{"row_id": "S-IQ3/speed/a1/run-001", "arm": "S-IQ3", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "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-IQ3/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": 78.162, "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": 77.8, "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": 1871.9, "prompt_tok_s_client": 1838.8, "ttft_ms": 2195.958, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 2157.2, "prompt_per_token_ms": 0.534, "prompt_per_second": 1871.9, "predicted_n": 256, "predicted_ms": 3288.6, "predicted_per_token_ms": 12.846, "predicted_per_second": 77.8, "draft_n": 236, "draft_n_accepted": 156}, "short_completion": false, "cache_flag": null, "completion_sha256": "64b905b4cfef09a757075d73bb32e00ca58d2c5574565568104b16f79c7389f0", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.871, "hits": 128839, "lookups": 147989, "drafts_accepted": 156, "drafts_offered": 236, "engine_generated": 256, "engine_decode_tok_s": 77.8, "engine_prompt_tok_s": 1871.9, "reused": 0, "metrics": {"hit_rate": 0.871, "decode_tok_s": 77.8, "engine_generated": 256, "prompt_ms": 2157.2, "decode_ms": 3288.6, "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": 1790988428.786, "t1": 1790988434.245, "seconds": 5.459, "host_samples": 6, "gpu_samples": 6}, "board": {"joules": 569.22, "mean_w": 113.77, "max_w": 137.37, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 380.75, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 5.0}, "data_reads_per_s": {"mean": 8524.4, "max": 47643.0}, "data_read_mb_per_s": {"mean": 36.15}, "temps": {"gpu_max_c": 56.0, "ssd_max_c": 57.85}, "vram": {"max_used_mib": 23835.0, "baseline_mib": 15.0, "growth_mib": 23820.0, "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": 53166870528, "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": 3.823, "user_psi_full_avg10_max": 1.27, "psi_full_avg10_max": 2.43, "unit_psi_full_avg10_max": 0.19, "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-IQ3/speed/a1/run-002", "arm": "S-IQ3", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "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-IQ3/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": 91.424, "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": 91.3, "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": 2157.3, "prompt_tok_s_client": 2118.7, "ttft_ms": 1905.9, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1871.8, "prompt_per_token_ms": 0.464, "prompt_per_second": 2157.3, "predicted_n": 256, "predicted_ms": 2804.6, "predicted_per_token_ms": 10.955, "predicted_per_second": 91.3, "draft_n": 234, "draft_n_accepted": 162}, "short_completion": false, "cache_flag": null, "completion_sha256": "dc3b78defce89d71117bb35f4974bf84327a368944d6ea3f1170b1e7cd46e9fc", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.917, "hits": 137700, "lookups": 150161, "drafts_accepted": 162, "drafts_offered": 234, "engine_generated": 256, "engine_decode_tok_s": 91.3, "engine_prompt_tok_s": 2157.3, "reused": 0, "metrics": {"hit_rate": 0.917, "decode_tok_s": 91.3, "engine_generated": 256, "prompt_ms": 1871.8, "decode_ms": 2804.6, "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": 1790988434.246, "t1": 1790988438.941, "seconds": 4.695, "host_samples": 4, "gpu_samples": 4}, "board": {"joules": 394.54, "mean_w": 131.47, "max_w": 141.18, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 144.15, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 3.001}, "data_reads_per_s": {"mean": 441.7, "max": 806.9}, "data_read_mb_per_s": {"mean": 1.85}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": 55.85}, "vram": {"max_used_mib": 23835.0, "baseline_mib": 15.0, "growth_mib": 23820.0, "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": 53267578880, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.57, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.719, "user_psi_full_avg10_max": 0.85, "psi_full_avg10_max": 1.57, "unit_psi_full_avg10_max": 0.1, "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-IQ3/speed/a1/run-003", "arm": "S-IQ3", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "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-IQ3/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": 85.572, "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": 85.4, "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": 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": 2147.5, "prompt_tok_s_client": 2109.8, "ttft_ms": 1913.944, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1880.3, "prompt_per_token_ms": 0.466, "prompt_per_second": 2147.5, "predicted_n": 256, "predicted_ms": 2996.4, "predicted_per_token_ms": 11.705, "predicted_per_second": 85.4, "draft_n": 230, "draft_n_accepted": 161}, "short_completion": false, "cache_flag": null, "completion_sha256": "5963b58dd033f468155b47b6f59888dd45fed7a9836752b0ba9aeabc6edd7369", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.925, "hits": 138741, "lookups": 149963, "drafts_accepted": 161, "drafts_offered": 230, "engine_generated": 256, "engine_decode_tok_s": 85.4, "engine_prompt_tok_s": 2147.5, "reused": 0, "metrics": {"hit_rate": 0.925, "decode_tok_s": 85.4, "engine_generated": 256, "prompt_ms": 1880.3, "decode_ms": 2996.4, "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": 1790988438.944, "t1": 1790988443.838, "seconds": 4.894, "host_samples": 5, "gpu_samples": 5}, "board": {"joules": 505.36, "mean_w": 126.3, "max_w": 136.06, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 315.75, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 4.0}, "data_reads_per_s": {"mean": 193.6, "max": 951.9}, "data_read_mb_per_s": {"mean": 0.82}, "temps": {"gpu_max_c": 59.0, "ssd_max_c": 55.85}, "vram": {"max_used_mib": 23835.0, "baseline_mib": 15.0, "growth_mib": 23820.0, "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": 53267652608, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.47, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.641, "user_psi_full_avg10_max": 0.57, "psi_full_avg10_max": 1.05, "unit_psi_full_avg10_max": 0.07, "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-IQ3/speed/a1/run-004", "arm": "S-IQ3", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "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-IQ3/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": 87.343, "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": 87.1, "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": 2650.5, "prompt_tok_s_client": 2633.6, "ttft_ms": 12164.966, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 12087.4, "prompt_per_token_ms": 0.377, "prompt_per_second": 2650.5, "predicted_n": 256, "predicted_ms": 2938.4, "predicted_per_token_ms": 11.478, "predicted_per_second": 87.1, "draft_n": 214, "draft_n_accepted": 150}, "short_completion": false, "cache_flag": null, "completion_sha256": "e3cdde9fd2287da6e2ead8c329b64adbfc0655de9cf614bf07aec53e25b9827e", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.919, "hits": 134042, "lookups": 145842, "drafts_accepted": 150, "drafts_offered": 214, "engine_generated": 256, "engine_decode_tok_s": 87.1, "engine_prompt_tok_s": 2650.5, "reused": 0, "metrics": {"hit_rate": 0.919, "decode_tok_s": 87.1, "engine_generated": 256, "prompt_ms": 12087.4, "decode_ms": 2938.4, "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": 1790988443.839, "t1": 1790988458.924, "seconds": 15.085, "host_samples": 15, "gpu_samples": 15}, "board": {"joules": 1964.58, "mean_w": 140.29, "max_w": 154.36, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 714.07, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 14.002}, "data_reads_per_s": {"mean": 11571.0, "max": 61273.8}, "data_read_mb_per_s": {"mean": 48.46}, "temps": {"gpu_max_c": 61.0, "ssd_max_c": 54.85}, "vram": {"max_used_mib": 23835.0, "baseline_mib": 15.0, "growth_mib": 23820.0, "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": 53443387392, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.35, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.491, "user_psi_full_avg10_max": 0.31, "psi_full_avg10_max": 0.58, "unit_psi_full_avg10_max": 0.04, "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-IQ3/speed/a1/run-005", "arm": "S-IQ3", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "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-IQ3/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": 85.582, "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": 85.5, "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": 2655.1, "prompt_tok_s_client": 2641.0, "ttft_ms": 12130.806, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 12066.5, "prompt_per_token_ms": 0.377, "prompt_per_second": 2655.1, "predicted_n": 256, "predicted_ms": 2995.6, "predicted_per_token_ms": 11.702, "predicted_per_second": 85.5, "draft_n": 215, "draft_n_accepted": 156}, "short_completion": false, "cache_flag": null, "completion_sha256": "7cfa81e23a2f5e3630ef17329ac1de85e729c87a5d5349b0fbb68a27925d15c5", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.932, "hits": 135449, "lookups": 145279, "drafts_accepted": 156, "drafts_offered": 215, "engine_generated": 256, "engine_decode_tok_s": 85.5, "engine_prompt_tok_s": 2655.1, "reused": 0, "metrics": {"hit_rate": 0.932, "decode_tok_s": 85.5, "engine_generated": 256, "prompt_ms": 12066.5, "decode_ms": 2995.6, "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": 1790988458.925, "t1": 1790988474.036, "seconds": 15.111, "host_samples": 15, "gpu_samples": 15}, "board": {"joules": 1946.0, "mean_w": 138.97, "max_w": 148.87, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 580.05, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 14.002}, "data_reads_per_s": {"mean": 141.0, "max": 854.9}, "data_read_mb_per_s": {"mean": 0.62}, "temps": {"gpu_max_c": 63.0, "ssd_max_c": 53.85}, "vram": {"max_used_mib": 23835.0, "baseline_mib": 15.0, "growth_mib": 23820.0, "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": 53455491072, "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": 3.383, "user_psi_full_avg10_max": 0.07, "psi_full_avg10_max": 0.14, "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": 135.27, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-IQ3/speed/a1/run-006", "arm": "S-IQ3", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "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-IQ3/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": 87.054, "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": 86.9, "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": 2664.5, "prompt_tok_s_client": 2649.8, "ttft_ms": 12090.582, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 12024.2, "prompt_per_token_ms": 0.375, "prompt_per_second": 2664.5, "predicted_n": 256, "predicted_ms": 2946.0, "predicted_per_token_ms": 11.508, "predicted_per_second": 86.9, "draft_n": 219, "draft_n_accepted": 158}, "short_completion": false, "cache_flag": null, "completion_sha256": "b2dfe5e04ed2eb6cb9d367231da30f2f23e970fa983fa44b4d0fbe19e967808a", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.935, "hits": 136960, "lookups": 146485, "drafts_accepted": 158, "drafts_offered": 219, "engine_generated": 256, "engine_decode_tok_s": 86.9, "engine_prompt_tok_s": 2664.5, "reused": 0, "metrics": {"hit_rate": 0.935, "decode_tok_s": 86.9, "engine_generated": 256, "prompt_ms": 12024.2, "decode_ms": 2946.0, "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": 1790988474.037, "t1": 1790988489.058, "seconds": 15.02, "host_samples": 15, "gpu_samples": 15}, "board": {"joules": 1975.07, "mean_w": 141.03, "max_w": 148.81, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 639.31, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 14.002}, "data_reads_per_s": {"mean": 58.2, "max": 718.9}, "data_read_mb_per_s": {"mean": 0.24}, "temps": {"gpu_max_c": 66.0, "ssd_max_c": 52.85}, "vram": {"max_used_mib": 23835.0, "baseline_mib": 15.0, "growth_mib": 23820.0, "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": 53457080320, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 4.94, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.312, "user_psi_full_avg10_max": 0.01, "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": 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-IQ3/speed/a1/run-007", "arm": "S-IQ3", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "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-IQ3/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": 71.921, "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": 71.8, "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": 2716.0, "prompt_tok_s_client": 2703.7, "ttft_ms": 47355.706, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 47141.9, "prompt_per_token_ms": 0.368, "prompt_per_second": 2716.0, "predicted_n": 256, "predicted_ms": 3563.7, "predicted_per_token_ms": 13.921, "predicted_per_second": 71.8, "draft_n": 219, "draft_n_accepted": 155}, "short_completion": false, "cache_flag": null, "completion_sha256": "b78613d3b4c2dc2fb1eaeb6ce0a7841a44fb863ac58fb761381adabe016bf635", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.917, "hits": 133204, "lookups": 145239, "drafts_accepted": 155, "drafts_offered": 219, "engine_generated": 256, "engine_decode_tok_s": 71.8, "engine_prompt_tok_s": 2716.0, "reused": 0, "metrics": {"hit_rate": 0.917, "decode_tok_s": 71.8, "engine_generated": 256, "prompt_ms": 47141.9, "decode_ms": 3563.7, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.535, "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": 1790988489.062, "t1": 1790988539.964, "seconds": 50.902, "host_samples": 51, "gpu_samples": 51}, "board": {"joules": 7262.5, "mean_w": 145.13, "max_w": 158.88, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 2536.79, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 50.008}, "data_reads_per_s": {"mean": 13138.9, "max": 78780.5}, "data_read_mb_per_s": {"mean": 54.98}, "temps": {"gpu_max_c": 70.0, "ssd_max_c": 52.85}, "vram": {"max_used_mib": 23835.0, "baseline_mib": 15.0, "growth_mib": 23820.0, "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": 54070657024, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 4.31, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.685, "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-IQ3/speed/a1/run-008", "arm": "S-IQ3", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "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-IQ3/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": 69.765, "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": 69.7, "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": 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": 2705.0, "prompt_tok_s_client": 2692.3, "ttft_ms": 47557.425, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 47333.8, "prompt_per_token_ms": 0.37, "prompt_per_second": 2705.0, "predicted_n": 256, "predicted_ms": 3674.8, "predicted_per_token_ms": 14.355, "predicted_per_second": 69.7, "draft_n": 233, "draft_n_accepted": 142}, "short_completion": false, "cache_flag": null, "completion_sha256": "fed6b25ba6a65847ea03408e48820e12700166ce05c57d2229e415eab9c40232", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.924, "hits": 146228, "lookups": 158295, "drafts_accepted": 142, "drafts_offered": 233, "engine_generated": 256, "engine_decode_tok_s": 69.7, "engine_prompt_tok_s": 2705.0, "reused": 0, "metrics": {"hit_rate": 0.924, "decode_tok_s": 69.7, "engine_generated": 256, "prompt_ms": 47333.8, "decode_ms": 3674.8, "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": 1790988539.968, "t1": 1790988591.181, "seconds": 51.213, "host_samples": 52, "gpu_samples": 51}, "board": {"joules": 7276.2, "mean_w": 145.34, "max_w": 161.6, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 2551.46, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 51.007}, "data_reads_per_s": {"mean": 4942.0, "max": 44147.1}, "data_read_mb_per_s": {"mean": 20.67}, "temps": {"gpu_max_c": 71.0, "ssd_max_c": 53.85}, "vram": {"max_used_mib": 23835.0, "baseline_mib": 15.0, "growth_mib": 23820.0, "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": 54079365120, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.7, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.667, "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-IQ3/speed/a1/run-009", "arm": "S-IQ3", "engine": "strata", "engine_commit": "8fc40dde5e8897a097ec0ea6ea27f9313d171268", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "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-IQ3/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": 64.927, "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": 64.7, "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": 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": 2697.5, "prompt_tok_s_client": 2683.5, "ttft_ms": 47712.203, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 47464.9, "prompt_per_token_ms": 0.371, "prompt_per_second": 2697.5, "predicted_n": 256, "predicted_ms": 3958.1, "predicted_per_token_ms": 15.461, "predicted_per_second": 64.7, "draft_n": 253, "draft_n_accepted": 142}, "short_completion": false, "cache_flag": null, "completion_sha256": "76cf55b7c39acf0bc86efa3048de011c219bd15fef2fe9361496f73251095b7a", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9229999999999999, "hits": 154468, "lookups": 167425, "drafts_accepted": 142, "drafts_offered": 253, "engine_generated": 256, "engine_decode_tok_s": 64.7, "engine_prompt_tok_s": 2697.5, "reused": 0, "metrics": {"hit_rate": 0.923, "decode_tok_s": 64.7, "engine_generated": 256, "prompt_ms": 47464.9, "decode_ms": 3958.1, "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": 1790988591.184, "t1": 1790988642.824, "seconds": 51.64, "host_samples": 51, "gpu_samples": 52}, "board": {"joules": 7410.31, "mean_w": 145.1, "max_w": 151.49, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 2623.08, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 50.008}, "data_reads_per_s": {"mean": 4813.8, "max": 33406.6}, "data_read_mb_per_s": {"mean": 20.14}, "temps": {"gpu_max_c": 74.0, "ssd_max_c": 53.85}, "vram": {"max_used_mib": 23835.0, "baseline_mib": 15.0, "growth_mib": 23820.0, "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": 54079291392, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 6.19, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.664, "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"}}}