LLM post-training

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 3 of 4.

1. 1
   passed4h 24m into the runtook 9m 35s

   startup 109.4 s of 300 s; 1/1 gates passed; One wave of 32 concurrent public requests stood for the suite's 3,153 graded ones at that concurrency; MATH-500 and MMLU-Pro are held out.

   7 gates

   • ✓review prerequisites
   • ✓checkpoint audit
   • ✓startup · GET /health answered 200 after 109.4 s
   • ✓GPQA request as graded · {"finish_reason": "stop", "answer": "A", "content_characters": 11, "reasoning_characters": 29521, "content_end": "\n\nANSWER: A"}
   • ✓slowest of 32 concurrent requests · 32 public requests at once (16 GPQA, 16 hard LiveCodeBench) as graded (max_tokens 16384, temperature 0.7, 250 s each): slowest 216.1 s, median 186.9 s; 0 stopped at max_tokens, 0 timed out. Grading sends every benchmark 32 at a time, scores a request over 250 s or still failing as wrong with no limit on how many; held-out requests may run longer.
   • ✓concurrent requests answered · 32 well-formed replies
   • ✓healthy after requests
2. 2
   passed7h 11m into the runtook 9m 52s

   startup 108.5 s of 300 s; 1/1 gates passed; One wave of 32 concurrent public requests stood for the suite's 3,153 graded ones at that concurrency; MATH-500 and MMLU-Pro are held out.

   7 gates

   • ✓review prerequisites
   • ✓checkpoint audit
   • ✓startup · GET /health answered 200 after 108.5 s
   • ✓GPQA request as graded · {"finish_reason": "stop", "answer": "B", "content_characters": 840, "reasoning_characters": 31254, "content_end": ".\n\\]\n\nThus the decay proceeds via \\(|3,0,0\\rangle\\to|2,1,0\\rangle\\to|1,0,0\\rangle\\) with probability \\(1/3\\).\n\nANSWER: B"}
   • ✓slowest of 32 concurrent requests · 32 public requests at once (16 GPQA, 16 hard LiveCodeBench) as graded (max_tokens 16384, temperature 0.7, 250 s each): slowest 224.4 s, median 180.9 s; 3 stopped at max_tokens, 0 timed out. Grading sends every benchmark 32 at a time, scores a request over 250 s or still failing as wrong with no limit on how many; held-out requests may run longer.
   • ✓concurrent requests answered · 32 well-formed replies
   • ✓healthy after requests
3. 3
   passed8h 11m into the runtook 9m 2s

   startup 115.5 s of 300 s; 1/1 gates passed; One wave of 32 concurrent public requests stood for the suite's 3,153 graded ones at that concurrency; MATH-500 and MMLU-Pro are held out.

   7 gates

   • ✓review prerequisites
   • ✓checkpoint audit
   • ✓startup · GET /health answered 200 after 115.5 s
   • ✓GPQA request as graded · {"finish_reason": "stop", "answer": "C", "content_characters": 11, "reasoning_characters": 24541, "content_end": "\n\nANSWER: C"}
   • ✓slowest of 32 concurrent requests · 32 public requests at once (16 GPQA, 16 hard LiveCodeBench) as graded (max_tokens 16384, temperature 0.7, 250 s each): slowest 220.6 s, median 179.1 s; 0 stopped at max_tokens, 0 timed out. Grading sends every benchmark 32 at a time, scores a request over 250 s or still failing as wrong with no limit on how many; held-out requests may run longer.
   • ✓concurrent requests answered · 32 well-formed replies
   • ✓healthy after requests

The agent's last message before its session ended.