{"row_id": "S-HEAD/speed/a1/run-001", "arm": "S-HEAD", "engine": "strata", "engine_commit": "36fa455e579b23a9c909c2c6fe1bddd9e51cb8ca", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-36fa455e/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-HEAD/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": 80.014, "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": 79.6, "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": 1994.5, "prompt_tok_s_client": 1954.2, "ttft_ms": 2066.288, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 2024.6, "prompt_per_token_ms": 0.501, "prompt_per_second": 1994.5, "predicted_n": 256, "predicted_ms": 3216.3, "predicted_per_token_ms": 12.564, "predicted_per_second": 79.6, "draft_n": 222, "draft_n_accepted": 161}, "short_completion": false, "cache_flag": null, "completion_sha256": "29a39c1774f1b44c7f8b9a47a606d8471df0208370b0bc6fcfaed9316366d67a", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.87, "hits": 121590, "lookups": 139776, "drafts_accepted": 161, "drafts_offered": 222, "engine_generated": 256, "engine_decode_tok_s": 79.6, "engine_prompt_tok_s": 1994.5, "reused": 0, "metrics": {"hit_rate": 0.87, "decode_tok_s": 79.6, "engine_generated": 256, "prompt_ms": 2024.6, "decode_ms": 3216.3, "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": 1790988959.475, "t1": 1790988964.728, "seconds": 5.253, "host_samples": 5, "gpu_samples": 5}, "board": {"joules": 420.29, "mean_w": 105.04, "max_w": 142.55, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 288.96, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 4.001}, "data_reads_per_s": {"mean": 10017.9, "max": 47636.9}, "data_read_mb_per_s": {"mean": 42.54}, "temps": {"gpu_max_c": 53.0, "ssd_max_c": 54.85}, "vram": {"max_used_mib": 23747.0, "baseline_mib": 15.0, "growth_mib": 23732.0, "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": 53514407936, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 3.97, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.593, "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": 150.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-HEAD/speed/a1/run-002", "arm": "S-HEAD", "engine": "strata", "engine_commit": "36fa455e579b23a9c909c2c6fe1bddd9e51cb8ca", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-36fa455e/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-HEAD/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": 80.41, "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": 80.4, "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": 2313.4, "prompt_tok_s_client": 2270.4, "ttft_ms": 1778.526, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1745.5, "prompt_per_token_ms": 0.432, "prompt_per_second": 2313.4, "predicted_n": 256, "predicted_ms": 3185.4, "predicted_per_token_ms": 12.443, "predicted_per_second": 80.4, "draft_n": 217, "draft_n_accepted": 159}, "short_completion": false, "cache_flag": null, "completion_sha256": "f41fb58e77f6bc095606c0fe562edfb24466f076605a79d40dd69c895fac005d", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.92, "hits": 131484, "lookups": 142962, "drafts_accepted": 159, "drafts_offered": 217, "engine_generated": 256, "engine_decode_tok_s": 80.4, "engine_prompt_tok_s": 2313.4, "reused": 0, "metrics": {"hit_rate": 0.92, "decode_tok_s": 80.4, "engine_generated": 256, "prompt_ms": 1745.5, "decode_ms": 3185.4, "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": 1790988964.73, "t1": 1790988969.68, "seconds": 4.95, "host_samples": 5, "gpu_samples": 5}, "board": {"joules": 499.7, "mean_w": 124.91, "max_w": 136.07, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 349.77, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 4.001}, "data_reads_per_s": {"mean": 325.6, "max": 1270.9}, "data_read_mb_per_s": {"mean": 1.36}, "temps": {"gpu_max_c": 54.0, "ssd_max_c": 53.85}, "vram": {"max_used_mib": 23747.0, "baseline_mib": 15.0, "growth_mib": 23732.0, "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": 53520424960, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.65, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.472, "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": 146.38, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-HEAD/speed/a1/run-003", "arm": "S-HEAD", "engine": "strata", "engine_commit": "36fa455e579b23a9c909c2c6fe1bddd9e51cb8ca", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-36fa455e/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-HEAD/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": 83.649, "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": 83.5, "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": 2299.4, "prompt_tok_s_client": 2252.7, "ttft_ms": 1792.525, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1756.1, "prompt_per_token_ms": 0.435, "prompt_per_second": 2299.4, "predicted_n": 256, "predicted_ms": 3064.5, "predicted_per_token_ms": 11.971, "predicted_per_second": 83.5, "draft_n": 233, "draft_n_accepted": 160}, "short_completion": false, "cache_flag": null, "completion_sha256": "6b1ce0bf869df6b6ca7747aa1447a1dbbf61da8ad7dd9e5974c6940caff6a160", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.915, "hits": 137469, "lookups": 150198, "drafts_accepted": 160, "drafts_offered": 233, "engine_generated": 256, "engine_decode_tok_s": 83.5, "engine_prompt_tok_s": 2299.4, "reused": 0, "metrics": {"hit_rate": 0.915, "decode_tok_s": 83.5, "engine_generated": 256, "prompt_ms": 1756.1, "decode_ms": 3064.5, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.667, "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": 1790988969.685, "t1": 1790988974.526, "seconds": 4.841, "host_samples": 5, "gpu_samples": 5}, "board": {"joules": 513.05, "mean_w": 128.23, "max_w": 132.92, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 234.67, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 4.001}, "data_reads_per_s": {"mean": 435.1, "max": 1072.9}, "data_read_mb_per_s": {"mean": 1.82}, "temps": {"gpu_max_c": 56.0, "ssd_max_c": 52.85}, "vram": {"max_used_mib": 23747.0, "baseline_mib": 15.0, "growth_mib": 23732.0, "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": 53520490496, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 7.28, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.433, "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-HEAD/speed/a1/run-004", "arm": "S-HEAD", "engine": "strata", "engine_commit": "36fa455e579b23a9c909c2c6fe1bddd9e51cb8ca", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-36fa455e/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-HEAD/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": 79.56, "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": 79.4, "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": 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": 2863.9, "prompt_tok_s_client": 2843.7, "ttft_ms": 11266.444, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11187.0, "prompt_per_token_ms": 0.349, "prompt_per_second": 2863.9, "predicted_n": 256, "predicted_ms": 3224.2, "predicted_per_token_ms": 12.595, "predicted_per_second": 79.4, "draft_n": 229, "draft_n_accepted": 146}, "short_completion": false, "cache_flag": null, "completion_sha256": "0af6770ab0d206c8929273a6f90fc3f79700996eb0511a32f2e73d07e24467f7", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.925, "hits": 143596, "lookups": 155283, "drafts_accepted": 146, "drafts_offered": 229, "engine_generated": 256, "engine_decode_tok_s": 79.4, "engine_prompt_tok_s": 2863.9, "reused": 0, "metrics": {"hit_rate": 0.925, "decode_tok_s": 79.4, "engine_generated": 256, "prompt_ms": 11187.0, "decode_ms": 3224.2, "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": 1790988974.528, "t1": 1790988989.0, "seconds": 14.472, "host_samples": 15, "gpu_samples": 15}, "board": {"joules": 1989.64, "mean_w": 142.08, "max_w": 159.82, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 833.53, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 14.002}, "data_reads_per_s": {"mean": 11609.3, "max": 61051.3}, "data_read_mb_per_s": {"mean": 48.61}, "temps": {"gpu_max_c": 60.0, "ssd_max_c": 51.85}, "vram": {"max_used_mib": 23747.0, "baseline_mib": 15.0, "growth_mib": 23732.0, "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": 53698191360, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 8.43, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.209, "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-HEAD/speed/a1/run-005", "arm": "S-HEAD", "engine": "strata", "engine_commit": "36fa455e579b23a9c909c2c6fe1bddd9e51cb8ca", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-36fa455e/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-HEAD/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": 78.568, "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": 78.5, "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": 2874.8, "prompt_tok_s_client": 2856.1, "ttft_ms": 11217.268, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11144.4, "prompt_per_token_ms": 0.348, "prompt_per_second": 2874.8, "predicted_n": 256, "predicted_ms": 3261.4, "predicted_per_token_ms": 12.74, "predicted_per_second": 78.5, "draft_n": 211, "draft_n_accepted": 156}, "short_completion": false, "cache_flag": null, "completion_sha256": "aeed01158e35211670928afc00004092ec925272d0afd559a320adedc3dfb096", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9329999999999999, "hits": 134166, "lookups": 143814, "drafts_accepted": 156, "drafts_offered": 211, "engine_generated": 256, "engine_decode_tok_s": 78.5, "engine_prompt_tok_s": 2874.8, "reused": 0, "metrics": {"hit_rate": 0.933, "decode_tok_s": 78.5, "engine_generated": 256, "prompt_ms": 11144.4, "decode_ms": 3261.4, "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": 1790988989.003, "t1": 1790989003.466, "seconds": 14.463, "host_samples": 14, "gpu_samples": 14}, "board": {"joules": 1825.21, "mean_w": 140.37, "max_w": 149.29, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 664.25, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 13.002}, "data_reads_per_s": {"mean": 125.3, "max": 655.9}, "data_read_mb_per_s": {"mean": 0.52}, "temps": {"gpu_max_c": 63.0, "ssd_max_c": 50.85}, "vram": {"max_used_mib": 23747.0, "baseline_mib": 15.0, "growth_mib": 23732.0, "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": 53699686400, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 7.21, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.093, "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-HEAD/speed/a1/run-006", "arm": "S-HEAD", "engine": "strata", "engine_commit": "36fa455e579b23a9c909c2c6fe1bddd9e51cb8ca", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-36fa455e/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-HEAD/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": 92.126, "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": 92.0, "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": 2865.9, "prompt_tok_s_client": 2847.9, "ttft_ms": 11249.516, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11179.2, "prompt_per_token_ms": 0.349, "prompt_per_second": 2865.9, "predicted_n": 256, "predicted_ms": 2783.3, "predicted_per_token_ms": 10.872, "predicted_per_second": 92.0, "draft_n": 219, "draft_n_accepted": 155}, "short_completion": false, "cache_flag": null, "completion_sha256": "b2dfe5e04ed2eb6cb9d367231da30f2f23e970fa983fa44b4d0fbe19e967808a", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9329999999999999, "hits": 137756, "lookups": 147638, "drafts_accepted": 155, "drafts_offered": 219, "engine_generated": 256, "engine_decode_tok_s": 92.0, "engine_prompt_tok_s": 2865.9, "reused": 0, "metrics": {"hit_rate": 0.933, "decode_tok_s": 92.0, "engine_generated": 256, "prompt_ms": 11179.2, "decode_ms": 2783.3, "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": 1790989003.469, "t1": 1790989017.486, "seconds": 14.018, "host_samples": 14, "gpu_samples": 14}, "board": {"joules": 1846.78, "mean_w": 141.97, "max_w": 150.47, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 759.42, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 13.002}, "data_reads_per_s": {"mean": 82.9, "max": 647.9}, "data_read_mb_per_s": {"mean": 0.34}, "temps": {"gpu_max_c": 65.0, "ssd_max_c": 50.85}, "vram": {"max_used_mib": 23747.0, "baseline_mib": 15.0, "growth_mib": 23732.0, "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": 53699686400, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 6.29, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.118, "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-HEAD/speed/a1/run-007", "arm": "S-HEAD", "engine": "strata", "engine_commit": "36fa455e579b23a9c909c2c6fe1bddd9e51cb8ca", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-36fa455e/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-HEAD/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": 74.034, "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": 73.9, "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": 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": 2809.9, "prompt_tok_s_client": 2796.3, "ttft_ms": 45787.248, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 45566.5, "prompt_per_token_ms": 0.356, "prompt_per_second": 2809.9, "predicted_n": 256, "predicted_ms": 3462.0, "predicted_per_token_ms": 13.523, "predicted_per_second": 73.9, "draft_n": 218, "draft_n_accepted": 147}, "short_completion": false, "cache_flag": null, "completion_sha256": "e5c96183305f780a558c34ec2311bcd0bf699ba72049be9b49f682cf135892ad", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.917, "hits": 136803, "lookups": 149161, "drafts_accepted": 147, "drafts_offered": 218, "engine_generated": 256, "engine_decode_tok_s": 73.9, "engine_prompt_tok_s": 2809.9, "reused": 0, "metrics": {"hit_rate": 0.917, "decode_tok_s": 73.9, "engine_generated": 256, "prompt_ms": 45566.5, "decode_ms": 3462.0, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.349, "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": 1790989017.492, "t1": 1790989066.724, "seconds": 49.232, "host_samples": 49, "gpu_samples": 49}, "board": {"joules": 7046.68, "mean_w": 146.64, "max_w": 160.07, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 2373.26, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 48.006}, "data_reads_per_s": {"mean": 13687.5, "max": 78792.3}, "data_read_mb_per_s": {"mean": 57.34}, "temps": {"gpu_max_c": 69.0, "ssd_max_c": 51.85}, "vram": {"max_used_mib": 23747.0, "baseline_mib": 15.0, "growth_mib": 23732.0, "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": 54203224064, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.96, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.576, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.18, "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-HEAD/speed/a1/run-008", "arm": "S-HEAD", "engine": "strata", "engine_commit": "36fa455e579b23a9c909c2c6fe1bddd9e51cb8ca", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-36fa455e/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-HEAD/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": 72.293, "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": 72.1, "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": 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": 2784.8, "prompt_tok_s_client": 2771.1, "ttft_ms": 46204.345, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 45977.4, "prompt_per_token_ms": 0.359, "prompt_per_second": 2784.8, "predicted_n": 256, "predicted_ms": 3549.2, "predicted_per_token_ms": 13.864, "predicted_per_second": 72.1, "draft_n": 224, "draft_n_accepted": 151}, "short_completion": false, "cache_flag": null, "completion_sha256": "43f6a8939ddc5a17d459c6fb4bc5d1cc158d7e89a7056d0c71aae22b4a46ed73", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9209999999999999, "hits": 138006, "lookups": 149923, "drafts_accepted": 151, "drafts_offered": 224, "engine_generated": 256, "engine_decode_tok_s": 72.1, "engine_prompt_tok_s": 2784.8, "reused": 0, "metrics": {"hit_rate": 0.921, "decode_tok_s": 72.1, "engine_generated": 256, "prompt_ms": 45977.4, "decode_ms": 3549.2, "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": 1790989066.731, "t1": 1790989116.463, "seconds": 49.732, "host_samples": 50, "gpu_samples": 50}, "board": {"joules": 7143.6, "mean_w": 145.58, "max_w": 150.38, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 2425.21, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 49.007}, "data_reads_per_s": {"mean": 5163.7, "max": 44236.1}, "data_read_mb_per_s": {"mean": 21.8}, "temps": {"gpu_max_c": 70.0, "ssd_max_c": 52.85}, "vram": {"max_used_mib": 23747.0, "baseline_mib": 15.0, "growth_mib": 23732.0, "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": 54151458816, "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": 2.513, "user_psi_full_avg10_max": 0.0, "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": 135.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-HEAD/speed/a1/run-009", "arm": "S-HEAD", "engine": "strata", "engine_commit": "36fa455e579b23a9c909c2c6fe1bddd9e51cb8ca", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-36fa455e/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-HEAD/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": 75.329, "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": 75.2, "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": 2789.5, "prompt_tok_s_client": 2776.5, "ttft_ms": 46113.949, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 45899.7, "prompt_per_token_ms": 0.358, "prompt_per_second": 2789.5, "predicted_n": 256, "predicted_ms": 3402.8, "predicted_per_token_ms": 13.292, "predicted_per_second": 75.2, "draft_n": 245, "draft_n_accepted": 145}, "short_completion": false, "cache_flag": null, "completion_sha256": "d05cbf9f990e42ce79c3e76056c7e3d58e332bdaf2fd60f9fc66445db9d420ab", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9209999999999999, "hits": 149791, "lookups": 162606, "drafts_accepted": 145, "drafts_offered": 245, "engine_generated": 256, "engine_decode_tok_s": 75.2, "engine_prompt_tok_s": 2789.5, "reused": 0, "metrics": {"hit_rate": 0.921, "decode_tok_s": 75.2, "engine_generated": 256, "prompt_ms": 45899.7, "decode_ms": 3402.8, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.306, "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": 1790989116.466, "t1": 1790989165.966, "seconds": 49.5, "host_samples": 50, "gpu_samples": 49}, "board": {"joules": 7058.95, "mean_w": 146.89, "max_w": 161.83, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 2313.13, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 49.009}, "data_reads_per_s": {"mean": 4942.6, "max": 21667.3}, "data_read_mb_per_s": {"mean": 20.98}, "temps": {"gpu_max_c": 72.0, "ssd_max_c": 53.85}, "vram": {"max_used_mib": 23747.0, "baseline_mib": 15.0, "growth_mib": 23732.0, "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": 54153199616, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.67, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.529, "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"}}}