diff --git a/backend-compliance/compliance/services/llm_cascade.py b/backend-compliance/compliance/services/llm_cascade.py
index d7541f04..6fd9d190 100644
--- a/backend-compliance/compliance/services/llm_cascade.py
+++ b/backend-compliance/compliance/services/llm_cascade.py
@@ -119,7 +119,12 @@ async def _call_ollama(system: str, user: str,
             r.raise_for_status()
         return (r.json().get("message") or {}).get("content", "") or ""
     except Exception as e:
-        logger.warning("ollama cascade tier 1 failed: %s", e)
+        # P83-followup: explizit type+message loggen damit empty-message
+        # exceptions (z.B. ReadTimeout) diagnostizierbar sind.
+        logger.warning(
+            "ollama cascade tier 1 failed: %s (%s) model=%s base=%s",
+            str(e) or "(no message)", type(e).__name__, model, base,
+        )
         return ""
 
 
@@ -190,9 +195,21 @@ async def call_with_cascade(
     cached = _cache_get(key)
     if cached:
         cached["cached"] = True
+        logger.info("cascade cache HIT key=%s len=%d", key[-12:],
+                     len(cached.get("text", "")))
         return cached
 
     input_len = len(user)
+    # Pre-flight: warn if Tier 2 + Tier 3 are unconfigured so user knows
+    # we are de-facto running single-tier (cascade collapse).
+    if not (os.getenv("OVH_LLM_URL", "").strip()
+            and os.getenv("OVH_LLM_MODEL", "").strip()):
+        if not os.getenv("ANTHROPIC_API_KEY", "").strip():
+            logger.warning(
+                "cascade: Tier 2 (OVH) AND Tier 3 (Anthropic) unconfigured — "
+                "running on Tier 1 (Qwen) only. Set OVH_LLM_URL/MODEL/KEY "
+                "or ANTHROPIC_API_KEY to enable fallbacks."
+            )
     # Tier 1: Qwen lokal
     text = await _call_ollama(system, user, max_tokens=max_tokens)
     conf = _heuristic_confidence(text, input_len)
diff --git a/backend-compliance/compliance/services/tcf_vendor_authority.py b/backend-compliance/compliance/services/tcf_vendor_authority.py
index 429341e5..82e4f165 100644
--- a/backend-compliance/compliance/services/tcf_vendor_authority.py
+++ b/backend-compliance/compliance/services/tcf_vendor_authority.py
@@ -146,7 +146,7 @@ def lookup_tcf_authority(
             """
             SELECT cookie_name, actual_category, vendor_name
             FROM compliance.cookie_library
-            WHERE source = 'iab_tcf_v2'
+            WHERE source_name = 'iab_tcf_v2'
               AND LOWER(vendor_name) LIKE :pat
             LIMIT 5
             """
diff --git a/backend-compliance/compliance/services/vendor_llm_extractor.py b/backend-compliance/compliance/services/vendor_llm_extractor.py
index 10ecf14d..13acb14b 100644
--- a/backend-compliance/compliance/services/vendor_llm_extractor.py
+++ b/backend-compliance/compliance/services/vendor_llm_extractor.py
@@ -59,9 +59,17 @@ async def extract_vendors_via_llm(
     und passt in Qwen3-30b-a3b (128k Context) sowie OVH 120B.
     """
     if not cookie_text or len(cookie_text) < 500:
+        logger.info(
+            "LLM vendor extraction SKIP: cookie_text too short (%d chars, need >=500)",
+            len(cookie_text or ""),
+        )
         return []
     excerpt = cookie_text[:max_text_chars]
     user_prompt = f"Cookie-Richtlinie-Text:\n\n{excerpt}"
+    logger.info(
+        "LLM vendor extraction START: input=%d chars (excerpt %d), calling cascade",
+        len(cookie_text), len(excerpt),
+    )
 
     # P31: nutze tiered LLM-Cascade mit Cache (Qwen → OVH → Anthropic).
     # Re-Runs derselben Cookie-Doc landen im Valkey-Cache (7d TTL) und
@@ -76,13 +84,24 @@ async def extract_vendors_via_llm(
         vendors = _parse_vendor_list(res.get("text", ""))
         if vendors:
             logger.info(
-                "LLM vendor extraction (cascade %s, conf=%.2f, cached=%s): %d vendors",
+                "LLM vendor extraction OK (cascade %s, conf=%.2f, cached=%s): %d vendors",
                 res.get("source"), res.get("confidence", 0),
                 res.get("cached"), len(vendors),
             )
-        return vendors
+            return vendors
+        # Silent failure ist unbrauchbar fuer Diagnose: log explicit dass
+        # cascade returnte aber 0 Vendors (parse failed oder LLM gab nix).
+        logger.warning(
+            "LLM vendor extraction returned 0 vendors (cascade source=%s "
+            "text_len=%d below_threshold=%s) — fallback to direct calls",
+            res.get("source"), len(res.get("text", "") or ""),
+            res.get("below_threshold"),
+        )
     except Exception as e:
-        logger.warning("Cascade extract failed, fallback to direct Qwen: %s", e)
+        logger.warning(
+            "Cascade extract failed, fallback to direct Qwen: %s (%s)",
+            str(e) or "(no message)", type(e).__name__,
+        )
 
     # Fallback: alte direkte Logik
     content = await _call_ollama(user_prompt)