Files
Timmy-time-dashboard/config/moderation.yaml
Claude (Opus 4.6) 1697e55cdb
Some checks failed
Tests / lint (push) Has been cancelled
Tests / test (push) Has been cancelled
[claude] Add content moderation pipeline (Llama Guard + game-context prompts) (#1056) (#1059)
2026-03-23 02:14:42 +00:00

108 lines
4.1 KiB
YAML

# 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."