diff --git a/docs/MILESTONES.md b/docs/MILESTONES.md
index 88e4e87..3a7f11e 100644
--- a/docs/MILESTONES.md
+++ b/docs/MILESTONES.md
@@ -49,7 +49,8 @@ Status and scope for each phase of SentinelOps. Dates are approximate; this is a
 
 ## Known gaps (tracked, not blocking)
 - No Grafana/Prometheus scrape config yet — metrics are exposed but not yet centrally collected
-- Sustained 40-49% error rates across all services for over an hour post-recovery (not resolving over time as event counts grew 10x) — this looks like a real bug (possibly in the `/api/services` error-rate calculation, or an actual issue in the log producer/ingestion path) rather than a transient catch-up artifact. Needs investigation as the next priority after Milestone 4.
+- Fixed: log producer was assigning `level` independently of `status_code` (a 200 response could randomly be logged as ERROR) — root cause of inflated error-rate metrics. Fixed by deriving level from status_code.
+- Still open: even after that fix, ANOMALY_PROBABILITY (tuned 0.02 → 0.0015 based on burst-duration math) has not brought error rates down to the expected ~5% in live testing — actual observed rates remain 25-37%. The duty-cycle calculation may be missing a factor (e.g. multiple concurrent bursts per service, or burst state not resetting as assumed). Needs a fresh look with the producer's actual runtime behavior, not just the math.
 
 ## Recently closed
 - `anomaly-rust`: added `detect()` end-to-end tests, including a regression guard for the directional-severity bug
diff --git a/infra/docker-compose.yml b/infra/docker-compose.yml
index 3ef8778..c34cc4b 100644
--- a/infra/docker-compose.yml
+++ b/infra/docker-compose.yml
@@ -31,6 +31,9 @@ services:
     depends_on: [redpanda]
     environment:
       KAFKA_BROKERS: redpanda:9092
+      LOG_TOPIC: service-logs
+      EMIT_INTERVAL_SECONDS: "0.2"
+      ANOMALY_PROBABILITY: "0.0015"
     ports:
       - "8080:8080"
     networks: [sentinelops-net]
diff --git a/services/log-producer/__pycache__/producer.cpython-312.pyc b/services/log-producer/__pycache__/producer.cpython-312.pyc
new file mode 100644
index 0000000..f393096
Binary files /dev/null and b/services/log-producer/__pycache__/producer.cpython-312.pyc differ
diff --git a/services/log-producer/producer.py b/services/log-producer/producer.py
index 4ca21f0..01b590a 100644
--- a/services/log-producer/producer.py
+++ b/services/log-producer/producer.py
@@ -33,8 +33,14 @@
 KAFKA_BROKERS = os.getenv("KAFKA_BROKERS", "localhost:9092")
 TOPIC = os.getenv("LOG_TOPIC", "service-logs")
 EMIT_INTERVAL_SECONDS = float(os.getenv("EMIT_INTERVAL_SECONDS", "0.2"))
-ANOMALY_PROBABILITY = float(os.getenv("ANOMALY_PROBABILITY", "0.02"))
-
+# Target: anomaly bursts should occupy roughly 3-5% of total time per
+# service, not be the dominant traffic pattern. With burst length averaging
+# ~40 events (random.randint(20, 60)) and ~1.25 events/sec/service, a 0.02
+# probability produced bursts ~44% of the time — anomalies were effectively
+# "normal," which defeats statistical anomaly detection entirely (there's no
+# stable low baseline to detect deviation from). 0.0015 targets ~5% duty cycle:
+# mean gap = 1/0.0015 ≈ 667 events, duty = 40/(667+40) ≈ 5.7%.
+ANOMALY_PROBABILITY = float(os.getenv("ANOMALY_PROBABILITY", "0.0015"))
 SERVICES = ["checkout", "auth", "payments", "inventory"]
 ENDPOINTS = {
     "checkout": ["/cart/add", "/cart/checkout", "/cart/remove"],
@@ -42,7 +48,6 @@
     "payments": ["/charge", "/refund", "/payment-status"],
     "inventory": ["/stock/check", "/stock/reserve", "/stock/release"],
 }
-LOG_LEVELS = ["INFO", "INFO", "INFO", "WARN", "ERROR"]
 
 fake = Faker()
 
@@ -123,12 +128,18 @@ def build_event(service: str, anomaly_state: AnomalyState) -> LogEvent:
             anomaly_injected=True,
         )
 
-    # Normal baseline traffic
+    # Normal baseline traffic. `level` is derived from `status_code` (not
+    # independently randomized) so downstream error-rate calculations are
+    # meaningful: a 200 response must never be logged as ERROR, and vice versa.
     latency_ms = round(max(5, random.gauss(120, 40)), 2)
     status_code = 200 if random.random() > 0.03 else random.choice([400, 404])
-    level = random.choice(LOG_LEVELS)
+    if status_code >= 500:
+        level = "ERROR"
+    elif status_code >= 400:
+        level = "WARN"
+    else:
+        level = "INFO"
     message = f"{service} handled {endpoint} for user {fake.uuid4()[:8]}"
-
     return LogEvent(
         event_id=str(uuid.uuid4()),
         timestamp=datetime.now(UTC).isoformat(),