| R1/R2 functional checks |
native preflight |
covered |
R2 native preflight passed 6/6 |
none |
| disabled-thinking support |
baseline failure, candidate failure |
rejected |
default/max thinking is the matched operating mode |
both profiles leaked reasoning into strict visible-output checks |
| Oracle |
R2 native rounds, restored native rounds |
covered |
R2 and fresh restored-original controls both passed 54/54 across three rounds; R1 failed strict 17/18 |
no causal or performance comparison |
| vision and context |
R2 vision, restored vision, R2 context, restored context |
covered |
R2 and fresh restored-original controls passed vision 18/18 and separate retrieval/tool-context probes at 216307/110760 actual tokens |
no full-window C4 claim |
| coding quality |
R2 native SWE, restored native SWE, independent control review |
rejected |
R2 and restored original each resolved 3/5, below the 4/5 gate |
matching resolved cases do not establish equivalent detailed grader failures |
| reserve |
R2 sampled reserve, restored telemetry |
covered for R2 only |
R2 sampled minimum free VRAM stayed above its 4096 MiB challenger gate |
200 ms sampling only; restored original minimum was 1725/1745 MiB under its retained historical waiver and is not assessed against the challenger gate |
| capacity, stability, DFlash, performance |
none |
not run |
none |
coding gate failed |
| exact restoration |
restoration, final status, primary router receipt |
covered |
exact incumbent identity, direct/routed 6/6 acceptance, zero OOM, and primary router readmission restored |
separate secondary tier remained unavailable; it is not a primary-control failure |