Play 61 — Content Moderation V2
Advanced multi-modal content moderation — Azure Content Safety for text+image+video, custom category blocklists, per-category severity thresholds, severity-based routing (auto-block/human-review/flag), Service Bus review queues with SLA-based priority, appeal workflows with different-reviewer requirement, and streaming moderation.
Architecture
Full architecture details:
architecture.md
How It Differs from Play 10 (Content Moderation v1)
| Aspect | Play 10 (v1) | Play 61 (V2) |
|---|---|---|
| Modalities | Text only | Text + image + video |
| Categories | Built-in only | Built-in + custom blocklists |
| Routing | Binary (block/allow) | 4-tier: block/review/flag/allow |
| Review | No human review | Service Bus queue with SLA per category |
| Appeals | None | Full appeal workflow with different reviewer |
| Thresholds | Global | Per-category severity thresholds |
| Explanation | None | Category + severity + policy reference |
Key Metrics
| Metric | Target | Description |
|---|---|---|
| Precision (violence) | > 95% | Blocked content is actually violent |
| Recall (self-harm) | > 98% | Self-harm content not missed |
| False Positive Rate | < 3% | Safe content incorrectly blocked |
| Text Latency | < 200ms | Real-time text moderation |
| Queue Wait (self-harm) | < 15 min | Critical category SLA |
| Appeal Success Rate | 5-15% | Healthy appeal overturn range |
Cost Estimate
| Service | Dev | Prod | Enterprise |
|---|---|---|---|
| Azure AI Content Safety | $0 | $200 | $800 |
| Azure OpenAI | $50 | $400 | $1,500 |
| Cosmos DB | $5 | $150 | $600 |
| Azure Service Bus | $5 | $50 | $700 |
| Azure Blob Storage | $3 | $30 | $100 |
| Key Vault | $1 | $5 | $15 |
| Application Insights | $0 | $30 | $100 |
| Total | $64 | $865 | $3,815 |
Detailed breakdown with SKUs and optimization tips:
cost.json· Azure Pricing Calculator
WAF Alignment
| Pillar | Implementation |
|---|---|
| Responsible AI | Self-harm lowest thresholds, per-category tuning, cultural context |
| Security | Content Safety API, Key Vault, data not stored after moderation |
| Reliability | Service Bus for reliable queuing, dead-letter for failed reviews |
| Cost Optimization | gpt-4o-mini for explanations, skip LLM for clear cases |
| Operational Excellence | SLA-based review routing, appeal audit trail |
| Performance Efficiency | <200ms text, parallel multi-modal, video frame sampling |
FAI Manifest
| Field | Value |
|---|---|
| Play | 61-content-moderation-v2 |
| Version | 1.0.0 |
| Knowledge | T2-Responsible-AI, F1-GenAI-Foundations, O2-Agent-Coding, T3-Production-Patterns, R3-Deterministic-AI |
| WAF Pillars | responsible-ai, security, reliability, operational-excellence |
| Groundedness | ≥ 85% |
| Safety | 0 violations max |
Last updated on