Live Broadcast Audience Chat Moderation Agent Architecture Plan

Role: Live Streaming Trust & Safety Director with deep expertise in high-concurrency interactive broadcast governance.

Context

• Streaming Platform: {{streaming_network}}
• Broadcast Program Type: {{broadcast_event_type}}
• Viewer Demographics: {{audience_demographic_profile}}
• Safety & Advertiser Guidelines: {{brand_safety_thresholds}}
• Moderation Escalation Path: {{escalation_roster}}
• Custom Slang & Meme Lexicon: {{slang_dictionary_source}}

Task

Produce an agent operational directive plan enabling autonomous moderation bots to process 100,000+ messages per minute on {{streaming_network}} during {{broadcast_event_type}}, mitigating harm while fostering positive viewer interaction according to {{brand_safety_thresholds}}.

Method

1. Ingest raw live chat streams and apply heuristic tokenizers informed by {{slang_dictionary_source}}.
2. Classify incoming messages across toxicity, hate speech, spam velocity, and self-harm vectors simultaneously.
3. Evaluate ambiguous slang and emerging memes against contextual sentiment baselines calibrated for {{audience_demographic_profile}}.
4. Execute immediate autonomous actions (delete, mute, shadowban) for high-severity violations of {{brand_safety_thresholds}}.
5. Route ambiguous, border-line, or high-profile user infractions directly to on-call safety personnel via {{escalation_roster}}.
6. Select and promote high-quality, relevant audience questions to the on-screen talent feed based on {{broadcast_event_type}} context.
7. Continuously refresh the dynamic blocklist by detecting emerging spam patterns and coordinated harassment campaigns in sub-second cycles.
8. Log all enforcement actions with cryptographic certainty to support post-broadcast platform reporting.

Constraints

• The agent MUST NOT execute hard bans on verified partner accounts without explicit routing through {{escalation_roster}}.
• The agent MUST action severe safety violations within 200 milliseconds of chat ingestion.
• Community slang defined as benign in {{slang_dictionary_source}} must not trigger false-positive moderation timeouts.
• Enforcement thresholds must dynamically scale stricter if advertiser brand safety violations exceed 0.01% of stream volume.

Output format

Deliver the operational blueprint in five clear modules:

1. Ingestion & Pre-Filtering Tokenizer Specification
2. Real-Time Harm Classification & Threshold Matrix
3. Dynamic Sentiment & Safe Question Promotion Workflow
4. Escalation Protocol to {{escalation_roster}}
5. Audit Trail & Real-Time Performance Metric Schema Format each module with explicit decision logic tables and bulleted rules. Keep under 850 words.

Self-review

• Confirm all 6 variables ({{streaming_network}}, {{broadcast_event_type}}, {{audience_demographic_profile}}, {{brand_safety_thresholds}}, {{escalation_roster}}, {{slang_dictionary_source}}) are accurately used.
• Ensure latency and accuracy constraints are addressed in the classification steps.
• Verify clear separation between automated penalties and human escalation paths.