Model · OpenAI
GPT-6 Luna
GPT-6 Luna2 bench decks · 5/5 problems correct on canonical boards · 5 audited cells.
methodology + notes
How to read. Cell scores are peak fraction of the board roofline (Hard / most CUDA), geomean milliseconds for CUDA Native Sparse Attention, or speedup vs the torch baseline (Mega). Each cell has an unlimited agent budget; interrupted or overlapping resumes are documented in the audit. Audit chips come from the human/subagent reward-hack review of every published cell; scores from flagged sessions render dimmed — they don't count toward the charts.
Board summary bars are each score relative to the best published model on that board (1.00 = board leader); the printed number is the bench-native score.
Methodology. Rank per bench: valid passes (audited-clean correct cells / problems) desc, Cells with no audit annotation are shown but never scored. then mean normalized performance over the FULL problem deck (cell score / board best per problem; fail/invalid/missing cells count as 0) desc. Hack badge = flagged audited sessions / total audited sessions for that model; flagged = annotation verdict reward_hack | contamination | rubric_leak | suspect, or megakernel_authentic false (mega). Verdicts come from per-run audit YAMLs, not static lint. Hack rate is displayed, never a sort key. Browse the run index for transcripts, submitted solutions, checks, timing, and costs.
RTX PRO 6000· canonical board
all audited codex attempts
Every attempted cell stays visible, including correctness failures, hardware mismatches, contaminated runs, and audit rejects. Only publishable results contribute to the board above.
Luna wrote one cooperative CUDA launch for the four-layer Kimi-Linear decode. It unpacks int4 weights inside custom GEMVs, parallelizes MLA cache reads across context chunks, and keeps KDA, RoPE, routed and shared MoE, and state updates inside the launch. The final isolated regrade passed the full six-case checker.
audited runRTX PRO 6000· canonical board
all audited codex attempts
Every attempted cell stays visible, including correctness failures, hardware mismatches, contaminated runs, and audit rejects. Only publishable results contribute to the board above.
Luna wrote CUDA token packing and weighted reduction around routed and always-on shared experts. PyTorch batched matrix multiplies perform the expert projections. Each forward allocates fresh routing and output buffers and computes from the current inputs and weights. The isolated regrade passed on tetra's RTX PRO 6000.
audited runLuna wrote a prefix-mean CUDA kernel and a sparse attention kernel that repeatedly selects the causal top eight blocks, merges the local window, and computes attention from the selected live keys and values. The kernels allocate new output and prefix buffers per call. The final isolated check passed all graded shapes and numeric stress cases.
audited runOne authored cooperative CUDA launch per decode step runs all four Qwen3 layers, including projection GEMVs, RoPE and cache append, chunked grouped-query attention, and the MLP. The prefill path executes the same kernel. Only input-independent RoPE tables are cached. The language gate passes CUDA evidence despite reporting framework unknown.
audited runLuna captured the Grid + MinGRU rollout in a CUDA graph with custom environment, observation, and MinGRU activation kernels. Autocast PyTorch matrix multiplies perform the gate projections. The graph keeps live parameter pointers and reinitializes agents, food, and RNG state on each replay.
audited run