Shadow Recall
97.6%
Retrieves knowledge in
8 tool calls
Blind Bug Hunting
+25.4
percentage points vs baseline
20 repos × 100 synthetic bugs · wins
15/20
50 real bugs:
88%
right module,
22%
exact function
No problem statement, no hints
Feature Ideation
+10.4
percentage points alignment
Insight
+0.40
User Impact
+0.24
Anticipates
shipped features