200 instances were attempted in four 50-instance chunks. The fair comparison below uses 162 paired instances after excluding native infra/pre-agent/container failures and manual generation/run outlier pairs from both arms.
For 150:200, token totals were recomputed from nested per-response usage records in the trajectory JSON files because aggregate model stats only recorded API calls.
| Chunk | Fair pairs | Baseline solved | GitNexus solved | Solved delta | Token delta | API call delta |
|---|---|---|---|---|---|---|
| 0:50 | 47 | 25/47 | 28/47 | +3 | 27.67% fewer | 17.20% fewer |
| 50:100 | 45 | 35/45 | 36/45 | +1 | 25.02% fewer | 9.18% fewer |
| 100:150 | 44 | 34/44 | 31/44 | -3 | 34.49% fewer | 12.29% fewer |
| 150:200 | 26 | 22/26 | 21/26 | -1 | 38.11% fewer | 15.25% fewer |
| Combined | 162 | 116/162 | 116/162 | +0 | 30.45% fewer | 13.26% fewer |
Excluded from both arms: astropy__astropy-14508, django__django-11149, django__django-11239, django__django-11749, django__django-12406, django__django-13033, django__django-13279, django__django-13343, django__django-13401, django__django-13513, django__django-13590, django__django-13786, django__django-11734, django__django-13112, django__django-15098, django__django-15103, django__django-15104, django__django-15127, django__django-15128, django__django-15161, django__django-15252, django__django-15268, django__django-15277, django__django-15278, django__django-15280, django__django-15315, django__django-15368, django__django-15375, django__django-15380, django__django-15382, django__django-15467, django__django-15499, django__django-15503, django__django-15525, django__django-15554, django__django-15561, django__django-15563, django__django-15569. Manual generation/run outlier exclusions: django__django-11734, django__django-13112.