Pre-Launch Compliance
AI Model Card
ContentPulse AI Model Card — EU AI Act Article 50 transparency + California AI Executive Order disclosures
Sections
Product Description
Product Name
ContentPulse by Aurora AI Solutions Studio UG
Version & Classification Date
Version 1.0 (Pre-Launch) | Classification Date: April 30, 2026
What It Does
ContentPulse is an AI-powered content multiplication platform for B2B agencies. It ingests long-form input (audio, video, transcripts, URLs) and generates platform-native variants — LinkedIn posts, X threads, blog posts, podcast scripts, and more. Every generated piece is held in a human review queue and requires explicit agency approval before publishing. Voice profiles are learned per-client from engagement metrics; no demographic data is collected.
AI Systems Used
| System | Function | Provider |
|---|---|---|
| Content Generation | Generates platform-native content variants from a single source piece; voice-profile aware | Anthropic Claude Sonnet 4.6 |
| Content Multiplication Engine | Maps source segments to target platforms; computes content multiplication score | Anthropic Claude Sonnet 4.6 |
| Voice Profile Learning | Per-client voice profile built from agency content history + engagement signals | Anthropic Claude Sonnet 4.6 |
| A+ Ghostwriting (Write-as-Client) | Suite-only mode that generates content using a client-specific voice profile | Anthropic Claude Sonnet 4.6 |
| Audio Transcription | Converts uploaded audio (MP3 / WAV / podcast episodes) to text with speaker diarization | Deepgram Nova-3 |
| Video Multimodal (Roadmap) | Transcript + visual semantics from MP4 / MOV uploads; ships in Sprint 7.8 | Google Gemini Files API |
| Engagement Readers | Pulls engagement events (impressions, reactions, comments) from publishing platforms | YouTube / X / Apple Podcasts APIs + LinkedIn email-bridge |
Model Configuration
Primary Model: Anthropic Claude Sonnet 4.6 (claude-sonnet-4-6)
Transcription: Deepgram Nova-3 (~$0.26/hr, diarization included)
Video (roadmap): Google Gemini Files API
Temperature (Generation): 0.7 (creative)
Temperature (Voice Profile Build): 0.3 (deterministic)
Data Processed
Processed Data Types
- • Agency-uploaded media (audio, video, transcripts, URLs) processed via Deepgram / Gemini and stored in Supabase EU
- • Agency content history + voice-profile artifacts (per-client when on Suite)
- • Engagement metrics from publishing platforms (impressions, reactions, comments)
- • Agency owner email + account metadata for communications
- • LinkedIn email-bridge notification payloads (parsed and discarded after URN extraction)
NOT Collected
- • PII of an agency's end-clients (only what the agency explicitly uploads)
- • Biometric data, demographic data, or protected-class information
- • Health, employment, or financial-services records
Risk Classification
| Framework | Classification | Notes |
|---|---|---|
| EU AI Act | Limited Risk (Article 50) | Content generation requires “Generated by ContentPulse” transparency disclosure on every consumed surface |
| California AI Executive Order | Moderate Risk | AI content generation triggers transparency + provenance metadata obligations |
| Colorado AI Act | Potentially Applicable | Effective June 30, 2026. Case-by-case “consequential decisions” assessment by deployer. |
| UK / ICO AI Guidance | Limited Risk | AI content generation triggers transparency obligations under UK GDPR Article 22 when used in an automated decision context |
| GDPR / EU Data Protection | Compliant | EU hosting (Frankfurt / eu-central-1), data minimization, right-to-delete (Art 17) endpoint live |
Human-in-the-Loop Controls
Every AI-generated piece is held in the agency's review queue. No piece reaches a publishing platform without explicit human approval.
Review Queue
All generated pieces — posts, threads, blog drafts, scripts — appear in /queue labeled “Generated by ContentPulse” pending review
Per-Piece Edit + Approve
Agency owner can edit, regenerate, approve, or reject every piece individually before any publishing action
Manual Publish Path
ContentPulse does not auto-publish. Agencies copy approved content into the publishing platform of their choice (or use the platform's native compose URL)
No Autonomous Publishing
Zero generated content is published to any external platform without explicit human approval and a manual publish action.
Transparency Measures
Article 50 Disclosure on Every Surface
A “Generated by ContentPulse” label is shown on the public share page (/share/[token]), the per-piece detail in /queue, and the composer output in /repurpose
Source Lineage Display
Every generated piece shows the source segment it was derived from (Content Map sidebar in the composer + per-piece source-segment panel)
Audit Logging
All generations + voice-profile updates are logged with input data, model version, and timestamp
Public Model Card
This page is publicly accessible and linked from the ContentPulse landing footer, the Aurora site footer, and the in-app footer
Bias & Fairness
Voice-Learning Foundation
Per-client voice profiles are learned from the agency's own engagement metrics (impressions, reactions, comments) — never from demographic, racial, gender, or protected-class data.
Generation Prompt Review
Generation prompts are reviewed for fairness before deployment. Refusal classes (see Content Policy) cover political disinformation, fabricated PII, and impersonation of named individuals without consent.
Provider-Level Safety
Anthropic Claude Sonnet inherits Anthropic's constitutional AI safety training. Deepgram Nova-3 transcription is reviewed for diarization-quality bias on accented speech.
Future Work (Post-Launch)
Synthetic-test cohorts for measuring fairness across client segments are planned for Launch+30 → +60.
Limitations & Known Issues
Generation Quality: Output quality depends on source-input quality. Short, low-context inputs yield lower-fidelity variants.
Voice Profile Maturity: Per-client voice profiles need a minimum of 5 published-and-engaged pieces before recursive rebuilds kick in.
Transcription Accuracy: Depends on audio quality, speaker diarization, accents, and background noise.
Language Support: Voice profile + generation are English-first in v1.0; other languages supported on a best-effort basis.
Engagement Bridge: The LinkedIn email-bridge requires the agency to set a forwarding rule. Dropped emails (no URN match) are silently discarded; reach-out is via the in-app banner.
Contact & Review
Contact for Questions
Review Cycle
Quarterly assessment of model performance, bias indicators, and regulatory changes
Next review: July 2026
Regulatory Monitoring
Active monitoring for:
- • EU AI Act Article 50 (effective Aug 2, 2026 — compliance target)
- • California AI Executive Order + CA AI Transparency Act (compliance target)
- • Colorado AI Act (effective June 30, 2026 — case-by-case)
- • UK GDPR Article 22 + ICO AI Guidance