{"row_id": "S-IQ3/speed/a2/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/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": 78.008, "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": 77.6, "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": 1897.3, "prompt_tok_s_client": 1859.8, "ttft_ms": 2171.175, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 2128.3, "prompt_per_token_ms": 0.527, "prompt_per_second": 1897.3, "predicted_n": 256, "predicted_ms": 3298.5, "predicted_per_token_ms": 12.885, "predicted_per_second": 77.6, "draft_n": 247, "draft_n_accepted": 159}, "short_completion": false, "cache_flag": null, "completion_sha256": "83316581364a0d7a32aadd157727670690c843988addb1ac33abda79fd85706b", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.873, "hits": 133644, "lookups": 153006, "drafts_accepted": 159, "drafts_offered": 247, "engine_generated": 256, "engine_decode_tok_s": 77.6, "engine_prompt_tok_s": 1897.3, "reused": 0, "metrics": {"hit_rate": 0.873, "decode_tok_s": 77.6, "engine_generated": 256, "prompt_ms": 2128.3, "decode_ms": 3298.5, "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": 1791012325.266, "t1": 1791012330.706, "seconds": 5.44, "host_samples": 6, "gpu_samples": 6}, "board": {"joules": 566.69, "mean_w": 113.11, "max_w": 136.84, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 363.21, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 5.004}, "data_reads_per_s": {"mean": 8547.8, "max": 47637.5}, "data_read_mb_per_s": {"mean": 36.1}, "temps": {"gpu_max_c": 41.0, "ssd_max_c": 40.85}, "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": 53331390464, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 1.85, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.836, "user_psi_full_avg10_max": 1.75, "psi_full_avg10_max": 2.34, "unit_psi_full_avg10_max": 0.88, "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/a2/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/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": 91.276, "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": 91.1, "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": 2190.2, "prompt_tok_s_client": 2158.1, "ttft_ms": 1871.071, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1843.7, "prompt_per_token_ms": 0.457, "prompt_per_second": 2190.2, "predicted_n": 256, "predicted_ms": 2809.9, "predicted_per_token_ms": 10.976, "predicted_per_second": 91.1, "draft_n": 217, "draft_n_accepted": 154}, "short_completion": false, "cache_flag": null, "completion_sha256": "ac942ab2f1e99e9adc63f7e9c63bf7390fefb510a01d0d1b0e5b53af15a94e99", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9179999999999999, "hits": 133745, "lookups": 145639, "drafts_accepted": 154, "drafts_offered": 217, "engine_generated": 256, "engine_decode_tok_s": 91.1, "engine_prompt_tok_s": 2190.2, "reused": 0, "metrics": {"hit_rate": 0.918, "decode_tok_s": 91.1, "engine_generated": 256, "prompt_ms": 1843.7, "decode_ms": 2809.9, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.51, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1791012330.708, "t1": 1791012335.373, "seconds": 4.665, "host_samples": 4, "gpu_samples": 4}, "board": {"joules": 381.4, "mean_w": 127.02, "max_w": 134.41, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 222.44, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 3.005}, "data_reads_per_s": {"mean": 301.3, "max": 631.4}, "data_read_mb_per_s": {"mean": 1.36}, "temps": {"gpu_max_c": 44.0, "ssd_max_c": 39.85}, "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": 53371977728, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 3.9, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.83, "user_psi_full_avg10_max": 0.96, "psi_full_avg10_max": 1.4, "unit_psi_full_avg10_max": 0.48, "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/a2/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/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": 93.03, "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": 92.8, "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": 2196.6, "prompt_tok_s_client": 2162.3, "ttft_ms": 1867.43, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1838.3, "prompt_per_token_ms": 0.455, "prompt_per_second": 2196.6, "predicted_n": 256, "predicted_ms": 2759.3, "predicted_per_token_ms": 10.779, "predicted_per_second": 92.8, "draft_n": 222, "draft_n_accepted": 161}, "short_completion": false, "cache_flag": null, "completion_sha256": "bb872dbe1f3284f09fa775ba2eb1a8e0ffc8f2f679fe721bc6d3663b6154a950", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9209999999999999, "hits": 133487, "lookups": 144912, "drafts_accepted": 161, "drafts_offered": 222, "engine_generated": 256, "engine_decode_tok_s": 92.8, "engine_prompt_tok_s": 2196.6, "reused": 0, "metrics": {"hit_rate": 0.921, "decode_tok_s": 92.8, "engine_generated": 256, "prompt_ms": 1838.3, "decode_ms": 2759.3, "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": 1791012335.375, "t1": 1791012339.983, "seconds": 4.609, "host_samples": 5, "gpu_samples": 5}, "board": {"joules": 507.27, "mean_w": 126.67, "max_w": 133.78, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 190.75, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 4.003}, "data_reads_per_s": {"mean": 355.0, "max": 1143.8}, "data_read_mb_per_s": {"mean": 1.48}, "temps": {"gpu_max_c": 46.0, "ssd_max_c": 39.85}, "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": 53372035072, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 6.57, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.826, "user_psi_full_avg10_max": 0.64, "psi_full_avg10_max": 0.94, "unit_psi_full_avg10_max": 0.32, "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": 137.16, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-IQ3/speed/a2/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/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": 90.292, "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": 90.1, "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": 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": 2694.8, "prompt_tok_s_client": 2678.2, "ttft_ms": 11962.383, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11888.9, "prompt_per_token_ms": 0.371, "prompt_per_second": 2694.8, "predicted_n": 256, "predicted_ms": 2842.2, "predicted_per_token_ms": 11.102, "predicted_per_second": 90.1, "draft_n": 212, "draft_n_accepted": 154}, "short_completion": false, "cache_flag": null, "completion_sha256": "803ceebbcd49e5ff13a1646d3675215eb4f5e9a0ed765584f94330f154e71d1b", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9259999999999999, "hits": 134001, "lookups": 144710, "drafts_accepted": 154, "drafts_offered": 212, "engine_generated": 256, "engine_decode_tok_s": 90.1, "engine_prompt_tok_s": 2694.8, "reused": 0, "metrics": {"hit_rate": 0.926, "decode_tok_s": 90.1, "engine_generated": 256, "prompt_ms": 11888.9, "decode_ms": 2842.2, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.51, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1791012339.985, "t1": 1791012354.771, "seconds": 14.787, "host_samples": 15, "gpu_samples": 15}, "board": {"joules": 1970.63, "mean_w": 140.58, "max_w": 149.28, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 602.07, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 14.005}, "data_reads_per_s": {"mean": 11598.1, "max": 96457.0}, "data_read_mb_per_s": {"mean": 48.58}, "temps": {"gpu_max_c": 52.0, "ssd_max_c": 38.85}, "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": 53547872256, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 7.6, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.599, "user_psi_full_avg10_max": 0.35, "psi_full_avg10_max": 0.51, "unit_psi_full_avg10_max": 0.21, "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/a2/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/a2/strata-config.json", "--port", "withheld-2026-10-13"], "ctx": "32K", "repeat": 1, "run": 5, "slice_sha256": "b72b07858a845e76331ec38d46abae3c09d3e1cb6e4d968e887fa8b132c64977", "prompt_sha256": "c8a1ccb2149bdf9595fe8d97b7183fd9efd61c32259e907606ddbac7710955b2", "nonce": "Ticket 97542637.", "requests_since_start": 5, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 95.57, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 95.3, "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": 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": 2715.0, "prompt_tok_s_client": 2699.0, "ttft_ms": 11870.137, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11800.4, "prompt_per_token_ms": 0.368, "prompt_per_second": 2715.0, "predicted_n": 256, "predicted_ms": 2685.3, "predicted_per_token_ms": 10.489, "predicted_per_second": 95.3, "draft_n": 212, "draft_n_accepted": 159}, "short_completion": false, "cache_flag": null, "completion_sha256": "17e03871bda508f10ba1eb0913b50b3be1c8663c9611f8f7b01c756611911d27", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9309999999999999, "hits": 132900, "lookups": 142747, "drafts_accepted": 159, "drafts_offered": 212, "engine_generated": 256, "engine_decode_tok_s": 95.3, "engine_prompt_tok_s": 2715.0, "reused": 0, "metrics": {"hit_rate": 0.931, "decode_tok_s": 95.3, "engine_generated": 256, "prompt_ms": 11800.4, "decode_ms": 2685.3, "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": 1791012354.773, "t1": 1791012369.311, "seconds": 14.539, "host_samples": 14, "gpu_samples": 14}, "board": {"joules": 1869.76, "mean_w": 143.6, "max_w": 154.46, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 431.8, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 13.003}, "data_reads_per_s": {"mean": 112.4, "max": 775.2}, "data_read_mb_per_s": {"mean": 0.47}, "temps": {"gpu_max_c": 53.0, "ssd_max_c": 38.85}, "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": 53559898112, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 8.3, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.494, "user_psi_full_avg10_max": 0.08, "psi_full_avg10_max": 0.12, "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/a2/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/a2/strata-config.json", "--port", "withheld-2026-10-13"], "ctx": "32K", "repeat": 2, "run": 6, "slice_sha256": "b72b07858a845e76331ec38d46abae3c09d3e1cb6e4d968e887fa8b132c64977", "prompt_sha256": "d6c964051c51baa9ef5cb009be0a921422c2b28350d2dbea610a6038552bdc18", "nonce": "Ticket 24502846.", "requests_since_start": 6, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 92.534, "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": 92.4, "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": 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": 2704.0, "prompt_tok_s_client": 2689.1, "ttft_ms": 11914.124, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11848.3, "prompt_per_token_ms": 0.37, "prompt_per_second": 2704.0, "predicted_n": 256, "predicted_ms": 2770.8, "predicted_per_token_ms": 10.823, "predicted_per_second": 92.4, "draft_n": 211, "draft_n_accepted": 155}, "short_completion": false, "cache_flag": null, "completion_sha256": "016156593999cc72fff7632c90fddca26c6b243cf4491a6b5eb99b68da675fd0", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9329999999999999, "hits": 135157, "lookups": 144898, "drafts_accepted": 155, "drafts_offered": 211, "engine_generated": 256, "engine_decode_tok_s": 92.4, "engine_prompt_tok_s": 2704.0, "reused": 0, "metrics": {"hit_rate": 0.933, "decode_tok_s": 92.4, "engine_generated": 256, "prompt_ms": 11848.3, "decode_ms": 2770.8, "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": 1791012369.313, "t1": 1791012383.983, "seconds": 14.67, "host_samples": 15, "gpu_samples": 15}, "board": {"joules": 1986.9, "mean_w": 141.89, "max_w": 149.89, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 617.93, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 14.005}, "data_reads_per_s": {"mean": 52.6, "max": 416.9}, "data_read_mb_per_s": {"mean": 0.26}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": 39.85}, "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": 53559898112, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 9.17, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.559, "user_psi_full_avg10_max": 0.02, "psi_full_avg10_max": 0.03, "unit_psi_full_avg10_max": 0.0, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 138.65, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "S-IQ3/speed/a2/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/a2/strata-config.json", "--port", "withheld-2026-10-13"], "ctx": "128K", "repeat": 0, "run": 7, "slice_sha256": "286827657f5df11459f4ee550246f7e0e06ce7af3238fe74a53f0aa4c12676ac", "prompt_sha256": "2491e963527b8a39f7fff1aa2d04dab5eb4dd3ca12c7417c451f6216c665972c", "nonce": "Ticket 14239051.", "requests_since_start": 7, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 81.517, "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": 81.4, "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": 2739.8, "prompt_tok_s_client": 2727.0, "ttft_ms": 46951.299, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 46732.7, "prompt_per_token_ms": 0.365, "prompt_per_second": 2739.8, "predicted_n": 256, "predicted_ms": 3146.2, "predicted_per_token_ms": 12.29, "predicted_per_second": 81.4, "draft_n": 223, "draft_n_accepted": 154}, "short_completion": false, "cache_flag": null, "completion_sha256": "8a647c5bea2d66201bb081245115cadb90e6360a739449773b6aa689b5917cfa", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9059999999999999, "hits": 134085, "lookups": 148069, "drafts_accepted": 154, "drafts_offered": 223, "engine_generated": 256, "engine_decode_tok_s": 81.4, "engine_prompt_tok_s": 2739.8, "reused": 0, "metrics": {"hit_rate": 0.906, "decode_tok_s": 81.4, "engine_generated": 256, "prompt_ms": 46732.7, "decode_ms": 3146.2, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.51, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1791012383.987, "t1": 1791012434.067, "seconds": 50.08, "host_samples": 50, "gpu_samples": 50}, "board": {"joules": 7140.61, "mean_w": 145.58, "max_w": 151.14, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 1727.11, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 49.012}, "data_reads_per_s": {"mean": 13414.1, "max": 78761.6}, "data_read_mb_per_s": {"mean": 56.13}, "temps": {"gpu_max_c": 61.0, "ssd_max_c": 43.85}, "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": 54015680512, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 6.61, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 3.038, "user_psi_full_avg10_max": 9.64, "psi_full_avg10_max": 9.52, "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"}}}