{"row_id": "S-Q2/speed/a2/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/a2/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": 120.301, "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": 119.4, "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": 2043.0, "prompt_tok_s_client": 2004.7, "ttft_ms": 2014.281, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1976.5, "prompt_per_token_ms": 0.489, "prompt_per_second": 2043.0, "predicted_n": 256, "predicted_ms": 2144.6, "predicted_per_token_ms": 8.377, "predicted_per_second": 119.4, "draft_n": 211, "draft_n_accepted": 158}, "short_completion": false, "cache_flag": null, "completion_sha256": "93fbe774adcc88b5c1a68b48148c610a99c7b6cc9e940f08f410d79747ca5eb4", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9540000000000001, "hits": 138160, "lookups": 144814, "drafts_accepted": 158, "drafts_offered": 211, "engine_generated": 256, "engine_decode_tok_s": 119.4, "engine_prompt_tok_s": 2043.0, "reused": 0, "metrics": {"hit_rate": 0.954, "decode_tok_s": 119.4, "engine_generated": 256, "prompt_ms": 1976.5, "decode_ms": 2144.6, "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": 1791004206.541, "t1": 1791004210.675, "seconds": 4.134, "host_samples": 4, "gpu_samples": 4}, "board": {"joules": 346.25, "mean_w": 115.37, "max_w": 150.56, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 170.75, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 3.004}, "data_reads_per_s": {"mean": 12411.0, "max": 47629.9}, "data_read_mb_per_s": {"mean": 52.13}, "temps": {"gpu_max_c": 45.0, "ssd_max_c": 44.85}, "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": 38028767232, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 1.83, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 16.498, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.44, "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": 150.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-Q2/speed/a2/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/a2/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": 134.88, "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": 134.2, "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": 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": 2399.3, "prompt_tok_s_client": 2352.2, "ttft_ms": 1716.663, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1683.0, "prompt_per_token_ms": 0.417, "prompt_per_second": 2399.3, "predicted_n": 256, "predicted_ms": 1906.9, "predicted_per_token_ms": 7.449, "predicted_per_second": 134.2, "draft_n": 207, "draft_n_accepted": 159}, "short_completion": false, "cache_flag": null, "completion_sha256": "031a4e2ac934ba1ec7c42982f795ddf5419dcaa2870d99b5b002da27bfaaf50d", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.983, "hits": 142408, "lookups": 144817, "drafts_accepted": 159, "drafts_offered": 207, "engine_generated": 256, "engine_decode_tok_s": 134.2, "engine_prompt_tok_s": 2399.2, "reused": 0, "metrics": {"hit_rate": 0.983, "decode_tok_s": 134.2, "engine_generated": 256, "prompt_ms": 1683.0, "decode_ms": 1906.9, "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": 1791004210.679, "t1": 1791004214.287, "seconds": 3.607, "host_samples": 4, "gpu_samples": 3}, "board": {"joules": 261.86, "mean_w": 130.8, "max_w": 138.75, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 213.01, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 3.002}, "data_reads_per_s": {"mean": 447.9, "max": 1559.8}, "data_read_mb_per_s": {"mean": 1.87}, "temps": {"gpu_max_c": 48.0, "ssd_max_c": 44.85}, "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": 38788243456, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 2.91, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 16.443, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.29, "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/a2/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/a2/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": 131.727, "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": 131.3, "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": 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": 2435.8, "prompt_tok_s_client": 2390.1, "ttft_ms": 1689.486, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1657.8, "prompt_per_token_ms": 0.411, "prompt_per_second": 2435.8, "predicted_n": 256, "predicted_ms": 1949.1, "predicted_per_token_ms": 7.614, "predicted_per_second": 131.3, "draft_n": 215, "draft_n_accepted": 161}, "short_completion": false, "cache_flag": null, "completion_sha256": "491eedddfda01befb2ed0d83e68927833cd03c87f6bff19318019d8d36144ee0", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.986, "hits": 146829, "lookups": 148904, "drafts_accepted": 161, "drafts_offered": 215, "engine_generated": 256, "engine_decode_tok_s": 131.3, "engine_prompt_tok_s": 2435.8, "reused": 0, "metrics": {"hit_rate": 0.986, "decode_tok_s": 131.3, "engine_generated": 256, "prompt_ms": 1657.8, "decode_ms": 1949.1, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.695, "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": 1791004214.289, "t1": 1791004217.914, "seconds": 3.625, "host_samples": 4, "gpu_samples": 4}, "board": {"joules": 411.51, "mean_w": 137.13, "max_w": 148.37, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 237.4, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 3.002}, "data_reads_per_s": {"mean": 48.0, "max": 104.0}, "data_read_mb_per_s": {"mean": 0.2}, "temps": {"gpu_max_c": 49.0, "ssd_max_c": 43.85}, "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": 38900469760, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 4.45, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 16.442, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.2, "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": 140.51, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-Q2/speed/a2/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/a2/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": 114.676, "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": 114.5, "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": 251, "content_frames": 250, "frames": 252, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2814.1, "prompt_tok_s_client": 2797.0, "ttft_ms": 11454.452, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11384.9, "prompt_per_token_ms": 0.355, "prompt_per_second": 2814.1, "predicted_n": 256, "predicted_ms": 2236.7, "predicted_per_token_ms": 8.737, "predicted_per_second": 114.5, "draft_n": 222, "draft_n_accepted": 151}, "short_completion": false, "cache_flag": null, "completion_sha256": "cbd5a947d83e822fc1dfd52e2629bbe3db30d05a7d9440d807ff4f13d6acfadc", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.977, "hits": 151762, "lookups": 155348, "drafts_accepted": 151, "drafts_offered": 222, "engine_generated": 256, "engine_decode_tok_s": 114.5, "engine_prompt_tok_s": 2814.1, "reused": 0, "metrics": {"hit_rate": 0.977, "decode_tok_s": 114.5, "engine_generated": 256, "prompt_ms": 11384.9, "decode_ms": 2236.7, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.438, "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": 1791004217.918, "t1": 1791004231.596, "seconds": 13.678, "host_samples": 13, "gpu_samples": 14}, "board": {"joules": 1829.96, "mean_w": 140.73, "max_w": 149.78, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 596.81, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 12.002}, "data_reads_per_s": {"mean": 13293.0, "max": 96060.5}, "data_read_mb_per_s": {"mean": 55.66}, "temps": {"gpu_max_c": 55.0, "ssd_max_c": 43.85}, "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": 39076601856, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.73, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 16.256, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.13, "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": 140.51, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-Q2/speed/a2/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/a2/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": 115.243, "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": 115.1, "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": 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": 2832.9, "prompt_tok_s_client": 2815.4, "ttft_ms": 11379.512, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11309.3, "prompt_per_token_ms": 0.353, "prompt_per_second": 2832.9, "predicted_n": 256, "predicted_ms": 2225.1, "predicted_per_token_ms": 8.692, "predicted_per_second": 115.1, "draft_n": 230, "draft_n_accepted": 153}, "short_completion": false, "cache_flag": null, "completion_sha256": "fcc40b9dd9faaf3ccccddba30417eaa768f21e00ef3181e4bb4c3516b063d7af", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.988, "hits": 157181, "lookups": 159142, "drafts_accepted": 153, "drafts_offered": 230, "engine_generated": 256, "engine_decode_tok_s": 115.1, "engine_prompt_tok_s": 2832.9, "reused": 0, "metrics": {"hit_rate": 0.988, "decode_tok_s": 115.1, "engine_generated": 256, "prompt_ms": 11309.3, "decode_ms": 2225.1, "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": 1791004231.598, "t1": 1791004245.19, "seconds": 13.592, "host_samples": 14, "gpu_samples": 13}, "board": {"joules": 1707.24, "mean_w": 142.22, "max_w": 148.8, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 808.25, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 13.003}, "data_reads_per_s": {"mean": 232.7, "max": 1317.5}, "data_read_mb_per_s": {"mean": 0.97}, "temps": {"gpu_max_c": 58.0, "ssd_max_c": 43.85}, "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": 39076601856, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.25, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 16.225, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.03, "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/a2/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/a2/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": 120.873, "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": 252, "content_frames": 251, "frames": 253, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2801.1, "prompt_tok_s_client": 2784.3, "ttft_ms": 11506.668, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11437.7, "prompt_per_token_ms": 0.357, "prompt_per_second": 2801.1, "predicted_n": 256, "predicted_ms": 2122.1, "predicted_per_token_ms": 8.289, "predicted_per_second": 120.6, "draft_n": 218, "draft_n_accepted": 156}, "short_completion": false, "cache_flag": null, "completion_sha256": "ad3731db3e0f4f863ead68ecfec7f4562f954459694d9eb03974a005a166c46b", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.988, "hits": 150691, "lookups": 152456, "drafts_accepted": 156, "drafts_offered": 218, "engine_generated": 256, "engine_decode_tok_s": 120.6, "engine_prompt_tok_s": 2801.1, "reused": 0, "metrics": {"hit_rate": 0.988, "decode_tok_s": 120.6, "engine_generated": 256, "prompt_ms": 11437.7, "decode_ms": 2122.1, "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": 1791004245.193, "t1": 1791004258.81, "seconds": 13.616, "host_samples": 13, "gpu_samples": 14}, "board": {"joules": 1826.51, "mean_w": 140.46, "max_w": 152.66, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 812.83, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 12.002}, "data_reads_per_s": {"mean": 47.8, "max": 359.0}, "data_read_mb_per_s": {"mean": 0.2}, "temps": {"gpu_max_c": 60.0, "ssd_max_c": 43.85}, "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": 38959595520, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.23, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 16.244, "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/a2/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/a2/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": 106.432, "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": 106.2, "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": 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": 2853.5, "prompt_tok_s_client": 2840.0, "ttft_ms": 45083.304, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 44870.0, "prompt_per_token_ms": 0.35, "prompt_per_second": 2853.5, "predicted_n": 256, "predicted_ms": 2409.9, "predicted_per_token_ms": 9.414, "predicted_per_second": 106.2, "draft_n": 223, "draft_n_accepted": 146}, "short_completion": false, "cache_flag": null, "completion_sha256": "4bc259d99b8b091b57be87863c197d1322c96e3730c27f1e93a35d8d34d1a401", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9890000000000001, "hits": 158378, "lookups": 160191, "drafts_accepted": 146, "drafts_offered": 223, "engine_generated": 256, "engine_decode_tok_s": 106.2, "engine_prompt_tok_s": 2853.5, "reused": 0, "metrics": {"hit_rate": 0.989, "decode_tok_s": 106.2, "engine_generated": 256, "prompt_ms": 44870.0, "decode_ms": 2409.9, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.327, "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": 1791004258.818, "t1": 1791004306.298, "seconds": 47.48, "host_samples": 48, "gpu_samples": 47}, "board": {"joules": 6703.9, "mean_w": 145.69, "max_w": 149.51, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 2520.77, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 47.008}, "data_reads_per_s": {"mean": 13975.7, "max": 75404.9}, "data_read_mb_per_s": {"mean": 58.48}, "temps": {"gpu_max_c": 65.0, "ssd_max_c": 46.85}, "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": 39577665536, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 6.75, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 15.185, "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/a2/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/a2/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": 102.519, "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": 102.3, "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": 247, "content_frames": 246, "frames": 248, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2843.2, "prompt_tok_s_client": 2830.2, "ttft_ms": 45238.774, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 45032.2, "prompt_per_token_ms": 0.352, "prompt_per_second": 2843.2, "predicted_n": 256, "predicted_ms": 2502.9, "predicted_per_token_ms": 9.777, "predicted_per_second": 102.3, "draft_n": 242, "draft_n_accepted": 143}, "short_completion": false, "cache_flag": null, "completion_sha256": "66abcb854a8e045eb578ede226761996af166ada6f6f3ad5bc44fb62772408f9", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9890000000000001, "hits": 167808, "lookups": 169710, "drafts_accepted": 143, "drafts_offered": 242, "engine_generated": 256, "engine_decode_tok_s": 102.3, "engine_prompt_tok_s": 2843.2, "reused": 0, "metrics": {"hit_rate": 0.989, "decode_tok_s": 102.3, "engine_generated": 256, "prompt_ms": 45032.2, "decode_ms": 2502.9, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.265, "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": 1791004306.304, "t1": 1791004354.031, "seconds": 47.727, "host_samples": 48, "gpu_samples": 48}, "board": {"joules": 6852.98, "mean_w": 145.61, "max_w": 176.79, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 2489.14, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 47.008}, "data_reads_per_s": {"mean": 5355.8, "max": 44806.2}, "data_read_mb_per_s": {"mean": 22.4}, "temps": {"gpu_max_c": 68.0, "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": 39582085120, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.71, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 15.538, "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/a2/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/a2/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": 102.201, "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": 100.5, "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": 2734.7, "prompt_tok_s_client": 2719.8, "ttft_ms": 47075.85, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 46819.2, "prompt_per_token_ms": 0.366, "prompt_per_second": 2734.7, "predicted_n": 256, "predicted_ms": 2547.9, "predicted_per_token_ms": 9.953, "predicted_per_second": 100.5, "draft_n": 218, "draft_n_accepted": 142}, "short_completion": false, "cache_flag": null, "completion_sha256": "1f082e66dc3ff2678956e787751bcabbbd3d6c872274f0cfd1f85651a6044122", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9890000000000001, "hits": 157107, "lookups": 158797, "drafts_accepted": 142, "drafts_offered": 218, "engine_generated": 256, "engine_decode_tok_s": 100.5, "engine_prompt_tok_s": 2734.7, "reused": 0, "metrics": {"hit_rate": 0.989, "decode_tok_s": 100.5, "engine_generated": 256, "prompt_ms": 46819.2, "decode_ms": 2547.9, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.246, "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": 1791004354.035, "t1": 1791004403.607, "seconds": 49.571, "host_samples": 49, "gpu_samples": 50}, "board": {"joules": 6732.13, "mean_w": 137.19, "max_w": 209.52, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 4155.46, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 48.008}, "data_reads_per_s": {"mean": 5029.9, "max": 33411.8}, "data_read_mb_per_s": {"mean": 21.04}, "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": 39582085120, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 7.07, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 13.714, "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": 148.38, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}