diff --git a/docs/PREDICTIVE_RESOURCE_ALLOCATION.md b/docs/PREDICTIVE_RESOURCE_ALLOCATION.md
index a16e7f5..f6d8a6c 100644
--- a/docs/PREDICTIVE_RESOURCE_ALLOCATION.md
+++ b/docs/PREDICTIVE_RESOURCE_ALLOCATION.md
@@ -12,8 +12,8 @@ The predictor reads two data sources:
 2. **Heartbeat logs** (`heartbeat/ticks_*.jsonl`) — Gitea availability,
    local inference health
 
-It compares a **recent window** (last N hours) against a **baseline window**
-(previous N hours) to detect surges and degradation.
+It compares a **recent window** (last N hours of activity) against the **previous active window**
+(previous N hours ending at the most recent event before the current window) so sparse telemetry still yields a meaningful baseline.
 
 ## Output Contract
 
diff --git a/scripts/predictive_resource_allocator.py b/scripts/predictive_resource_allocator.py
index 29df0c5..4546dd2 100644
--- a/scripts/predictive_resource_allocator.py
+++ b/scripts/predictive_resource_allocator.py
@@ -90,13 +90,19 @@ def compute_rates(
 
     latest = max(_parse_ts(r["timestamp"]) for r in rows)
     recent_cutoff = latest - timedelta(hours=horizon_hours)
-    baseline_cutoff = latest - timedelta(hours=horizon_hours * 2)
 
     recent = [r for r in rows if _parse_ts(r["timestamp"]) >= recent_cutoff]
-    baseline = [
-        r for r in rows
-        if baseline_cutoff <= _parse_ts(r["timestamp"]) < recent_cutoff
-    ]
+
+    earlier = [r for r in rows if _parse_ts(r["timestamp"]) < recent_cutoff]
+    if earlier:
+        previous_latest = max(_parse_ts(r["timestamp"]) for r in earlier)
+        previous_cutoff = previous_latest - timedelta(hours=horizon_hours)
+        baseline = [
+            r for r in earlier
+            if _parse_ts(r["timestamp"]) >= previous_cutoff
+        ]
+    else:
+        baseline = []
 
     recent_rate = len(recent) / max(horizon_hours, 1)
     baseline_rate = (