GPT-6 Astra Pro

OpenAI gpt-6-astra-pro through OpenRouter (codex harness, CLI 0.140.0), xhigh effort, unlimited budget, RTX PRO 6000: a single cooperative persistent megakernel using classic grid.sync() phase barriers, scalar-FMA int4 dequant-GEMVs, and wmma tensor cores reserved for the absorbed-MLA attention. Post-session trusted grade on the same box 2026-09-06 (check then benchmark via trusted_entrypoint.py after the agent exited, gpu_lock.log lines 148-153): 24.8188x geomean over the untouched baseline (23.65x / 24.95x / 25.91x at ctx 2048 / 8192 / 16384, 0.323-0.358 ms/tok); the agent's final in-run self-measurement was 24.7390. check.log shows out/S/cache = 1.0000 on all six seed/ctx combos against the 0.98 gate. All 8 template files byte-identical, zero foreign reads (bwrap masked outputs/runs and DEVLOG.md), no clock commands, one ninja install (pip failed, uv pip succeeded), zero lock contention, no key strings in the archive. Session ended on a deliberate task_complete at 5h05m after three transient CLI reconnects (all recovered). Audit-time isolated sequential regrade on the same box 2026-09-07 (clocks reset with nvidia-smi -rgc, sole GPU owner): 24.8043x geomean (23.66x / 24.95x / 25.85x, 0.323 / 0.336 / 0.359 ms/tok), -0.06% vs the 24.8188 post-session grade, which is kept as benchmark.contended.log. Overwrite probe (probe.log): primed step cos(ref,sol)=1.0000; a continuation step with the hidden buffer overwritten in place at the same data_ptr changes the output (cos(out1,out2)=0.0335) and matches the reference at 1.0000; a fresh state at seed+1 with the buffer overwritten again gives cos(out2,out3)=-0.0342 and matches at 1.0000. OpenRouter cost $83.75.

GPT-6 Astra Pro through OpenRouter (codex CLI 0.140.0), xhigh effort, unlimited budget, RTX PRO 6000, 1h33m session to a voluntary stop after PASS + RESULT OK, two context compactions and three recovered OpenRouter reconnects: a hand-rolled SM120 grouped MoE — mma.sync m16n8k16 bf16 with fp32 accumulate, ldmatrix + XOR-swizzled smem, cp.async 2-stage pipeline, hist/scan/scatter token packing (cub BlockScan), gate/up fused in one GEMM with SiLU*up in registers, fp32 atomicAdd top-k weighted combine into an L2-resident buffer, and a CUDA-core GEMV decode path for T<=4. Dispatch: 64x64x64/256-thread tiles, switching to 128x128x64/512-thread with B-fragment preload at tokens*topk >= 192*E (T>=6144). Graded 0.0946 geomean (per shape 0.3443 / 0.3453 / 0.0024 / 0.4720 / 0.0524 / 0.0998); the agent's own final official run also said 0.0946 and per-shape latencies match the trusted grade within 0.1 ms. Audit-time isolated sequential regrade on the same box 2026-09-07 (clocks reset, sole GPU owner): 0.0945 (per shape 0.3439 / 0.3454 / 0.0024 / 0.4718 / 0.0524 / 0.0998; ms 10.79 / 10.83 / 0.372 / 15.73 / 8.85 / 9.08), the 0.0946 post-session grade kept as benchmark.contended.log. Overwrite probe (probe.log): primed cos(ref,sol)=1.0000; in-place input overwrite at the same data_ptr changes the output (cos(out1,out2)=0.0004) and matches the reference at 1.0000; in-place weight overwrite on the same parameter buffers changes it again (cos(out2,out3)=0.0000) and matches at 1.0000; fresh inputs 1.0000. OpenRouter cost $41.68. All template files byte-identical, no foreign run reads, no clock commands (read-only nvidia-smi queries), six NVIDIA-docs web searches, one uv install of ninja into the bench venv. Reads of src/eval grader source and bench AGENTS.md/SPEC.md were documentation-only. 268k output tokens, 12.1M input (10.3M cached). Sixth on the problem behind fable-5-1 0.1017 (whose YAML names or-opus 0.1072 as leader).

GPT-6 Astra Pro through OpenRouter (codex CLI 0.140.0), xhigh effort, unlimited budget, RTX PRO 6000, 2h22m to a voluntary stop. Two-kernel query-major NSA in CUDA with inline PTX: key_means writes fp32 per-position causal prefix means plus block means; tensor_attention (one CTA per 64-query block per head) re-runs selection per CTA - tf32 MMA over a two-part split of the fp32 mean keys (~20-bit ranking), exact fp32 prefix dot for the home block, 8-round argmax top-8 with the reference (imp, block-id) tie-break - unions the 4 warps' masks with the window blocks in smem, then cp.async double-buffered bf16 mma.m16n8k16 online softmax over the union tiles. Selection is entirely in CUDA; the only torch on the path is empty allocations. The check.py blind spot (S<=384 selects every block; the blocks>8 selection path is never hit by the trusted check) is covered by the agent's own tests: reference oracle at S=577/641/1055 and a vectorized oracle with the exact tie-break on all six deck shapes at scales 0.01/1.0/8.0, worst max_abs 0.125 at scale 8, plus memcheck/racecheck clean. Templates byte-identical, no foreign reads, no clock commands, doc-only web searches, one OpenRouter reconnect pair recovered at ~23:52. In-session final benchmark 0.3944; the trusted post-session isolated regrade on a quiet GPU gives 0.3941 (per shape 0.2932/0.4597/0.5438/0.6156/0.2169/ 0.3829; ms 0.117/0.303/0.505/0.893/0.079/0.193). Audit-time isolated sequential regrade 2026-09-07 (clocks reset, sole GPU owner): 0.3946 (per shape 0.2899/0.4635/0.5397/0.6164/0.2188/0.3860; ms 0.119/0.301/0.509/0.892/0.079/0.191), the 0.3941 grade kept as benchmark.contended.log. Overwrite probe (probe.log): primed 1.0000; in-place q/k/v overwrite at the same data_ptr changes the output (cos(out1,out2)=-0.0041) and matches the reference at 1.0000; the weight-overwrite step is a no-op at cos 1.0000 because the reference Model has no parameters; fresh inputs 1.0000. Long-context probe past the check.py select-everything regime (probe_long.log): S=1024/1500/2048 x D=64/128 with seeds 42/123/7/991 against reference.nsa_attend, all six ok at cos 0.999998 and max abs diff <= 0.0081 against the 0.1 bf16 gate, LONG_CTX_OK. OpenRouter cost $55.54. On this problem the run trails claude-fable-5-1 (1.0627), claude-opus-5 (1.0367) and claude-fable-5 (0.7266) - the gap is design, not hygiene: Fable reads each selected K/V block once per head (key-major gather with flash-decoding partials) and replays three launches as a CUDA graph, while this kernel re-reads every selected block per query tile and re-runs selection per CTA, which is why the deficit widens from 2.0x at S=2048 to 4.6x at S=8192. Cost not reported (OpenRouter via codex, total_cost_usd null).

GPT-6 Astra Pro through OpenRouter (codex CLI 0.140.0 harness), xhigh effort, unlimited budget, RTX PRO 6000, 1h53m to a voluntary stop. One persistent cooperative megakernel: 376 blocks (2 per SM) x 256 threads, all decode steps x all four layers in a single cudaLaunchCooperativeKernel, six grid.sync barriers per layer (QKV | attention partials | cross-split combine | O-proj+residual | gate/up | down+residual). The 125.8 MB bf16 weight set is streamed with plain __ldg (no eviction policy); the KV cache streams evict-first via inline-PTX ld.global.cs. Attention is true split-KV flash-decoding: 47 splits per kv head with exact online softmax (running max + rescale, no fixed-shift trick), fp32 partials in global workspace, exact 16-block cross-split combine; ctx 131072 uses the same path. RMSNorm/RoPE/SwiGLU fully fused; residuals fp32 in-layer, bf16 between layers; prefill is real, run through the same kernel. Fresh h_out (clone) and fresh seeded noise every call; engine caches hold only weight/cache data_ptr tables, RoPE table, and workspace, and weights are re-read from GPU memory every launch. Agent verified batched torch.randn equals the reference's per-step draw before relying on it. In-run grade 0.0488; isolated sequential trusted grade after session end 0.0492 (+0.8%): 6,545 tok/s at ctx 2048 down to 634 tok/s at ctx 131072. Audit-time isolated regrade 2026-09-07 (clocks reset, sole GPU owner): 0.0490 (6,430 / 4,435 / 2,001 / 634 tok/s at ctx 2048 / 8192 / 32768 / 131072; per shape 0.1286 / 0.0887 / 0.0400 / 0.0127), the 0.0492 grade kept as benchmark.contended.log. Probe (probe.log): primed ctx 2048 dec 16 cos(ref,sol)=1.0000; a second decode_steps call with seed+1 on the same model gives a different output (cos(s1,s1b)=0.3321) that matches its own reference at 1.0000; an in-place weight overwrite on the same parameter buffers flips the output (cos(s1,s2)=-0.0194) and matches the new reference at 1.0000. OpenRouter cost $53.83. All template files byte-identical, patches only to solution.py, no foreign reads, no clock commands, a failed `pip install ninja` (nothing installed), and sanctioned fetching of the MegaQwen baseline repo only. Third on the problem behind claude-opus-5 0.0655 and claude-fable-5-1 0.0643; the gap is the two short shapes, where 24 cooperative grid syncs per token over 376 blocks lose to the leaders' 5-barrier/188-block designs.

GPT-6 Astra Pro through OpenRouter (codex CLI 0.140.0 harness), xhigh effort, unlimited budget, RTX PRO 6000, 2h01m to a voluntary stop, 23 solution.py revisions from a first PASS benchmark at 0.2732 (03:08Z) to 0.6836 self-measured (04:50Z); trusted post-session grade 0.6834; audit-time isolated regrade 2026-09-07 (clocks reset, sole GPU owner) 0.6837 (per shape 0.5306 / 0.7774 / 0.6950 / 0.7624; 79.6M / 116.6M / 104.2M / 114.4M sps), the 0.6834 grade kept as benchmark.contended.log. load_inline CUDA extension with two graded rollout paths: above 4096 envs, one fused_policy<48,16> launch per step (48 envs per 512-thread block, MinGRU state round-tripping through a global in-place buffer, the kernel boundary serving as the grid barrier the LCG any-hit rule needs); at or below 4096, a cooperative persistent_policy<32,16> with grid.sync() per step and state held in registers/shared (128 blocks on the 188-SM card for shape 0). Layer 1 folded algebraically into a 768x5 composed encoder recomputed in fp32 on every call; layers 2 and 3 run mma.sync.m16n8k16 with split-fp16 hi+lo operands, three MMAs per tile; heads as fp32 warp-reduction dot products; greedy argmax keeps the first index on ties like torch.argmax. Deferred one-step LCG respawn via a flags[] array honors the reference any-hit rule exactly. No caching anywhere: weights re-packed every run(), fresh outputs and a fresh CPU-randint init every call. Agent's own extended validation: positions and rewards bit-exact vs reference over 18 in-place weight draws at odd env counts (1 to 4609), max abs logit error 1.7e-8 against the 1e-3 gate. Audit probe at benchmark scale (probe.log; check.py itself only runs 128 envs): all four deck shapes at seeds 42 and 123 (not benchmark.py's 2026+trial seeds), positions bit-exact for every env (0 diverged of 4096 / 16384 / 65536 / 8192), rewards equal, max abs logit error 1.9e-8, exercising both the persistent path (4096 envs) and the per-step-launch path (8192 / 16384 / 65536); then an in-place weight overwrite on the same parameter buffers and a re-run: positions bit-exact and logits cos 1.00000 against the new reference, ALL_POSITIONS_EQUAL. OpenRouter cost $53.87. Templates byte-identical, no foreign archive reads, three nvidia-smi query-only calls, nine provider-side web searches all against docs.nvidia.com PTX pages. Third on the RTX PRO 6000 board for this problem behind claude-opus-5 (1.961) and claude-fable-5-1 (0.7093), ahead of gemini-3.8-flash-high (0.3637); the geomean deficit to Fable is entirely shape 0 (0.5275 vs 0.6726), while shapes 1 and 3 beat Fable (0.7781 vs 0.7036 and 0.7618 vs 0.7433).