Confusion Matrix
Top-15 most frequent misclassifications on 468 valid predictions. 83 hallucination/parse failures excluded.
| Actual DCA | Predicted DCA | Count | |
|---|---|---|---|
121RIGHT THROUGH | -> | 110CROSS TRAFFIC(INTERSECTIONS ONLY) | 42 |
130REAR END(VEHICLES IN SAME LANE) | -> | 110CROSS TRAFFIC(INTERSECTIONS ONLY) | 30 |
113RIGHT NEAR (INTERSECTIONS ONLY) | -> | 110CROSS TRAFFIC(INTERSECTIONS ONLY) | 21 |
120HEAD ON (NOT OVERTAKING) | -> | 130REAR END(VEHICLES IN SAME LANE) | 11 |
160VEHICLE COLLIDES WITH VEHICLE PARKED ON LEFT OF ROAD | -> | 130REAR END(VEHICLES IN SAME LANE) | 11 |
147VEHICLE STRIKES ANOTHER VEH WHILE EMERGING FROM DRIVEWAY | -> | 130REAR END(VEHICLES IN SAME LANE) | 9 |
132RIGHT REAR. | -> | 110CROSS TRAFFIC(INTERSECTIONS ONLY) | 8 |
130REAR END(VEHICLES IN SAME LANE) | -> | 120HEAD ON (NOT OVERTAKING) | 6 |
131LEFT REAR | -> | 110CROSS TRAFFIC(INTERSECTIONS ONLY) | 5 |
171LEFT OFF CARRIAGEWAY INTO OBJECT/PARKED VEHICLE | -> | 175OFF END OF ROAD/T-INTERSECTION. | 5 |
133LANE SIDE SWIPE (VEHICLES IN PARALLEL LANES) | -> | 130REAR END(VEHICLES IN SAME LANE) | 5 |
173RIGHT OFF CARRIAGEWAY INTO OBJECT/PARKED VEHICLE | -> | 166STRUCK OBJECT ON CARRIAGEWAY | 5 |
133LANE SIDE SWIPE (VEHICLES IN PARALLEL LANES) | -> | 110CROSS TRAFFIC(INTERSECTIONS ONLY) | 4 |
184OUT OF CONTROL ON CARRIAGEWAY (ON BEND) | -> | 174OUT OF CONTROL ON CARRIAGEWAY (ON STRAIGHT) | 4 |
180OFF CARRIAGEWAY ON RIGHT BEND | -> | 174OUT OF CONTROL ON CARRIAGEWAY (ON STRAIGHT) | 4 |
Key finding: 121 ->110 is the single largest confusion pair (42 cases). The model shows strong bias toward DCA 110 (cross-traffic / intersection) -- the most frequent category in the training distribution. This is a calibration failure, not just a coverage gap.