AR to diffusion conversion

After grading, a reviewer model reads the submission's files and the agent's trajectory and checks each of the task's rules. A failed check makes the score 0. Step numbers link into the trajectory.

LLM adversarial review is not proof of arbitrary program behavior. Host storage and task-specific deterministic checks remain required.

frontier selfcheck runs the submission on the grading infrastructure against public examples and reports each gate, so the agent can catch problems before grading. Each run gets four. This agent used 4 of 4.

1. 1
   passed7h 31m into the runtook 3m 45s

   startup 86.5 s of 300 s; 1/1 gates passed; The suite's 32 concurrent requests per copy, up to 3600 s each, were not sent; one request at its settings was.

   5 gates

   • ✓review prerequisites
   • ✓startup · GET /health answered 200 after 86.5 s
   • ✓single-stream speed · 8 of 8 measured prompts, sent one at a time after an untimed warm-up (max_tokens 1024, temperature 1), each reply's reasoning and answer tokens counted with the pinned openbmb/MiniCPM5-2B tokenizer over end-to-end request time (sandbox clock), a miss re-sent once; slowest 525 tokens/s (249 tokens in 0.47 s); per prompt: 978, 525, 720, 1223, 853, 1042, 924, 1011. Public prompts: 7 GPQA and 2 hard LiveCodeBench; grading draws 4 GPQA, 3 level-5 MATH-500 (held out) and 2 hard LiveCodeBench, so its replies differ.
   • ✓GPQA request as graded (self-check cap 291 s; grading allows 3600 s) · {"finish_reason": "stop", "answer": null, "note": "no answer letter found (ANSWER: X on its own line): grading scores it wrong", "content_characters": 539, "content_end": "mpute the relative entropy measure.\n\nNow, we need to calculate the nG for phi = -pi/4 and alpha = 0.5.\n\nANSWER: answer\n\n"}
   • ✓healthy after requests
2. 2
   passed7h 39m into the runtook 3m 18s

   startup 80.5 s of 300 s; 1/1 gates passed; The suite's 32 concurrent requests per copy, up to 3600 s each, were not sent; one request at its settings was.

   5 gates

   • ✓review prerequisites
   • ✓startup · GET /health answered 200 after 80.5 s
   • ✓single-stream speed · 8 of 8 measured prompts, sent one at a time after an untimed warm-up (max_tokens 1024, temperature 1), each reply's reasoning and answer tokens counted with the pinned openbmb/MiniCPM5-2B tokenizer over end-to-end request time (sandbox clock), a miss re-sent once; slowest 566 tokens/s (133 tokens in 0.24 s); per prompt: 1673, 1087, 566, 1081, 928, 690, 1062, 1127. Public prompts: 7 GPQA and 2 hard LiveCodeBench; grading draws 4 GPQA, 3 level-5 MATH-500 (held out) and 2 hard LiveCodeBench, so its replies differ.
   • ✓GPQA request as graded (self-check cap 291 s; grading allows 3600 s) · {"finish_reason": "stop", "answer": null, "note": "no answer letter found (ANSWER: X on its own line): grading scores it wrong", "content_characters": 668, "content_end": ", 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2"}
   • ✓healthy after requests
3. 3
   passed7h 44m into the runtook 4m 51s

   startup 105.7 s of 300 s; 1/1 gates passed; The suite's 32 concurrent requests per copy, up to 3600 s each, were not sent; one request at its settings was.

   5 gates

   • ✓review prerequisites
   • ✓startup · GET /health answered 200 after 105.7 s
   • ✓single-stream speed · 8 of 8 measured prompts, sent one at a time after an untimed warm-up (max_tokens 1024, temperature 1), each reply's reasoning and answer tokens counted with the pinned openbmb/MiniCPM5-2B tokenizer over end-to-end request time (sandbox clock), a miss re-sent once; slowest 584 tokens/s (251 tokens in 0.43 s); per prompt: 835, 1221, 794, 1059, 972, 584, 1002, 1249. Public prompts: 7 GPQA and 2 hard LiveCodeBench; grading draws 4 GPQA, 3 level-5 MATH-500 (held out) and 2 hard LiveCodeBench, so its replies differ.
   • ✓GPQA request as graded (self-check cap 291 s; grading allows 3600 s) · {"finish_reason": "stop", "answer": null, "note": "no answer letter found (ANSWER: X on its own line): grading scores it wrong", "content_characters": 378, "content_end": "\nArrange the substances in order of increasing the weight fraction of the yield of the para-isomer:\nOH < Cl < CO < NO <2"}
   • ✓healthy after requests
4. 4
   passed7h 52m into the runtook 4m 15s

   startup 86.1 s of 300 s; 1/1 gates passed; The suite's 32 concurrent requests per copy, up to 3600 s each, were not sent; one request at its settings was.

   5 gates

   • ✓review prerequisites
   • ✓startup · GET /health answered 200 after 86.1 s
   • ✓single-stream speed · 8 of 8 measured prompts, sent one at a time after an untimed warm-up (max_tokens 1024, temperature 1), each reply's reasoning and answer tokens counted with the pinned openbmb/MiniCPM5-2B tokenizer over end-to-end request time (sandbox clock), a miss re-sent once; slowest 688 tokens/s (250 tokens in 0.36 s); per prompt: 688, 1186, 1470, 791, 1241, 1042, 1160, 1888. Public prompts: 7 GPQA and 2 hard LiveCodeBench; grading draws 4 GPQA, 3 level-5 MATH-500 (held out) and 2 hard LiveCodeBench, so its replies differ.
   • ✓GPQA request as graded (self-check cap 291 s; grading allows 3600 s) · {"finish_reason": "stop", "answer": "A", "content_characters": 402, "content_end": " CH3CH2C(H)(CH3)C(H)(CH3)COOH has a doublet of of triplets of quartets and a doublet of triplets of quartets.\n\nANSWER: A"}
   • ✓healthy after requests

The agent's last message before its session ended.