diff --git a/.env.example b/.env.example
index fff4a27..97b21e6 100644
--- a/.env.example
+++ b/.env.example
@@ -42,6 +42,23 @@
 # TURNSTONE_EMBED_MODEL=BAAI/bge-small-en-v1.5
 # TURNSTONE_EMBED_DEVICE=cpu
 
+# --- Anomaly scoring pipeline (IDS / watchdog) ---
+# Batch-scores every ingested log entry after each glean cycle.
+# Any HuggingFace text-classification model works; the byviz classifier (already
+# required by the diagnose pipeline) is the recommended starting point.
+# Detections above the threshold are inserted into the detections table and
+# surfaced in the Security Alerts tab.
+#
+# Set TURNSTONE_ANOMALY_MODEL to enable; leave unset to disable (safe default).
+# TURNSTONE_ANOMALY_MODEL=byviz/bylastic_classification_logs
+# TURNSTONE_ANOMALY_DEVICE=cpu          # or "cuda" / "mps" for GPU inference
+# TURNSTONE_ANOMALY_THRESHOLD=0.80      # confidence floor for detection insertion
+# TURNSTONE_ANOMALY_INTERVAL=0          # standalone loop (0 = glean-triggered only)
+#
+# HuggingFace model cache — share with the host to avoid re-downloading models.
+# HF_HOME=/hf_cache                     # inside container (set in docker-compose)
+# HF_CACHE_PATH=/Library/Assets/LLM    # host bind-mount source (docker-compose only)
+
 # --- Air-gapped / offline deployment ---
 # Set to 1 to block all HuggingFace hub network access at runtime.
 # Pre-download models to ~/.cache/huggingface/ before deploying — see docs/air-gapped-deployment.md.
diff --git a/docker-compose.yml b/docker-compose.yml
index 8c9bf29..d197bc1 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -41,9 +41,23 @@ services:
       TURNSTONE_GLEAN_INTERVAL: ${TURNSTONE_GLEAN_INTERVAL:-900}
       TURNSTONE_SOURCE_HOST: ${TURNSTONE_SOURCE_HOST:-}
       TURNSTONE_SUBMIT_ENDPOINT: ${TURNSTONE_SUBMIT_ENDPOINT:-}
+      # --- Multi-agent diagnose pipeline ---
+      TURNSTONE_MULTI_AGENT_DIAGNOSE: ${TURNSTONE_MULTI_AGENT_DIAGNOSE:-false}
+      TURNSTONE_CLASSIFIER_MODEL: ${TURNSTONE_CLASSIFIER_MODEL:-}
+      TURNSTONE_EMBED_BACKEND: ${TURNSTONE_EMBED_BACKEND:-}
+      TURNSTONE_EMBED_MODEL: ${TURNSTONE_EMBED_MODEL:-}
+      TURNSTONE_EMBED_DEVICE: ${TURNSTONE_EMBED_DEVICE:-cpu}
+      # --- Anomaly scoring pipeline ---
+      TURNSTONE_ANOMALY_MODEL: ${TURNSTONE_ANOMALY_MODEL:-}
+      TURNSTONE_ANOMALY_DEVICE: ${TURNSTONE_ANOMALY_DEVICE:-cpu}
+      TURNSTONE_ANOMALY_THRESHOLD: ${TURNSTONE_ANOMALY_THRESHOLD:-0.75}
+      TURNSTONE_ANOMALY_INTERVAL: ${TURNSTONE_ANOMALY_INTERVAL:-0}
+      # --- HuggingFace model cache ---
+      HF_HOME: /hf_cache
     volumes:
       - ./patterns:/app/patterns:ro
       - ./data:/app/data  # optional: persists SQLite files if DATABASE_URL unset
+      - ${HF_CACHE_PATH:-/Library/Assets/LLM}:/hf_cache:ro  # shared model cache
 
 volumes:
   turnstone_pgdata:
diff --git a/docker-standalone.sh b/docker-standalone.sh
index 7098fa8..8d45406 100755
--- a/docker-standalone.sh
+++ b/docker-standalone.sh
@@ -62,7 +62,10 @@ set -euo pipefail
 REPO_DIR="${HOME}/turnstone"
 DATA_DIR="${REPO_DIR}/data"
 PATTERNS_DIR="${REPO_DIR}/patterns"
-HF_CACHE_DIR="${REPO_DIR}/hf-cache"   # persists downloaded ML models across restarts
+# HF_CACHE_DIR: override to a shared cache directory to avoid re-downloading models.
+# Example (Heimdall, where byviz/bylastic_classification_logs is already cached):
+#   export HF_CACHE_DIR=/Library/Assets/LLM
+HF_CACHE_DIR="${HF_CACHE_DIR:-${REPO_DIR}/hf-cache}"
 
 TZ="${TZ:-America/Los_Angeles}"
 
@@ -83,6 +86,16 @@ TZ="${TZ:-America/Los_Angeles}"
 #   bash ~/turnstone/docker-standalone.sh
 #
 
+# ── Anomaly scoring pipeline (IDS / watchdog) ────────────────────────────────
+# Set TURNSTONE_ANOMALY_MODEL to enable automatic anomaly scoring after each
+# glean run.  The byviz classifier (already used by the diagnose pipeline) is
+# a good default — it's cached alongside the other models.
+#
+#   export TURNSTONE_ANOMALY_MODEL=byviz/bylastic_classification_logs
+#   export TURNSTONE_ANOMALY_THRESHOLD=0.80   # confidence floor (default 0.75)
+#   bash ~/turnstone/docker-standalone.sh
+#
+
 # ── Multi-agent diagnose pipeline ────────────────────────────────────────────
 # Enable the 5-stage ML pipeline to get smarter diagnose results.
 #
@@ -134,6 +147,10 @@ docker run -d \
   -e TURNSTONE_EMBED_BACKEND="${TURNSTONE_EMBED_BACKEND:-sentence_transformers}" \
   -e TURNSTONE_EMBED_MODEL="${TURNSTONE_EMBED_MODEL:-sentence-transformers/all-MiniLM-L6-v2}" \
   -e TURNSTONE_EMBED_DEVICE="${TURNSTONE_EMBED_DEVICE:-cpu}" \
+  -e TURNSTONE_ANOMALY_MODEL="${TURNSTONE_ANOMALY_MODEL:-}" \
+  -e TURNSTONE_ANOMALY_DEVICE="${TURNSTONE_ANOMALY_DEVICE:-cpu}" \
+  -e TURNSTONE_ANOMALY_THRESHOLD="${TURNSTONE_ANOMALY_THRESHOLD:-0.75}" \
+  -e TURNSTONE_ANOMALY_INTERVAL="${TURNSTONE_ANOMALY_INTERVAL:-0}" \
   localhost/turnstone:latest
 
 echo ""