{"row_id": "S-HEAD/speed/a2/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/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": 76.731, "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": 76.4, "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": 1983.1, "prompt_tok_s_client": 1943.5, "ttft_ms": 2077.721, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 2036.2, "prompt_per_token_ms": 0.504, "prompt_per_second": 1983.1, "predicted_n": 256, "predicted_ms": 3351.9, "predicted_per_token_ms": 13.093, "predicted_per_second": 76.4, "draft_n": 231, "draft_n_accepted": 160}, "short_completion": false, "cache_flag": null, "completion_sha256": "02d9206e8edf9acc11db0b7264c2839bf5d800a61f73521207ff62bb32723ddd", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.873, "hits": 126656, "lookups": 145088, "drafts_accepted": 160, "drafts_offered": 231, "engine_generated": 256, "engine_decode_tok_s": 76.4, "engine_prompt_tok_s": 1983.1, "reused": 0, "metrics": {"hit_rate": 0.873, "decode_tok_s": 76.4, "engine_generated": 256, "prompt_ms": 2036.2, "decode_ms": 3351.9, "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": 1791007197.651, "t1": 1791007203.052, "seconds": 5.401, "host_samples": 6, "gpu_samples": 6}, "board": {"joules": 525.13, "mean_w": 104.93, "max_w": 147.4, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 375.37, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 5.003}, "data_reads_per_s": {"mean": 8572.7, "max": 47650.2}, "data_read_mb_per_s": {"mean": 36.39}, "temps": {"gpu_max_c": 45.0, "ssd_max_c": 42.85}, "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": 53163311104, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 2.52, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.99, "user_psi_full_avg10_max": 3.22, "psi_full_avg10_max": 4.46, "unit_psi_full_avg10_max": 0.47, "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/a2/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/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": 86.292, "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": 86.2, "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": 2326.6, "prompt_tok_s_client": 2282.4, "ttft_ms": 1769.186, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1735.6, "prompt_per_token_ms": 0.43, "prompt_per_second": 2326.6, "predicted_n": 256, "predicted_ms": 2969.7, "predicted_per_token_ms": 11.6, "predicted_per_second": 86.2, "draft_n": 236, "draft_n_accepted": 157}, "short_completion": false, "cache_flag": null, "completion_sha256": "7b824d095cc81409609a68f1f23002bc7a38516b6033e8c065c9619713692a8e", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9129999999999999, "hits": 138772, "lookups": 152046, "drafts_accepted": 157, "drafts_offered": 236, "engine_generated": 256, "engine_decode_tok_s": 86.2, "engine_prompt_tok_s": 2326.5, "reused": 0, "metrics": {"hit_rate": 0.913, "decode_tok_s": 86.2, "engine_generated": 256, "prompt_ms": 1735.6, "decode_ms": 2969.7, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.586, "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": 1791007203.055, "t1": 1791007207.779, "seconds": 4.724, "host_samples": 4, "gpu_samples": 4}, "board": {"joules": 392.5, "mean_w": 130.13, "max_w": 148.79, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 247.83, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 3.003}, "data_reads_per_s": {"mean": 564.2, "max": 1056.0}, "data_read_mb_per_s": {"mean": 2.36}, "temps": {"gpu_max_c": 48.0, "ssd_max_c": 41.85}, "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": 53264244736, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 4.55, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.828, "user_psi_full_avg10_max": 1.77, "psi_full_avg10_max": 2.45, "unit_psi_full_avg10_max": 0.25, "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/a2/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/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": 90.2, "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": 90.0, "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": 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": 2271.1, "prompt_tok_s_client": 2225.7, "ttft_ms": 1814.287, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1778.0, "prompt_per_token_ms": 0.44, "prompt_per_second": 2271.1, "predicted_n": 256, "predicted_ms": 2844.4, "predicted_per_token_ms": 11.111, "predicted_per_second": 90.0, "draft_n": 226, "draft_n_accepted": 159}, "short_completion": false, "cache_flag": null, "completion_sha256": "dd8c25032b410d8e96d70f1e44e1c08610c0eb6ec83d0703e5aed3e61b8d289f", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.925, "hits": 137088, "lookups": 148186, "drafts_accepted": 159, "drafts_offered": 226, "engine_generated": 256, "engine_decode_tok_s": 90.0, "engine_prompt_tok_s": 2271.1, "reused": 0, "metrics": {"hit_rate": 0.925, "decode_tok_s": 90.0, "engine_generated": 256, "prompt_ms": 1778.0, "decode_ms": 2844.4, "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": 1791007207.782, "t1": 1791007212.423, "seconds": 4.642, "host_samples": 5, "gpu_samples": 5}, "board": {"joules": 514.47, "mean_w": 128.58, "max_w": 133.2, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 210.69, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 4.003}, "data_reads_per_s": {"mean": 333.4, "max": 862.9}, "data_read_mb_per_s": {"mean": 1.53}, "temps": {"gpu_max_c": 48.0, "ssd_max_c": 41.85}, "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": 53264244736, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 4.58, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.128, "user_psi_full_avg10_max": 1.18, "psi_full_avg10_max": 1.82, "unit_psi_full_avg10_max": 0.17, "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/a2/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/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": 80.806, "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": 80.7, "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": 2898.5, "prompt_tok_s_client": 2881.0, "ttft_ms": 11120.496, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11053.2, "prompt_per_token_ms": 0.345, "prompt_per_second": 2898.5, "predicted_n": 256, "predicted_ms": 3173.0, "predicted_per_token_ms": 12.395, "predicted_per_second": 80.7, "draft_n": 220, "draft_n_accepted": 151}, "short_completion": false, "cache_flag": null, "completion_sha256": "e3cdde9fd2287da6e2ead8c329b64adbfc0655de9cf614bf07aec53e25b9827e", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9159999999999999, "hits": 135299, "lookups": 147649, "drafts_accepted": 151, "drafts_offered": 220, "engine_generated": 256, "engine_decode_tok_s": 80.7, "engine_prompt_tok_s": 2898.5, "reused": 0, "metrics": {"hit_rate": 0.916, "decode_tok_s": 80.7, "engine_generated": 256, "prompt_ms": 11053.2, "decode_ms": 3173.0, "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": 1791007212.426, "t1": 1791007226.702, "seconds": 14.276, "host_samples": 14, "gpu_samples": 14}, "board": {"joules": 1806.64, "mean_w": 138.93, "max_w": 148.54, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 818.5, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 13.006}, "data_reads_per_s": {"mean": 12414.8, "max": 96572.1}, "data_read_mb_per_s": {"mean": 51.99}, "temps": {"gpu_max_c": 55.0, "ssd_max_c": 41.85}, "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": 53551611904, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.44, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.13, "user_psi_full_avg10_max": 4.64, "psi_full_avg10_max": 4.82, "unit_psi_full_avg10_max": 0.09, "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"}}}