PR for #987: Implement content moderation pipeline (Llama Guard + game-context prompts) #1038

gemini · 2026-03-22T23:13:22Z

gemini commented

2026-03-22 23:13:22 +00:00

Closes #987

Source

#946 — Integration Architecture: Eight Deep Dives (content moderation section)
#982 — Session Crystallization Playbook

Objective

Build a real-time content moderation pipeline for Timmy's narration output to prevent a Neuro-sama-style incident while handling Morrowind's mature themes appropriately.

Architecture

Moderation and TTS preprocessing run in parallel: while Llama Guard checks the sentence, Kokoro tokenizes and phonemizes. On pass, TTS fires immediately. On fail, a contextual fallback narration replaces flagged content.

Scope

Model Options (ranked)

Llama Guard 3 1B (INT4): Best for speed. <30ms per sentence on GPU. Run alongside TTS.
ShieldGemma 2B: Best for accuracy. +10.8% AU-PRC over Llama Guard. Threshold tuning via probability scores.
NeMo Guardrails: Framework for parallel guardrails with custom Colang rules (~0.5s with 5 GPU rails).
LEG (Lightweight Explainable Guardrail): Emerging, <8ms, 7x faster.

Game-Context Awareness

System prompt instructs narrator to describe slavery as "game mechanic and historical worldbuilding"
Drugs described as "in-game consumable items"
Per-game moderation threshold profiles whitelist expected vocabulary: "Skooma", "slave", "Morag Tong"
Never editorialize on real-world parallels

Fallback Strategy

On moderation failure: pre-generated contextual fallback narration (per game scene type)
Visible "filtered" indicator (transparency-as-entertainment, per Neuro-sama resolution)

Key Design Notes

Neuro-sama earned a 2-week Twitch ban for unfiltered AI output — moderation is non-negotiable
Latency budget: <30ms for moderation, parallel with TTS preprocessing
Morrowind's themes require nuanced filtering, not blanket censorship

#966 — Three-tier metabolic LLM router (narration output feeds into this)
#959 — Narration templates (content that needs moderation)

Closes #987 ## Source - #946 — Integration Architecture: Eight Deep Dives (content moderation section) - #982 — Session Crystallization Playbook ## Objective Build a real-time content moderation pipeline for Timmy's narration output to prevent a Neuro-sama-style incident while handling Morrowind's mature themes appropriately. ## Architecture Moderation and TTS preprocessing run **in parallel**: while Llama Guard checks the sentence, Kokoro tokenizes and phonemizes. On pass, TTS fires immediately. On fail, a contextual fallback narration replaces flagged content. ## Scope ### Model Options (ranked) 1. **Llama Guard 3 1B (INT4):** Best for speed. <30ms per sentence on GPU. Run alongside TTS. 2. **ShieldGemma 2B:** Best for accuracy. +10.8% AU-PRC over Llama Guard. Threshold tuning via probability scores. 3. **NeMo Guardrails:** Framework for parallel guardrails with custom Colang rules (~0.5s with 5 GPU rails). 4. **LEG (Lightweight Explainable Guardrail):** Emerging, <8ms, 7x faster. ### Game-Context Awareness - System prompt instructs narrator to describe slavery as "game mechanic and historical worldbuilding" - Drugs described as "in-game consumable items" - Per-game moderation threshold profiles whitelist expected vocabulary: "Skooma", "slave", "Morag Tong" - Never editorialize on real-world parallels ### Fallback Strategy - On moderation failure: pre-generated contextual fallback narration (per game scene type) - Visible "filtered" indicator (transparency-as-entertainment, per Neuro-sama resolution) ## Key Design Notes - Neuro-sama earned a 2-week Twitch ban for unfiltered AI output — moderation is non-negotiable - Latency budget: <30ms for moderation, parallel with TTS preprocessing - Morrowind's themes require nuanced filtering, not blanket censorship ## Related - #966 — Three-tier metabolic LLM router (narration output feeds into this) - #959 — Narration templates (content that needs moderation)

Timmy closed this pull request

2026-03-23 15:03:05 +00:00

Tests / lint (push) Has been cancelled

Details

Tests / test (push) Has been cancelled

Details

Tests / lint (pull_request) Successful in 19s

Details

Tests / test (pull_request) Failing after 25m3s

Details

Pull request closed

Please reopen this pull request to perform a merge.

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: Rockachopa/Timmy-time-dashboard#1038