{ "study_id": "beam-pi-programbench-20261009", "model": "Beam-501B-A23B", "space_id": "burtenshaw/beam-pi-programbench", "project": "beam-pi-programbench-20261009", "global_api_token_cap": 130000000, "category_caps": { "benchmark": 120000000, "validation": 10000000 }, "daily_headroom_tokens": 20000000, "max_inflight_requests": 5, "episode_wall_seconds": 7200, "snapshot_seconds": [ 30, 60, 120, 300, 600, 1200, 2400, 4800, 7200 ], "reference_prefill_tps": 10000, "reference_decode_tps": 265, "context_accounting": "estimated fresh input; full provider prompt at window transitions; actual completion tokens including reasoning", "tasks": [ { "instance_id": "burntsushi__ripgrep.3b7fd44" }, { "instance_id": "jqlang__jq.b33a763" }, { "instance_id": "burntsushi__xsv.f430466" }, { "instance_id": "cmatsuoka__figlet.202a0a8" }, { "instance_id": "sharkdp__hexyl.2e26437" } ], "runs": [ { "run_id": "xsv-single-r1", "task": "burntsushi__xsv.f430466", "condition": "single", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 1000000 }, { "run_id": "xsv-peers-r1", "task": "burntsushi__xsv.f430466", "condition": "peers", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 200000 }, { "run_id": "xsv-async-r1", "task": "burntsushi__xsv.f430466", "condition": "async", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 200000 }, { "run_id": "hexyl-peers-r1", "task": "sharkdp__hexyl.2e26437", "condition": "peers", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 200000 }, { "run_id": "hexyl-async-r1", "task": "sharkdp__hexyl.2e26437", "condition": "async", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 200000 }, { "run_id": "hexyl-single-r1", "task": "sharkdp__hexyl.2e26437", "condition": "single", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 1000000 }, { "run_id": "ripgrep-async-r1", "task": "burntsushi__ripgrep.3b7fd44", "condition": "async", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 200000 }, { "run_id": "ripgrep-single-r1", "task": "burntsushi__ripgrep.3b7fd44", "condition": "single", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 1000000 }, { "run_id": "ripgrep-peers-r1", "task": "burntsushi__ripgrep.3b7fd44", "condition": "peers", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 200000 }, { "run_id": "figlet-single-r1", "task": "cmatsuoka__figlet.202a0a8", "condition": "single", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 1000000 }, { "run_id": "figlet-peers-r1", "task": "cmatsuoka__figlet.202a0a8", "condition": "peers", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 200000 }, { "run_id": "figlet-async-r1", "task": "cmatsuoka__figlet.202a0a8", "condition": "async", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 200000 }, { "run_id": "jq-peers-r1", "task": "jqlang__jq.b33a763", "condition": "peers", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 200000 }, { "run_id": "jq-async-r1", "task": "jqlang__jq.b33a763", "condition": "async", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 200000 }, { "run_id": "jq-single-r1", "task": "jqlang__jq.b33a763", "condition": "single", "category": "benchmark", "api_token_cap": 8000000, "context_token_cap": 1000000, "agent_context_token_cap": 1000000 }, { "run_id": "validation-single", "task": "synthetic-tool-and-git-validation", "condition": "single", "category": "validation", "api_token_cap": 3000000, "context_token_cap": 100000, "agent_context_token_cap": 100000 }, { "run_id": "validation-peers", "task": "synthetic-tool-and-git-validation", "condition": "peers", "category": "validation", "api_token_cap": 3000000, "context_token_cap": 100000, "agent_context_token_cap": 50000 }, { "run_id": "validation-async", "task": "synthetic-tool-and-git-validation", "condition": "async", "category": "validation", "api_token_cap": 3000000, "context_token_cap": 100000, "agent_context_token_cap": 50000 } ] }