diff --git a/utils/llm/README.md b/utils/llm/README.md
index 713e4aa..842b6e0 100644
--- a/utils/llm/README.md
+++ b/utils/llm/README.md
@@ -141,7 +141,7 @@ All configuration is via environment variables:
 
 | Variable | Default | Description |
 |---|---|---|
-| `LLM_PROVIDER` | `venice` | Provider name: `venice`, `openai`, `anthropic`, or custom |
+| `LLM_PROVIDER` | `venice` | Provider name: `venice`, `groq`, `openai`, `anthropic`, or custom |
 | `LLM_API_KEY` | *(required)* | API key for the LLM provider |
 | `LLM_MODEL` | `grok-41-fast` | Model identifier |
 | `LLM_BASE_URL` | *(per provider)* | API base URL (not needed for anthropic) |
@@ -154,6 +154,7 @@ All configuration is via environment variables:
 | Provider | Base URL | Default Model | Package |
 |---|---|---|---|
 | Venice.ai | `https://api.venice.ai/api/v1` | `grok-41-fast` | `openai` |
+| Groq | `https://api.groq.com/openai/v1` | `openai/gpt-oss-safeguard-20b` | `openai` |
 | OpenAI | `https://api.openai.com/v1` | `gpt-4o-mini` | `openai` |
 | Anthropic | *(native API)* | `claude-haiku-4-5-20251001` | `anthropic` |
 | Custom | Set `LLM_BASE_URL` | Set `LLM_MODEL` | `openai` |
diff --git a/utils/llm/ai_explainer.py b/utils/llm/ai_explainer.py
index ae7cda7..69f330c 100644
--- a/utils/llm/ai_explainer.py
+++ b/utils/llm/ai_explainer.py
@@ -28,8 +28,9 @@
 - Asset/token flow changes
 - State changes and their impact
 - Risk assessment (LOW/MEDIUM/HIGH/CRITICAL)
-- Any concerns or notable observations
+- Any concerns or notable observations"""
 
+FORMAT_REMINDER = """
 Format your response exactly as:
 TLDR: <your short summary>
 
@@ -138,9 +139,27 @@ def _build_prompt(
     if simulation:
         parts.append(f"\n--- Simulation Results ---\n{_format_simulation_context(simulation)}")
 
+    parts.append(FORMAT_REMINDER)
+
     return "\n".join(parts)
 
 
+def _find_marker(text: str, keyword: str) -> tuple[int, int]:
+    """Find a section marker like 'TLDR:' or '### DETAIL' and return (start_of_marker, start_of_content).
+
+    Handles variations: 'KEYWORD:', '## KEYWORD', '**KEYWORD**', '**KEYWORD:**', etc.
+    Returns (-1, -1) if not found.
+    """
+    import re
+
+    heading = r"#{1,4}"  # fmt: skip
+    pattern = rf"(?:^|\n)\s*(?:{heading}\s+)?(?:\*{{2}})?{keyword}(?:\*{{2}})?[:\s]*"
+    match = re.search(pattern, text, re.IGNORECASE)
+    if match:
+        return match.start(), match.end()
+    return -1, -1
+
+
 def _parse_explanation(raw: str) -> Explanation:
     """Parse LLM response into summary and detail sections.
 
@@ -151,28 +170,23 @@ def _parse_explanation(raw: str) -> Explanation:
         <detailed analysis>
 
     Falls back gracefully if the LLM doesn't follow the format exactly.
+    Handles markdown-style headers like '### DETAIL' or '**TLDR:**'.
     """
-    # Try to split on DETAIL: marker
-    upper = raw.upper()
-    detail_idx = upper.find("DETAIL:")
-    tldr_idx = upper.find("TLDR:")
-
-    if tldr_idx != -1 and detail_idx != -1:
-        # Both markers found — extract each section
-        tldr_start = tldr_idx + len("TLDR:")
-        summary = raw[tldr_start:detail_idx].strip()
-        detail = raw[detail_idx + len("DETAIL:") :].strip()
+    tldr_start, tldr_content = _find_marker(raw, "TLDR")
+    detail_start, detail_content = _find_marker(raw, "DETAIL")
+
+    if tldr_start != -1 and detail_start != -1:
+        summary = raw[tldr_content:detail_start].strip()
+        detail = raw[detail_content:].strip()
         return Explanation(summary=summary, detail=detail)
 
-    if tldr_idx != -1:
-        # Only TLDR found — everything after it is the summary, no detail
-        summary = raw[tldr_idx + len("TLDR:") :].strip()
+    if tldr_start != -1:
+        summary = raw[tldr_content:].strip()
         return Explanation(summary=summary, detail="")
 
-    if detail_idx != -1:
-        # Only DETAIL found — everything before is summary
-        summary = raw[:detail_idx].strip()
-        detail = raw[detail_idx + len("DETAIL:") :].strip()
+    if detail_start != -1:
+        summary = raw[:detail_start].strip()
+        detail = raw[detail_content:].strip()
         return Explanation(summary=summary, detail=detail)
 
     # No markers — use full response as summary (backward compatible)
diff --git a/utils/llm/factory.py b/utils/llm/factory.py
index ca3abca..d855337 100644
--- a/utils/llm/factory.py
+++ b/utils/llm/factory.py
@@ -8,6 +8,7 @@
 
 Provider defaults:
     venice: base_url=https://api.venice.ai/api/v1, model=llama-3.3-70b
+    groq: base_url=https://api.groq.com/openai/v1, model=openai/gpt-oss-safeguard-20b
     openai: base_url=https://api.openai.com/v1, model=gpt-4o-mini
     anthropic: model=claude-haiku-4-5-20251001 (uses native Anthropic API)
     Custom: Set LLM_BASE_URL and LLM_MODEL explicitly.
@@ -26,6 +27,10 @@
         "base_url": "https://api.venice.ai/api/v1",
         "model": "grok-41-fast",
     },
+    "groq": {
+        "base_url": "https://api.groq.com/openai/v1",
+        "model": "openai/gpt-oss-safeguard-20b",
+    },
     "openai": {
         "base_url": "https://api.openai.com/v1",
         "model": "gpt-4o-mini",