligations for AI-generated likeness use. The per-state right-of-publicity policy library is operator-counsel-maintained (the same library the generative creative production sibling skill reads from). Anchor 5: Per-platform editorial policy + per-platform trademark complaint systems. Each major platform (Meta, TikTok, LinkedIn, YouTube, Google, X, Pinterest, Snapchat, Reddit) maintains a prohibited-content list (hate speech, medical misinformation, election misinformation, certain financial-claim categories, etc) and a trademark complaint system (Meta Brand Rights Protection, TikTok IP Reporting, LinkedIn IP Reporting, YouTube Content ID + IP complaint, Google Trademark Policy, X IP Reporting). The per-platform forbidden-phrase overlay enforces the platform-specific bans before content reaches the platform. Broader gate also enforced: COPPA + California AADC + Connecticut SB 3 + DSA Article 28 child protection + ADA Title III + WCAG 2.2 AA + NIST AI RMF + ISO 42001 via policy-as-code (OPA Rego + AWS Cedar + Casbin + Cerbos + Oso). WORM audit trail (AWS S3 Object Lock + GCS retention + Azure Blob immutable + Snowflake Time Travel) with per-statute retention (FTC 7yr + Lanham trademark-statute-of-limitations variable + state right-of-publicity variable + per-platform variable + IRS 7yr + state variable) per operator counsel policy.

Question 6

What does the engagement look like across Tier 1 → Tier 2 → Tier 3, and what does the Tier 3 reporting cycle commit to?

Accepted Answer

Tier 1 AI Readiness Assessment (2-3 weeks, diagnostic): audits the operator current forbidden-phrase posture against the 3-skill bundle + 5-anchor compliance gate; deliverable is a gap-pack report identifying which phrase categories are uncovered, which downstream subscribers lack the propagation receiver, which match-cascade layers are absent, which per-state right-of-publicity policies are unrepresented in the library, which per-vertical claims-allowlist categories need population, which per-platform editorial policies need codification, and a recommended remediation sequence for Tier 2. Tier 2 AI Swarm Setup Sprint (4-8 weeks): builds the 3-skill bundle on the brand-spec-authoring agent, wires brand-consistency vendors (Acrolinx + Writer + Grammarly Business + IBM Watson Tone Analyzer — operator chooses), LLM observability (LangSmith + Langfuse + Galileo + Patronus AI + Helicone + Braintrust + Humanloop — operator chooses), AI content vendors (Jasper + Copy.ai + AdCreative.ai + Anyword + Pencil + Writesonic + Persado + Phrasee + Hyperwrite — operator chooses), trademark monitoring (BrandVerity + Markify + Corsearch + TrademarkNow + CompuMark + Tradesparq — operator chooses), legal research (Westlaw + Lexis+ + Bloomberg Law + Practical Law + Compliance.ai + LawGeex — operator chooses), C2PA Content Credentials integration, DAM (Frontify + Bynder + Brandfolder + Aprimo + Lytho + Templafy — operator chooses), event broker for cross-skill propagation, configures the operator-counsel-approved match-cascade thresholds, configures the operator-counsel-approved block severity routing, configures the operator-counsel-approved remediation suggestion library, wires policy-as-code + WORM-storage, runs 30-day shadow + canary period before flipping to enforce-mode. Tier 3 Fractional CMO with AI Swarm (6-month minimum, 1-2 days/wk embedded): continues operating with weekly Curate cadence reviews, monthly per-subscriber propagation SLA audits, quarterly match-cascade calibration against operator-realized false-positive + false-negative rates, quarterly per-state right-of-publicity policy refresh, quarterly per-vertical claims-allowlist refresh, quarterly per-platform editorial policy refresh, quarterly compliance evidence packages. Tier 3 reporting is a 6-workstream pre-engagement-baseline reporting cycle (per-category curation coverage trend + per-subscriber propagation acknowledgment trend + per-layer match-cascade calibration trend + per-severity block routing trend + per-state right-of-publicity policy library currency + WORM audit-trail completeness) measured against the operator’s pre-engagement baseline. Each workstream surfaces trend direction and the gap to operator-defined targets. Reporting carries explicit caveats: brand-consistency + LLM-observability + AI-content + trademark-monitoring + legal-research vendor SLA + per-platform editorial policy updates (Meta + TikTok + LinkedIn + YouTube + Google + X + Pinterest + Snapchat + Reddit update on own cadence) + Lanham Act case-law developments + FTC Endorsement Guides amendments + per-state right-of-publicity statute changes + Tennessee ELVIS Act implementing guidance + per--regulator + state liquor-board + state medical/dental/pharmacy board rule changes + SAG-AFTRA agreement amendments sit outside Completions control. Attorney-client privilege preservation across forbidden-phrase library + match-cascade thresholds + per-state right-of-publicity policy + per-vertical claims-allowlist + per-platform editorial policy library is maintained per operator counsel policy.

Question 7

Who owns the forbidden-phrase library, the match-cascade thresholds, the propagation routing, and the audit trail?

Accepted Answer

Operator owns every artifact. The forbidden-phrase library lives in operator code repo, operator-counsel + operator-brand-team-maintained with versioned signed updates. The per-category phrase taxonomy lives in operator code repo. The per-layer match-cascade thresholds (exact + fuzzy + regex + semantic + taxonomy) live in operator code repo, operator-counsel + operator-data-science-team-tuned against operator-realized FP/FN rates. The per-severity block routing rules + per-remediation suggestion library + per-state right-of-publicity policy library + per-vertical claims-allowlist library + per-platform editorial policy library all live in operator code repo, operator-counsel-maintained. The brand-consistency, LLM-observability, AI-content, trademark-monitoring, legal-research, and DAM vendor subscriptions all run under operator billing on operator-controlled accounts. The event broker (Apache Kafka, AWS Kinesis, Google Pub/Sub, Azure Event Hubs, AWS EventBridge — operator chooses) runs under operator cloud. The WORM audit trail lives on operator-controlled cloud storage. The policy-as-code policies (OPA Rego, AWS Cedar, Casbin, Cerbos, Oso) live in operator code repo, counsel-aligned. The Lanham Act + FTC + per-vertical regulatory + per-state right-of-publicity + per-platform editorial policy compliance evidence records are operator-counsel-maintained. Completions owns the orchestration knowledge — how to design the phrase taxonomy across the 25+ categories for the operator’s actual vertical + jurisdiction + brand-portfolio mix, how to tune the match-cascade thresholds against operator counsel’s FP/FN tolerance, how to wire propagation routing for the operator’s actual subscriber-agent inventory, how to compose Block severity routing with the per-vertical compliance overlay + per-platform editorial policy + per-state right-of-publicity policy, how to design substitution + rewrite remediation that satisfies operator counsel substantiation requirements — and that knowledge transfers under the Tier 3 transition path (30-60 days at engagement end with full hand-off of the library taxonomy, the match-cascade thresholds, the propagation routing, the block severity routing, the remediation library, and the compliance evidence-package generation playbook). Completions credentials revoke on engagement-end.

Forbidden phrase library for multi-brand AI-generated content operations — Curate + Propagate + Block 3-skill bundle with a layered match cascade and a 5-anchor compliance gate

The real ecosystem this sits above

Brand consistency + LLM observability

AI content generation

Trademark monitoring

Legal research + DAM

Watermarking + event broker

Policy-as-code + WORM storage + GRC

Frequently asked