[claude] Add content moderation pipeline (Llama Guard + game-context prompts) (#1056) (#1059)

config/moderation.yaml

@@ -0,0 +1,107 @@
+# Content Moderation Profiles
+# Per-game moderation configuration for the AI narrator pipeline.
+#
+# Each profile defines:
+#   - vocabulary_whitelist: Game terms safe in context (won't trigger moderation)
+#   - context_prompt: System prompt framing for the narrator
+#   - threshold: Confidence threshold — flags below this pass through
+#   - fallbacks: Pre-generated safe narration by scene type
+#
+# Model options (from research):
+#   llama-guard3:1b  — Speed (<30ms/sentence, INT4 quantized)
+#   shieldgemma:2b   — Accuracy (+10.8% AU-PRC, ~50-100ms)
+#
+# Override guard model via MODERATION_GUARD_MODEL env var.
+
+# ── Guard model selection ────────────────────────────────────────────────────
+guard_model: "llama-guard3:1b"
+
+# ── Streaming disclosure notes ───────────────────────────────────────────────
+# YouTube: Use "Altered or synthetic content" toggle
+# Twitch: Standard community guidelines (no specific AI disclosure req as of 2026-03)
+
+# ── Game Profiles ────────────────────────────────────────────────────────────
+profiles:
+
+  morrowind:
+    display_name: "The Elder Scrolls III: Morrowind"
+    threshold: 0.85
+    vocabulary_whitelist:
+      - Skooma
+      - Moon Sugar
+      - slave
+      - slavery
+      - Morag Tong
+      - Dark Brotherhood
+      - Telvanni
+      - Camonna Tong
+      - smuggler
+      - assassin
+      - Sixth House
+      - Corprus
+      - Dagoth Ur
+      - Nerevarine
+      - Balmora
+      - Vivec
+      - Almsivi
+      - Ordinators
+      - Ashlanders
+      - outlander
+      - N'wah
+    context_prompt: >
+      You are narrating gameplay of The Elder Scrolls III: Morrowind.
+      Morrowind contains mature themes including slavery, drug use
+      (Skooma/Moon Sugar), assassin guilds (Morag Tong, Dark Brotherhood),
+      and political intrigue. Treat these as game mechanics and historical
+      worldbuilding within the game's fictional universe. Never editorialize
+      on real-world parallels. Narrate events neutrally as a game
+      commentator would.
+    fallbacks:
+      combat: "The battle rages on in the ashlands of Vvardenfell."
+      dialogue: "The conversation continues between the characters."
+      exploration: "The Nerevarine presses onward through the landscape."
+      quest: "The quest unfolds as the hero navigates Morrowind's politics."
+      default: "The adventure continues in Morrowind."
+
+  skyrim:
+    display_name: "The Elder Scrolls V: Skyrim"
+    threshold: 0.85
+    vocabulary_whitelist:
+      - Skooma
+      - Dark Brotherhood
+      - Thieves Guild
+      - Stormcloak
+      - Imperial
+      - Dragonborn
+      - Dovahkiin
+      - Daedra
+      - Thalmor
+      - bandit
+      - assassin
+      - Forsworn
+      - necromancer
+    context_prompt: >
+      You are narrating gameplay of The Elder Scrolls V: Skyrim.
+      Skyrim features civil war, thieves guilds, assassin organizations,
+      and fantasy violence. Treat all content as in-game fiction.
+      Never draw real-world parallels. Narrate as a neutral game
+      commentator.
+    fallbacks:
+      combat: "Steel clashes as the battle continues in the wilds of Skyrim."
+      dialogue: "The conversation plays out in the cold northern land."
+      exploration: "The Dragonborn ventures further into the province."
+      default: "The adventure continues in Skyrim."
+
+  default:
+    display_name: "Generic Game"
+    threshold: 0.80
+    vocabulary_whitelist: []
+    context_prompt: >
+      You are narrating gameplay. Describe in-game events as a neutral
+      game commentator. Never reference real-world violence, politics,
+      or controversial topics. Stay focused on game mechanics and story.
+    fallbacks:
+      combat: "The action continues on screen."
+      dialogue: "The conversation unfolds between characters."
+      exploration: "The player explores the game world."
+      default: "The gameplay continues."