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

2026-03-23 02:14:42 +00:00
parent 092c982341
commit 1697e55cdb
6 changed files with 1013 additions and 0 deletions
--- a/src/config.py
+++ b/src/config.py
@@ -99,6 +99,14 @@ class Settings(BaseSettings):
    anthropic_api_key: str = ""
    claude_model: str = "haiku"

+    # ── Content Moderation ──────────────────────────────────────────────
+    # Three-layer moderation pipeline for AI narrator output.
+    # Uses Llama Guard via Ollama with regex fallback.
+    moderation_enabled: bool = True
+    moderation_guard_model: str = "llama-guard3:1b"
+    # Default confidence threshold — per-game profiles can override.
+    moderation_threshold: float = 0.8
+
    # ── Spark Intelligence ────────────────────────────────────────────────
    # Enable/disable the Spark cognitive layer.
    # When enabled, Spark captures swarm events, runs EIDOS predictions,