Evidence record
KernelGen 1P
Astra card snapshot, September 2026 · GPT-6 Luna · 2026-09-03
Reported result · numeric
21.83%
Mean kernel optimization reward · percent
lab reported · extraction review: agent checked
Source and extraction
Published 2026-09-03
- GPT-6 Astra System CardAppendix A.8.1.3.2, Figure 85 (PDF printed page 152); exact bar labels; section 10.1.3.2Original source ↗
Evaluation setup
| aggregate method | Mean rubric reward |
|---|---|
| hardware | OpenAI first-party hardware; exact hardware not disclosed |
| tool access | Kernel development environment, benchmark harness, reference materials, performance and correctness tests |
| comparability caveats | Per-model output-token budgets for the appendix bars are not printed. Figure 53 plots separate budget-dependent results. |
Not reported: split, task count, task snapshot, attempts per task, run count, selection rule, token budget, wall clock budget, training budget, inference budget, monetary cost, internet access, filtering, scaffold, evaluator version, human intervention, task exclusions, contamination concerns.
Comparability
Not compared with other results.
Limitations
- Per-model output-token budgets for the appendix bars are not printed. Figure 53 plots separate budget-dependent results.
- Internal suite; no cross-lab percentage comparison.
- Not a demonstration of recursive improvement.