Question 1

What does done-for-you borderline routing actually deliver?

Accepted Answer

Completions builds and operates a borderline-routing meta-skill on the governance-router agent for multi-location, multi-vertical, multi-jurisdictional operators running 5,000-50,000 AI-generated marketing artifacts per month. Borderline classification: every output passes through a multi-model LLM-as-judge ensemble (OpenAI GPT-5, Anthropic Claude Opus 4.7, Google Gemini Ultra, Meta Llama-3.1-405B, Mistral Large, and Cohere Command R+) that returns a tier — confident-pass, borderline-pass, borderline-fail, or confident-fail — together with a threshold-band confidence interval, an explainability trace, a citation to the operator compliance rule library, and a trend annotation against the prior 7-day band distribution. Risk-tier routing: borderline outputs are routed across six risk tiers — catastrophic, serious, moderate, minor, negligible, false-alarm — into ten possible routing decisions (auto-approve, auto-reject, queue-for-human-review, queue-for-counsel-review, queue-for-brand-officer-review, queue-for-compliance-officer-review, queue-for-CEO-escalation, queue-for-board-escalation, queue-for-insurance-escalation, queue-for-regulator-disclosure) with per-tier SLA, per-tier escalation path, and per-tier audit-trail emission. Reviewer assignment: routing decisions land in the right reviewer queue (auto-review, per-vertical counsel, per-vertical brand officer, per-vertical compliance officer, per-vertical corporate marketing, per-vertical corporate operations, per-vertical corporate legal, per-vertical CEO, per-vertical board, per-vertical insurance, per-vertical regulator-disclosure) with workload balancing, fatigue decay, and expertise-routing applied so the right person sees the right output at the right SLA. Cross-skill feedback: reviewer decisions feed back into the upstream skills that produced the artifacts (master-record canonicalization, compliance-overlay-manager, GBP photo management, JSON-LD generation, GBP posting, franchisee content moderation queue, and borderline-routing itself) so the upstream LLM prompts, rule citations, and confidence thresholds tighten over time under operator-data-science-team supervision. The compliance overlay covers HIPAA, FDA OPDP, DEA, Metrc, DISCUS, FDA tobacco, state licensing boards, state attorneys general, FTC, Lanham Act, ADA, COPPA, state lemon law, Prop 65, WEEE/RoHS/REACH, per-jurisdiction breach-notification windows, and EU AI Act Article 22 transparency requirements where applicable. Operator owns the borderline registry inside operator data infrastructure (Snowflake, BigQuery, Databricks, or Redshift), the borderline-classification model code in the operator repo, the risk-tier routing configuration, the reviewer-assignment configuration, the cross-skill feedback model code, the LLM-as-judge prompt library, LLM API credentials under operator billing, the attorney-approved compliance overlay rule library, and the audit trail in operator-controlled WORM storage (AWS S3 Object Lock, Google Cloud Storage retention, Azure Blob immutable, or Snowflake Time Travel). Completions owns the orchestration knowledge.

Question 2

Why is borderline routing typically operator-side rather than LLM-vendor- or workflow-vendor-shipped?

Accepted Answer

Six engineering surfaces sit between operator data infrastructure and a working borderline-routing meta-skill, and they sit outside the design center of the LLM-as-judge, workflow, ticketing, and WORM-audit ecosystems that own the upstream and downstream primitives. Surface 1 — Multi-model LLM-as-judge orchestration: OpenAI, Anthropic, Google, Meta, Mistral, and Cohere each ship strong inference primitives, but running them as a coordinated ensemble with threshold-band calibration, explainability tracing, and rule-citation linkage to an operator compliance library is operator-side modeling. Surface 2 — Risk-tier routing: Temporal, Inngest, Trigger.dev, and Vercel Queues ship strong durable-workflow primitives, but the mapping from borderline tier to ten possible routing decisions across six risk tiers depends on operator counsel, brand, and operations policy. Surface 3 — Reviewer assignment: ServiceNow, Jira, Linear, and Pega ship strong queue and ticketing primitives, but workload balancing, fatigue decay, and expertise-routing across eleven reviewer roles per vertical is operator workforce-management modeling. Surface 4 — Cross-skill feedback to six upstream skills with Bayesian updating and counterfactual validation requires operator-data-science-team capacity. Surface 5 — WORM audit trail: AWS S3 Object Lock, Google Cloud Storage retention, Azure Blob immutable, and Snowflake Time Travel ship strong WORM primitives, but the audit-trail schema, retention windows, and evidentiary-quality discipline are operator-counsel-side. Surface 6 — Compliance overlay across HIPAA, FDA OPDP, DEA, Metrc, DISCUS, FDA tobacco, state licensing, state AG, FTC, Lanham, ADA, COPPA, state lemon law, Prop 65, WEEE/RoHS/REACH, per-jurisdiction breach-notification, and EU AI Act Article 22 transparency is operator-counsel-side rule maintenance. Completions runs orchestration across all six surfaces under one Tier 3 Fractional CMO with AI Swarm engagement; operator owns the artifacts and can in-house at any time.

Question 3

What does the engagement look like across Tier 1, Tier 2, and Tier 3?

Accepted Answer

Tier 1 AI Readiness Assessment (2-3 weeks, diagnostic): audits the six surfaces above against the operator stack — what LLM-as-judge models are wired, what workflow engine is in place, what ticketing system runs reviewer queues, what WORM storage holds the audit trail today; where the borderline backlog and counsel-escalation friction lands. Tier 2 AI Swarm Setup Sprint (4-8 weeks): builds the borderline-routing meta-skill on the governance-router agent, with the LLM-as-judge prompt library and the compliance overlay rule library reviewed by operator counsel and data-science teams. Tier 3 Fractional CMO with AI Swarm (6-month minimum, 1-2 days/wk embedded): continues operating the meta-skill end-to-end and coordinating with the upstream compliance-overlay-manager, GBP, JSON-LD, master-record, photo, and franchisee-moderation skills.

Question 4

Who owns the borderline registry, routing configuration, reviewer-assignment configuration, and audit trail?

Accepted Answer

Operator owns 100% of every artifact. The borderline registry sits inside the operator data warehouse (Snowflake, BigQuery, Databricks, or Redshift). The borderline-classification model code sits in the operator repo with operator-controlled deploy pipeline and is aligned with the operator data-science team and counsel. The risk-tier routing configuration is aligned with operator counsel, brand, and compliance officers. The reviewer-assignment configuration is aligned with operator counsel and the workforce-management team. The cross-skill feedback model code is aligned with the operator data-science team. The LLM-as-judge prompt library lives in the operator repo. LLM API credentials sit under operator billing. The compliance overlay rule library lives in the operator repo with attorney-approved updates and tracks HIPAA, FDA OPDP, DEA, Metrc, DISCUS, FDA tobacco, state licensing, state AG, FTC, Lanham, ADA, COPPA, state lemon law, Prop 65, WEEE/RoHS/REACH, per-jurisdiction breach-notification, and EU AI Act Article 22 transparency. The brand spec is versioned in the operator repo. The audit trail persists in operator-controlled WORM storage (AWS S3 Object Lock, Google Cloud Storage retention, Azure Blob immutable, or Snowflake Time Travel). Completions owns the orchestration knowledge — how to design borderline classification contracts, how to tune risk-tier routing, how to debug reviewer-assignment cascades, and how to coordinate with the compliance-overlay-manager and franchisee-moderation siblings. The operator can in-house at any time; Completions credentials revoke immediately on engagement-end.

Question 5

What does Completions commit to on a Tier 3 engagement?

Accepted Answer

Completions commits to a 6-workstream pre-engagement-baseline reporting cycle on the governance-router agent: (1) Borderline-Classification workstream — pre-engagement baseline of which LLM-as-judge models are wired today, threshold-band calibration state, and explainability coverage, then weekly reporting on classification coverage, threshold-band stability, and rule-citation completeness. (2) Risk-Tier Routing workstream — pre-engagement baseline of which of the six risk tiers and ten routing decisions exist today, then weekly reporting on routing-decision outcomes against per-tier SLA and escalation-path adherence. (3) Reviewer-Assignment workstream — pre-engagement baseline of which of the eleven reviewer roles are staffed today and current queue-depth, then weekly reporting on workload balance, fatigue indicators, expertise-routing accuracy, and SLA adherence per reviewer role. (4) Cross-Skill Feedback workstream — pre-engagement baseline of which upstream skills receive feedback today, then weekly reporting on feedback-loop closure rates and observed upstream-skill confidence-threshold tightening. (5) Compliance-Overlay workstream — pre-engagement baseline of which frameworks in the rule library are attorney-approved today, then weekly reporting on overlay completeness, breach-notification-window adherence where applicable, and EU AI Act Article 22 transparency-record persistence where applicable. (6) Audit-Trail + WORM workstream — pre-engagement baseline of WORM-storage discipline today, then weekly reporting on audit-trail completeness, evidentiary-quality, and operator-account-ownership confirmation. Caveats: classification confidence depends on per-vertical training-data availability and operator data-warehouse maturity; counsel-approval cycle times are outside Completions control and can lag overlay updates; LLM-vendor API rate limits and model deprecation are outside Completions control; per-jurisdiction breach-notification windows, EU AI Act Article 22 transparency-record requirements, state-licensing-board policy, and FTC policy can change without notice and require operator-counsel re-review of the overlay; WORM-storage retention windows are operator-counsel-policy decisions; reviewer-role staffing and fatigue management are operator workforce-management decisions; the audit trail persists to operator-controlled WORM storage on the operator cloud account.

Question 6

How does engagement end and what is the operator transition path?

Accepted Answer

Tier 3 engagements are 6-month minimum with 90-day notice. At engagement end, Completions transitions back to operator in-house in 30-60 days: operating-playbook hand-off + in-house staff training across 3-5 operator team members covering classification, risk-tier routing, reviewer assignment, cross-skill feedback, compliance-overlay management, and WORM-audit discipline + borderline registry hand-off + classification model code hand-off + routing configuration hand-off + reviewer-assignment configuration hand-off + cross-skill feedback model code hand-off + LLM-as-judge prompt library hand-off + LLM API credentials hand-off + compliance overlay rule library hand-off + audit-trail hand-off with WORM-storage operator-account-ownership confirmation; Completions credentials revoke immediately on engagement-end.

Done-for-you borderline routing for AI-generated marketing artifacts at multi-location, multi-vertical, multi-jurisdictional operators — a borderline-classification + risk-tier-routing + reviewer-assignment + cross-skill-feedback meta-skill on the governance-router agent.

Frequently asked