{"row_id": "S-Q2/speed/a1/run-001", "arm": "S-Q2", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "Q2_0", "file_sha256": {"1": "69820c02ec7d0b45ef2ebb19d6620299db749fe2aded7f39f93c6b88b199b720", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-8fc40dde/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-Q2/speed/a1/strata-config.json", "--port", "withheld-2026-10-13"], "ctx": "4K", "repeat": 0, "run": 1, "slice_sha256": "9c676dc1959897277fda62a2d77d01fc06c932a2e04641c753a24bde3e7d4a16", "prompt_sha256": "b42c6a181c3e02478e7e033585c5962f35eb046b680c0ae837ad8706a3c0674b", "nonce": "Ticket 85822155.", "requests_since_start": 1, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 116.204, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 115.5, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 249, "content_frames": 248, "frames": 250, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2054.2, "prompt_tok_s_client": 2019.0, "ttft_ms": 1999.962, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1965.7, "prompt_per_token_ms": 0.487, "prompt_per_second": 2054.2, "predicted_n": 256, "predicted_ms": 2217.2, "predicted_per_token_ms": 8.661, "predicted_per_second": 115.5, "draft_n": 210, "draft_n_accepted": 159}, "short_completion": false, "cache_flag": null, "completion_sha256": "93fbe774adcc88b5c1a68b48148c610a99c7b6cc9e940f08f410d79747ca5eb4", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9540000000000001, "hits": 138234, "lookups": 144850, "drafts_accepted": 159, "drafts_offered": 210, "engine_generated": 256, "engine_decode_tok_s": 115.5, "engine_prompt_tok_s": 2054.2, "reused": 0, "metrics": {"hit_rate": 0.954, "decode_tok_s": 115.5, "engine_generated": 256, "prompt_ms": 1965.7, "decode_ms": 2217.2, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.639, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1790988043.871, "t1": 1790988048.066, "seconds": 4.194, "host_samples": 4, "gpu_samples": 4}, "board": {"joules": 314.4, "mean_w": 104.55, "max_w": 133.18, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 176.56, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 3.0}, "data_reads_per_s": {"mean": 12401.5, "max": 47636.1}, "data_read_mb_per_s": {"mean": 52.08}, "temps": {"gpu_max_c": 53.0, "ssd_max_c": 52.85}, "vram": {"max_used_mib": 23809.0, "baseline_mib": 15.0, "growth_mib": 23794.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 38915321856, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 2.03, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 16.644, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.02, "unit_psi_full_avg10_max": 0.0, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 214}, "power_limit": {"min_w": 150.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-Q2/speed/a1/run-002", "arm": "S-Q2", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "Q2_0", "file_sha256": {"1": "69820c02ec7d0b45ef2ebb19d6620299db749fe2aded7f39f93c6b88b199b720", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-8fc40dde/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-Q2/speed/a1/strata-config.json", "--port", "withheld-2026-10-13"], "ctx": "4K", "repeat": 1, "run": 2, "slice_sha256": "9c676dc1959897277fda62a2d77d01fc06c932a2e04641c753a24bde3e7d4a16", "prompt_sha256": "9744fee0d5b25fee2aef36ad0bc7c5c4611f73d11f4e6bd9e5b84b17d93d319b", "nonce": "Ticket 42865306.", "requests_since_start": 2, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 131.504, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 131.2, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 249, "content_frames": 248, "frames": 250, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2354.1, "prompt_tok_s_client": 2313.0, "ttft_ms": 1745.757, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1715.3, "prompt_per_token_ms": 0.425, "prompt_per_second": 2354.1, "predicted_n": 256, "predicted_ms": 1951.4, "predicted_per_token_ms": 7.623, "predicted_per_second": 131.2, "draft_n": 205, "draft_n_accepted": 158}, "short_completion": false, "cache_flag": null, "completion_sha256": "c88f73b22047e4f40e5cff6efa81484c3d78a90db07ee2707146c7dcd9fc7406", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.983, "hits": 141814, "lookups": 144301, "drafts_accepted": 158, "drafts_offered": 205, "engine_generated": 256, "engine_decode_tok_s": 131.2, "engine_prompt_tok_s": 2354.1, "reused": 0, "metrics": {"hit_rate": 0.983, "decode_tok_s": 131.2, "engine_generated": 256, "prompt_ms": 1715.3, "decode_ms": 1951.4, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.612, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1790988048.067, "t1": 1790988051.752, "seconds": 3.685, "host_samples": 4, "gpu_samples": 4}, "board": {"joules": 363.45, "mean_w": 121.1, "max_w": 140.24, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 197.51, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 3.0}, "data_reads_per_s": {"mean": 459.9, "max": 1623.7}, "data_read_mb_per_s": {"mean": 1.92}, "temps": {"gpu_max_c": 53.0, "ssd_max_c": 51.85}, "vram": {"max_used_mib": 23811.0, "baseline_mib": 15.0, "growth_mib": 23796.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 38915321856, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 2.12, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 16.599, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.01, "unit_psi_full_avg10_max": 0.0, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 5}, "power_limit": {"min_w": 137.87, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-Q2/speed/a1/run-003", "arm": "S-Q2", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "Q2_0", "file_sha256": {"1": "69820c02ec7d0b45ef2ebb19d6620299db749fe2aded7f39f93c6b88b199b720", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-8fc40dde/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-Q2/speed/a1/strata-config.json", "--port", "withheld-2026-10-13"], "ctx": "4K", "repeat": 2, "run": 3, "slice_sha256": "9c676dc1959897277fda62a2d77d01fc06c932a2e04641c753a24bde3e7d4a16", "prompt_sha256": "214ec62ff2b1550ea59fc2fa7cc264e79ceed8841f5fd91e1709a8b73be2fa8f", "nonce": "Ticket 90569497.", "requests_since_start": 3, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 126.081, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 125.8, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 248, "content_frames": 247, "frames": 249, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2363.6, "prompt_tok_s_client": 2334.9, "ttft_ms": 1729.432, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1708.4, "prompt_per_token_ms": 0.423, "prompt_per_second": 2363.6, "predicted_n": 256, "predicted_ms": 2034.3, "predicted_per_token_ms": 7.946, "predicted_per_second": 125.8, "draft_n": 226, "draft_n_accepted": 154}, "short_completion": false, "cache_flag": null, "completion_sha256": "20c4194dc6a84c1a2bb645c55d8b2b90c7c7e7037a55de0c70ee9cc9de5fb002", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.985, "hits": 154104, "lookups": 156468, "drafts_accepted": 154, "drafts_offered": 226, "engine_generated": 256, "engine_decode_tok_s": 125.8, "engine_prompt_tok_s": 2363.6, "reused": 0, "metrics": {"hit_rate": 0.985, "decode_tok_s": 125.8, "engine_generated": 256, "prompt_ms": 1708.4, "decode_ms": 2034.3, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.51, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1790988051.755, "t1": 1790988055.507, "seconds": 3.752, "host_samples": 4, "gpu_samples": 4}, "board": {"joules": 424.81, "mean_w": 141.34, "max_w": 145.72, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 115.65, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 3.0}, "data_reads_per_s": {"mean": 132.0, "max": 200.0}, "data_read_mb_per_s": {"mean": 0.55}, "temps": {"gpu_max_c": 54.0, "ssd_max_c": 50.85}, "vram": {"max_used_mib": 23811.0, "baseline_mib": 15.0, "growth_mib": 23796.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 38804627456, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 2.7, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 16.535, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.0, "unit_psi_full_avg10_max": 0.0, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 1}, "power_limit": {"min_w": 138.14, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-Q2/speed/a1/run-004", "arm": "S-Q2", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "Q2_0", "file_sha256": {"1": "69820c02ec7d0b45ef2ebb19d6620299db749fe2aded7f39f93c6b88b199b720", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-8fc40dde/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-Q2/speed/a1/strata-config.json", "--port", "withheld-2026-10-13"], "ctx": "32K", "repeat": 0, "run": 4, "slice_sha256": "b72b07858a845e76331ec38d46abae3c09d3e1cb6e4d968e887fa8b132c64977", "prompt_sha256": "507b7bc9961d968eaf07467339f314a3c76a036b2542852f07b9a638adcd5ad5", "nonce": "Ticket 10426037.", "requests_since_start": 4, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 120.797, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 120.6, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 248, "content_frames": 247, "frames": 249, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2805.7, "prompt_tok_s_client": 2790.5, "ttft_ms": 11481.054, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11419.0, "prompt_per_token_ms": 0.356, "prompt_per_second": 2805.7, "predicted_n": 256, "predicted_ms": 2123.6, "predicted_per_token_ms": 8.295, "predicted_per_second": 120.6, "draft_n": 222, "draft_n_accepted": 156}, "short_completion": false, "cache_flag": null, "completion_sha256": "2dbb9d10ff4d31c93cc111f158b58610f72ef9ab74cd2ff5529f6df5599c32fa", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.982, "hits": 150513, "lookups": 153331, "drafts_accepted": 156, "drafts_offered": 222, "engine_generated": 256, "engine_decode_tok_s": 120.5, "engine_prompt_tok_s": 2805.7, "reused": 0, "metrics": {"hit_rate": 0.982, "decode_tok_s": 120.6, "engine_generated": 256, "prompt_ms": 11419.0, "decode_ms": 2123.6, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.56, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1790988055.508, "t1": 1790988069.1, "seconds": 13.592, "host_samples": 13, "gpu_samples": 13}, "board": {"joules": 1686.89, "mean_w": 140.39, "max_w": 149.4, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 489.78, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 12.002}, "data_reads_per_s": {"mean": 13290.9, "max": 96463.7}, "data_read_mb_per_s": {"mean": 55.65}, "temps": {"gpu_max_c": 60.0, "ssd_max_c": 50.85}, "vram": {"max_used_mib": 23811.0, "baseline_mib": 15.0, "growth_mib": 23796.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 39098351616, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 4.15, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 16.267, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.0, "unit_psi_full_avg10_max": 0.0, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 166}, "power_limit": {"min_w": 135.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-Q2/speed/a1/run-005", "arm": "S-Q2", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "Q2_0", "file_sha256": {"1": "69820c02ec7d0b45ef2ebb19d6620299db749fe2aded7f39f93c6b88b199b720", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-8fc40dde/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-Q2/speed/a1/strata-config.json", "--port", "withheld-2026-10-13"], "ctx": "32K", "repeat": 1, "run": 5, "slice_sha256": "b72b07858a845e76331ec38d46abae3c09d3e1cb6e4d968e887fa8b132c64977", "prompt_sha256": "c8a1ccb2149bdf9595fe8d97b7183fd9efd61c32259e907606ddbac7710955b2", "nonce": "Ticket 97542637.", "requests_since_start": 5, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 120.315, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 120.1, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 249, "content_frames": 248, "frames": 250, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2818.4, "prompt_tok_s_client": 2801.2, "ttft_ms": 11437.191, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11367.5, "prompt_per_token_ms": 0.355, "prompt_per_second": 2818.4, "predicted_n": 256, "predicted_ms": 2131.6, "predicted_per_token_ms": 8.327, "predicted_per_second": 120.1, "draft_n": 212, "draft_n_accepted": 154}, "short_completion": false, "cache_flag": null, "completion_sha256": "e2d451ddd80fb6823a0bcb3ee554534f3309aa32c6a1727747a50d9598ed5f21", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.985, "hits": 147592, "lookups": 149861, "drafts_accepted": 154, "drafts_offered": 212, "engine_generated": 256, "engine_decode_tok_s": 120.1, "engine_prompt_tok_s": 2818.4, "reused": 0, "metrics": {"hit_rate": 0.985, "decode_tok_s": 120.1, "engine_generated": 256, "prompt_ms": 11367.5, "decode_ms": 2131.6, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.51, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1790988069.103, "t1": 1790988082.66, "seconds": 13.557, "host_samples": 14, "gpu_samples": 14}, "board": {"joules": 1869.19, "mean_w": 143.65, "max_w": 156.63, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 469.21, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 13.002}, "data_reads_per_s": {"mean": 170.3, "max": 1188.9}, "data_read_mb_per_s": {"mean": 0.71}, "temps": {"gpu_max_c": 60.0, "ssd_max_c": 49.85}, "vram": {"max_used_mib": 23811.0, "baseline_mib": 15.0, "growth_mib": 23796.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 39098761216, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 3.95, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 16.319, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.0, "unit_psi_full_avg10_max": 0.0, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 135.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-Q2/speed/a1/run-006", "arm": "S-Q2", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "Q2_0", "file_sha256": {"1": "69820c02ec7d0b45ef2ebb19d6620299db749fe2aded7f39f93c6b88b199b720", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-8fc40dde/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-Q2/speed/a1/strata-config.json", "--port", "withheld-2026-10-13"], "ctx": "32K", "repeat": 2, "run": 6, "slice_sha256": "b72b07858a845e76331ec38d46abae3c09d3e1cb6e4d968e887fa8b132c64977", "prompt_sha256": "d6c964051c51baa9ef5cb009be0a921422c2b28350d2dbea610a6038552bdc18", "nonce": "Ticket 24502846.", "requests_since_start": 6, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 127.018, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 126.7, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 246, "content_frames": 245, "frames": 247, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2823.0, "prompt_tok_s_client": 2806.1, "ttft_ms": 11417.407, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11349.0, "prompt_per_token_ms": 0.354, "prompt_per_second": 2823.0, "predicted_n": 256, "predicted_ms": 2021.0, "predicted_per_token_ms": 7.895, "predicted_per_second": 126.7, "draft_n": 211, "draft_n_accepted": 162}, "short_completion": false, "cache_flag": null, "completion_sha256": "0b79d77a41a936f1f974d6afebb35039d7c94df6a98e4d2b0bc6a0058cfe8239", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9890000000000001, "hits": 144811, "lookups": 146376, "drafts_accepted": 162, "drafts_offered": 211, "engine_generated": 256, "engine_decode_tok_s": 126.7, "engine_prompt_tok_s": 2823.0, "reused": 0, "metrics": {"hit_rate": 0.989, "decode_tok_s": 126.7, "engine_generated": 256, "prompt_ms": 11349.0, "decode_ms": 2021.0, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.723, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1790988082.663, "t1": 1790988096.088, "seconds": 13.425, "host_samples": 13, "gpu_samples": 13}, "board": {"joules": 1709.38, "mean_w": 142.22, "max_w": 150.23, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 497.29, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 12.002}, "data_reads_per_s": {"mean": 137.4, "max": 1163.9}, "data_read_mb_per_s": {"mean": 0.58}, "temps": {"gpu_max_c": 63.0, "ssd_max_c": 48.85}, "vram": {"max_used_mib": 23811.0, "baseline_mib": 15.0, "growth_mib": 23796.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 39098761216, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 4.01, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 16.338, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.0, "unit_psi_full_avg10_max": 0.0, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 141.56, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-Q2/speed/a1/run-007", "arm": "S-Q2", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "Q2_0", "file_sha256": {"1": "69820c02ec7d0b45ef2ebb19d6620299db749fe2aded7f39f93c6b88b199b720", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-8fc40dde/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-Q2/speed/a1/strata-config.json", "--port", "withheld-2026-10-13"], "ctx": "128K", "repeat": 0, "run": 7, "slice_sha256": "286827657f5df11459f4ee550246f7e0e06ce7af3238fe74a53f0aa4c12676ac", "prompt_sha256": "2491e963527b8a39f7fff1aa2d04dab5eb4dd3ca12c7417c451f6216c665972c", "nonce": "Ticket 14239051.", "requests_since_start": 7, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 106.57, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 106.3, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 250, "content_frames": 249, "frames": 251, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2855.8, "prompt_tok_s_client": 2842.5, "ttft_ms": 45044.507, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 44833.8, "prompt_per_token_ms": 0.35, "prompt_per_second": 2855.8, "predicted_n": 256, "predicted_ms": 2408.3, "predicted_per_token_ms": 9.407, "predicted_per_second": 106.3, "draft_n": 231, "draft_n_accepted": 153}, "short_completion": false, "cache_flag": null, "completion_sha256": "ad9d85a96d0f36d32d71949141abdd8a5f9a30e778e0886bbec5f61048004b56", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.988, "hits": 159092, "lookups": 161058, "drafts_accepted": 153, "drafts_offered": 231, "engine_generated": 256, "engine_decode_tok_s": 106.3, "engine_prompt_tok_s": 2855.8, "reused": 0, "metrics": {"hit_rate": 0.988, "decode_tok_s": 106.3, "engine_generated": 256, "prompt_ms": 44833.8, "decode_ms": 2408.3, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.485, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1790988096.092, "t1": 1790988143.53, "seconds": 47.438, "host_samples": 48, "gpu_samples": 48}, "board": {"joules": 6883.76, "mean_w": 146.29, "max_w": 150.93, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 1798.68, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 47.007}, "data_reads_per_s": {"mean": 13968.8, "max": 78781.9}, "data_read_mb_per_s": {"mean": 58.45}, "temps": {"gpu_max_c": 67.0, "ssd_max_c": 49.85}, "vram": {"max_used_mib": 23811.0, "baseline_mib": 15.0, "growth_mib": 23796.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 39599144960, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.97, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 15.692, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.0, "unit_psi_full_avg10_max": 0.0, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 580}, "power_limit": {"min_w": 144.17, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-Q2/speed/a1/run-008", "arm": "S-Q2", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "Q2_0", "file_sha256": {"1": "69820c02ec7d0b45ef2ebb19d6620299db749fe2aded7f39f93c6b88b199b720", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-8fc40dde/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-Q2/speed/a1/strata-config.json", "--port", "withheld-2026-10-13"], "ctx": "128K", "repeat": 1, "run": 8, "slice_sha256": "286827657f5df11459f4ee550246f7e0e06ce7af3238fe74a53f0aa4c12676ac", "prompt_sha256": "94613594c1c300728a39aa8436592de0bdf88c05edeceb88cd9c6794de1fa097", "nonce": "Ticket 81078402.", "requests_since_start": 8, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 112.755, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 112.5, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 250, "content_frames": 249, "frames": 251, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2848.5, "prompt_tok_s_client": 2834.8, "ttft_ms": 45166.011, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 44949.2, "prompt_per_token_ms": 0.351, "prompt_per_second": 2848.5, "predicted_n": 256, "predicted_ms": 2275.6, "predicted_per_token_ms": 8.889, "predicted_per_second": 112.5, "draft_n": 211, "draft_n_accepted": 150}, "short_completion": false, "cache_flag": null, "completion_sha256": "c4e3cbee36346d036891e7a3204f8a5a9137f3c8e056f6e11ec902836a27ac0d", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.991, "hits": 151357, "lookups": 152738, "drafts_accepted": 150, "drafts_offered": 211, "engine_generated": 256, "engine_decode_tok_s": 112.5, "engine_prompt_tok_s": 2848.5, "reused": 0, "metrics": {"hit_rate": 0.991, "decode_tok_s": 112.5, "engine_generated": 256, "prompt_ms": 44949.2, "decode_ms": 2275.6, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.415, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1790988143.535, "t1": 1790988190.963, "seconds": 47.428, "host_samples": 47, "gpu_samples": 47}, "board": {"joules": 6729.46, "mean_w": 146.09, "max_w": 150.46, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 1979.43, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 46.007}, "data_reads_per_s": {"mean": 5445.7, "max": 44485.1}, "data_read_mb_per_s": {"mean": 22.78}, "temps": {"gpu_max_c": 70.0, "ssd_max_c": 50.85}, "vram": {"max_used_mib": 23811.0, "baseline_mib": 15.0, "growth_mib": 23796.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 39605673984, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 4.86, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 15.299, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.0, "unit_psi_full_avg10_max": 0.0, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 7}, "power_limit": {"min_w": 135.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-Q2/speed/a1/run-009", "arm": "S-Q2", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "Q2_0", "file_sha256": {"1": "69820c02ec7d0b45ef2ebb19d6620299db749fe2aded7f39f93c6b88b199b720", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-8fc40dde/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-Q2/speed/a1/strata-config.json", "--port", "withheld-2026-10-13"], "ctx": "128K", "repeat": 2, "run": 9, "slice_sha256": "286827657f5df11459f4ee550246f7e0e06ce7af3238fe74a53f0aa4c12676ac", "prompt_sha256": "38d3c9a3177694d5b9006f1c813fddc9e6c5745048a59c545c41d199f780b364", "nonce": "Ticket 81461625.", "requests_since_start": 9, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 110.553, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 110.3, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 249, "content_frames": 248, "frames": 250, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2837.7, "prompt_tok_s_client": 2824.2, "ttft_ms": 45335.549, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 45120.3, "prompt_per_token_ms": 0.352, "prompt_per_second": 2837.7, "predicted_n": 256, "predicted_ms": 2320.9, "predicted_per_token_ms": 9.066, "predicted_per_second": 110.3, "draft_n": 214, "draft_n_accepted": 148}, "short_completion": false, "cache_flag": null, "completion_sha256": "1ea16a74cc30098069e6a09cd75105dad74aa7771e56f487f1e1a8f867a45eb4", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.992, "hits": 153493, "lookups": 154723, "drafts_accepted": 148, "drafts_offered": 214, "engine_generated": 256, "engine_decode_tok_s": 110.3, "engine_prompt_tok_s": 2837.7, "reused": 0, "metrics": {"hit_rate": 0.992, "decode_tok_s": 110.3, "engine_generated": 256, "prompt_ms": 45120.3, "decode_ms": 2320.9, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.37, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1790988190.965, "t1": 1790988238.607, "seconds": 47.643, "host_samples": 48, "gpu_samples": 48}, "board": {"joules": 6843.47, "mean_w": 145.46, "max_w": 150.38, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 1871.28, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 47.008}, "data_reads_per_s": {"mean": 5130.8, "max": 33319.9}, "data_read_mb_per_s": {"mean": 21.47}, "temps": {"gpu_max_c": 71.0, "ssd_max_c": 51.85}, "vram": {"max_used_mib": 23811.0, "baseline_mib": 15.0, "growth_mib": 23796.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 39607410688, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.29, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 15.407, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.0, "unit_psi_full_avg10_max": 0.0, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 2}, "power_limit": {"min_w": 135.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}