[
  {
    "id": "2026-07-11-mtp-workloads",
    "date": "2026-07-11",
    "machineId": "dgx-spark-a",
    "model": "Qwen3.6-35B-A3B · Q6_K · native MTP",
    "engine": "llama.cpp b9967 · single-slot quiet-box comparison",
    "task": "Speculative decoding · code versus creative writing",
    "metrics": {
      "code_off_tok_s": 64.1,
      "code_on_tok_s": 100.9,
      "creative_off_tok_s": 64.6,
      "creative_on_tok_s": 53.6
    },
    "source": "Local Spark tuning review, measured 2026-07-11; reviewed directly 2026-10-02.",
    "notes": "Historical local A/B with six draft positions. Code used temperature 0.3; creative writing used 0.7. MTP helped code but hurt creative throughput. This does not qualify a different model, runtime, or sampling policy.",
    "kind": "historical",
    "pageLastModified": "2026-07-11",
    "journalSlug": "2026-07-11-speculative-decoding",
    "methodology": "Same model and quiet hardware; one slot. Workload and sampling settings are explicitly different between the code and creative probes.",
    "limitations": "Repeat count and spread were not recorded in the reviewed summary; do not treat this as a universal speedup.",
    "suiteIds": [
      "concurrency"
    ]
  },
  {
    "id": "2026-07-19-agent-gemma",
    "date": "2026-07-19",
    "machineId": "lab-pool",
    "model": "Gemma-4-26B-A4B QAT · always-on agent",
    "engine": "llama.cpp · four 163840 slots · dated bench page",
    "task": "Jul 19 multimodal creative agent bakeoff",
    "metrics": {
      "single_tok_s": 60.4,
      "four_way_tok_s": 129.9,
      "tools_pct": 91.2,
      "blind_score": 8.35,
      "warm_prefix_s": 0.303,
      "deep_edge_s": 78.3
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical. Gemma kept as always-on agent. Goetia and Fable rejected. Warm 17K prefix 0.303 s (23.2× over cold). Deep edge 78.3 s · 119,593 prompt tokens. Catalog ~65 tok/s on the later agent-gemma on-demand id is a separate roster figure. Not live Oct 1 telemetry.",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-07-19-agent-gemma",
    "suiteIds": [
      "broad-tools",
      "blind-editor",
      "context-retrieval",
      "concurrency"
    ]
  },
  {
    "id": "2026-07-21-coder-2x256k",
    "date": "2026-07-21",
    "machineId": "lab-pool",
    "model": "Qwen3-Coder-Next · 2 × 262144",
    "engine": "llama.cpp · dated bench page",
    "task": "Jul 21 two deep coder slots",
    "metrics": {
      "single_tok_s": 52.6,
      "aggregate_tok_s": 76,
      "warm_turn_s": 4.3,
      "retrieval_tokens": 241000
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical qualification. Both deep sessions retrieved correctly near 241K. About 55.2 GiB. Warm-session affinity, not a 2× cold accelerator. Not live Oct 1 telemetry.",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-07-21-coder-lanes",
    "suiteIds": [
      "context-retrieval",
      "concurrency"
    ]
  },
  {
    "id": "2026-07-21-laguna-s-reject",
    "date": "2026-07-21",
    "machineId": "lab-pool",
    "model": "Laguna S 2.1 NVFP4 vs Qwen3-Coder-Next",
    "engine": "dated bench page · incl. TP=2",
    "task": "Jul 21 Laguna S first reject",
    "metrics": {
      "laguna_code_pct": 79.7,
      "coder_next_code_pct": 87.8,
      "laguna_c1_tok_s": 18.6,
      "coder_next_c4_agg_tok_s": 109.9,
      "laguna_c4_agg_tok_s": 57.8
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical. Not a quality or throughput upgrade. Native TP=2 deadlocked on 512-token generations. DFlash depth 15 cut four-way throughput 57.8 → 46.2 tok/s. Not live Oct 1 telemetry.",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-07-21-coder-lanes",
    "suiteIds": [
      "framework-code",
      "concurrency"
    ]
  },
  {
    "id": "2026-07-21-three-lane",
    "date": "2026-07-21",
    "machineId": "lab-pool",
    "model": "three-lane coder dispatcher",
    "engine": "capacity-aware proxy · dated bench page",
    "task": "Jul 21 three independent coder lanes",
    "metrics": {
      "aggregate_tok_s": 99.6,
      "per_stream_low_tok_s": 42,
      "per_stream_high_tok_s": 56.8,
      "http_409s": 0
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical. Two coding-peer slots plus one local lane. Four concurrent 256-token streams; excess queued, no 409s. Session affinity keeps cache local. Not live Oct 1 telemetry.",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-07-21-coder-lanes",
    "suiteIds": [
      "framework-code",
      "context-retrieval",
      "concurrency"
    ]
  },
  {
    "id": "2026-07-22-mage-flow",
    "date": "2026-07-22",
    "machineId": "lab-pool",
    "model": "Mage-Flow Turbo 4B · Z-Image-Turbo · Sana Sprint 1.6B",
    "engine": "local 1024² pipeline · dated bench page",
    "task": "Jul 22 image-gen gate",
    "metrics": {
      "mage_warm_s": 3.96,
      "mage_peak_gib": 17.03,
      "zimage_warm_s": 5.7,
      "zimage_peak_gib": 21.69,
      "sana_warm_s": 0.79,
      "sana_peak_gib": 10.19
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical. Mage was the only tested model combining high visual quality, sub-10 s 1024², and under-20 GiB peak. Not uncensored (life-drawing refusal). Later moved on demand — warm ~4 s, cold ~120 s. Not live Oct 1 telemetry.",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-07-22-image-and-refresh",
    "suiteIds": [
      "image-generation"
    ]
  },
  {
    "id": "2026-07-26-coder-revert",
    "date": "2026-07-26",
    "machineId": "lab-pool",
    "model": "Laguna S 2.1 NVFP4 vs Qwen3-Coder-Next UD-Q4_K_XL",
    "engine": "same-node A/B · dated bench page",
    "task": "Jul 26 Coder-Next restore / Laguna demote",
    "metrics": {
      "laguna_c1_tok_s": 17.7,
      "coder_next_c1_tok_s": 52.7,
      "laguna_devcode_pct": 87.2,
      "coder_next_devcode_pct": 90.9,
      "laguna_wall_s": 435.7,
      "coder_next_wall_s": 82
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical same-day A/B. Coder-Next 2.98× decode and 5.3× wall; 47 GiB reclaimed (101 vs 53.9 GiB). Laguna kept on demand for tools 75.0% then retired July 30. Not live Oct 1 telemetry.",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-07-26-coder-revert",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ]
  },
  {
    "id": "2026-07-27-gate-coder-next",
    "date": "2026-07-27",
    "machineId": "lab-pool",
    "model": "Qwen3-Coder-Next 80B-A3B · UD-Q4_K_XL",
    "engine": "llama.cpp · batch tuning gate at ubatch 2048",
    "task": "Archived coder gate · ub2048",
    "metrics": {
      "framework_pct": 87.8,
      "developer_tools_pct": 66.7,
      "c1_tok_s": 51.1,
      "c2_tok_s": 77,
      "c4_tok_s": 78.2,
      "four_request_mean_decode_tok_s": 40,
      "framework_wall_s": 88,
      "gate_first_response_s": 1,
      "bash_framework_pct": 88.9,
      "nuxt_framework_pct": 100,
      "repo_framework_pct": 100,
      "swift_framework_pct": 67.8,
      "tailwind_framework_pct": 77.8,
      "vue_framework_pct": 100
    },
    "source": "Retained numeric coder-gate summary, 2026-07-27; reviewed October 10, 2026.",
    "notes": "Measured historical gate. Single tuning gate. Six-run calibration median 86.4% is a separate summary.",
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ],
    "methodology": "Eight framework prompts scored by regex, twelve developer tool scenarios, and warm lighthouse-prose concurrency waves. Defaults: framework temperature 0.2 / output 4096; tools 0.1 / 1024; concurrency 0.7 / 512; thinking off. c1/c2/c4 divide total completion positions by whole-wave wall time. Mean per-request decode excludes each request’s TTFT. Gate first response is the short ping, not an independently measured model load time.",
    "limitations": "No code compilation or actual tool execution. One report does not establish repeat spread or context qualification. Different slot counts, runtime revisions and dates limit cross-engine comparisons."
  },
  {
    "id": "2026-07-29-gate-qwen35-base",
    "date": "2026-07-29",
    "machineId": "lab-pool",
    "model": "Qwen3.5-27B · Q4_K_M",
    "engine": "llama.cpp · matched base A/B",
    "task": "Archived coder gate · qwen35-27b-base-q4",
    "metrics": {
      "framework_pct": 84.3,
      "developer_tools_pct": 75,
      "c1_tok_s": 10.8,
      "c2_tok_s": 10.9,
      "c4_tok_s": 11,
      "four_request_mean_decode_tok_s": 11.1,
      "framework_wall_s": 484,
      "gate_first_response_s": 1,
      "bash_framework_pct": 88.9,
      "nuxt_framework_pct": 100,
      "repo_framework_pct": 50,
      "swift_framework_pct": 73.2,
      "tailwind_framework_pct": 88.9,
      "vue_framework_pct": 100
    },
    "source": "Retained numeric coder-gate summary, 2026-07-29; reviewed October 10, 2026.",
    "notes": "Measured historical gate. Matched base for Qwythos: no meaningful quality or throughput gain from the challenger.",
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ],
    "methodology": "Eight framework prompts scored by regex, twelve developer tool scenarios, and warm lighthouse-prose concurrency waves. Defaults: framework temperature 0.2 / output 4096; tools 0.1 / 1024; concurrency 0.7 / 512; thinking off. c1/c2/c4 divide total completion positions by whole-wave wall time. Mean per-request decode excludes each request’s TTFT. Gate first response is the short ping, not an independently measured model load time.",
    "limitations": "No code compilation or actual tool execution. One report does not establish repeat spread or context qualification. Different slot counts, runtime revisions and dates limit cross-engine comparisons."
  },
  {
    "id": "2026-07-29-gate-qwythos",
    "date": "2026-07-29",
    "machineId": "lab-pool",
    "model": "Empero Qwythos-27B-v1 · Q4_K_M",
    "engine": "llama.cpp · matched base A/B",
    "task": "Archived coder gate · qwythos-27b-q4",
    "metrics": {
      "framework_pct": 84.6,
      "developer_tools_pct": 75,
      "c1_tok_s": 11,
      "c2_tok_s": 11.1,
      "c4_tok_s": 11.1,
      "four_request_mean_decode_tok_s": 11.2,
      "framework_wall_s": 407,
      "gate_first_response_s": 1,
      "bash_framework_pct": 88.9,
      "nuxt_framework_pct": 100,
      "repo_framework_pct": 75,
      "swift_framework_pct": 73.2,
      "tailwind_framework_pct": 66.7,
      "vue_framework_pct": 100
    },
    "source": "Retained numeric coder-gate summary, 2026-07-29; reviewed October 10, 2026.",
    "notes": "Measured historical gate. Framework and tools within the matched base result; rejected for the historical coder/agent roles.",
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ],
    "methodology": "Eight framework prompts scored by regex, twelve developer tool scenarios, and warm lighthouse-prose concurrency waves. Defaults: framework temperature 0.2 / output 4096; tools 0.1 / 1024; concurrency 0.7 / 512; thinking off. c1/c2/c4 divide total completion positions by whole-wave wall time. Mean per-request decode excludes each request’s TTFT. Gate first response is the short ping, not an independently measured model load time.",
    "limitations": "No code compilation or actual tool execution. One report does not establish repeat spread or context qualification. Different slot counts, runtime revisions and dates limit cross-engine comparisons."
  },
  {
    "id": "2026-07-30-gate-laguna-xs",
    "date": "2026-07-30",
    "machineId": "lab-pool",
    "model": "Poolside Laguna-XS-2.1 33B · NVFP4",
    "engine": "vLLM · exact revision not frozen",
    "task": "Archived coder gate · coder-laguna-xs",
    "metrics": {
      "framework_pct": 83.9,
      "developer_tools_pct": 50,
      "c1_tok_s": 35.3,
      "c2_tok_s": 70.1,
      "c4_tok_s": 127.7,
      "four_request_mean_decode_tok_s": 32.3,
      "framework_wall_s": 141,
      "gate_first_response_s": 200,
      "bash_framework_pct": 88.9,
      "nuxt_framework_pct": 85.7,
      "repo_framework_pct": 100,
      "swift_framework_pct": 73.2,
      "tailwind_framework_pct": 66.7,
      "vue_framework_pct": 91.7
    },
    "source": "Retained numeric coder-gate summary, 2026-07-30; reviewed October 10, 2026.",
    "notes": "Measured historical gate. Rejected: slow cold readiness and single-stream latency. Four-slot vLLM wave is not a model-only comparison with two-slot llama.cpp.",
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ],
    "methodology": "Eight framework prompts scored by regex, twelve developer tool scenarios, and warm lighthouse-prose concurrency waves. Defaults: framework temperature 0.2 / output 4096; tools 0.1 / 1024; concurrency 0.7 / 512; thinking off. c1/c2/c4 divide total completion positions by whole-wave wall time. Mean per-request decode excludes each request’s TTFT. Gate first response is the short ping, not an independently measured model load time.",
    "limitations": "No code compilation or actual tool execution. One report does not establish repeat spread or context qualification. Different slot counts, runtime revisions and dates limit cross-engine comparisons."
  },
  {
    "id": "2026-07-30-gate-ornith",
    "date": "2026-07-30",
    "machineId": "lab-pool",
    "model": "Ornith-1.0-35B · UD-Q4_K_XL",
    "engine": "llama.cpp · tuned gate",
    "task": "Archived coder gate · coder-ornith",
    "metrics": {
      "framework_pct": 84.2,
      "developer_tools_pct": 75,
      "c1_tok_s": 61.6,
      "c2_tok_s": 93.9,
      "c4_tok_s": 94.2,
      "four_request_mean_decode_tok_s": 48.1,
      "framework_wall_s": 115,
      "gate_first_response_s": 27,
      "bash_framework_pct": 88.9,
      "nuxt_framework_pct": 100,
      "repo_framework_pct": 100,
      "swift_framework_pct": 53.5,
      "tailwind_framework_pct": 77.8,
      "vue_framework_pct": 100
    },
    "source": "Retained numeric coder-gate summary, 2026-07-30; reviewed October 10, 2026.",
    "notes": "Measured historical gate. One retained run. Five-run framework median 84.2%, Swift median 60.7% and framework standard deviation 5.55 points are separate summaries.",
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ],
    "methodology": "Eight framework prompts scored by regex, twelve developer tool scenarios, and warm lighthouse-prose concurrency waves. Defaults: framework temperature 0.2 / output 4096; tools 0.1 / 1024; concurrency 0.7 / 512; thinking off. c1/c2/c4 divide total completion positions by whole-wave wall time. Mean per-request decode excludes each request’s TTFT. Gate first response is the short ping, not an independently measured model load time.",
    "limitations": "No code compilation or actual tool execution. One report does not establish repeat spread or context qualification. Different slot counts, runtime revisions and dates limit cross-engine comparisons."
  },
  {
    "id": "2026-07-30-gate-qwen36",
    "date": "2026-07-30",
    "machineId": "lab-pool",
    "model": "Qwen3.6-35B-A3B · UD-Q4_K_XL",
    "engine": "llama.cpp · tuned rematch",
    "task": "Archived coder gate · coder-qwen36",
    "metrics": {
      "framework_pct": 81.7,
      "developer_tools_pct": 75,
      "c1_tok_s": 62.4,
      "c2_tok_s": 94.7,
      "c4_tok_s": 95.7,
      "four_request_mean_decode_tok_s": 49,
      "framework_wall_s": 112,
      "gate_first_response_s": 17,
      "bash_framework_pct": 100,
      "nuxt_framework_pct": 85.7,
      "repo_framework_pct": 100,
      "swift_framework_pct": 61.6,
      "tailwind_framework_pct": 44.4,
      "vue_framework_pct": 100
    },
    "source": "Retained numeric coder-gate summary, 2026-07-30; reviewed October 10, 2026.",
    "notes": "Measured historical gate. Original framework 75.2%; tuned rematch 81.7%. Historical rejection remained.",
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ],
    "methodology": "Eight framework prompts scored by regex, twelve developer tool scenarios, and warm lighthouse-prose concurrency waves. Defaults: framework temperature 0.2 / output 4096; tools 0.1 / 1024; concurrency 0.7 / 512; thinking off. c1/c2/c4 divide total completion positions by whole-wave wall time. Mean per-request decode excludes each request’s TTFT. Gate first response is the short ping, not an independently measured model load time.",
    "limitations": "No code compilation or actual tool execution. One report does not establish repeat spread or context qualification. Different slot counts, runtime revisions and dates limit cross-engine comparisons."
  },
  {
    "id": "2026-07-30-swe-ornith",
    "date": "2026-07-30",
    "machineId": "lab-pool",
    "model": "Ornith-1.0-35B · Qwen3.6-35B · Laguna-XS-2.1 · Coder-Next",
    "engine": "dated bench page · idle · thinking off",
    "task": "Jul 30 SWE-bench leaderboard review (local gate)",
    "metrics": {
      "ornith_devcode_median_pct": 84.2,
      "qwen36_devcode_pct": 81.7,
      "laguna_xs_devcode_pct": 83.9,
      "coder_next_devcode_pct": 86.4,
      "ornith_c1_tok_s": 61.6,
      "coder_next_c1_tok_s": 50.9
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical. All three published-rank challengers rejected as primary coder. Ornith took the tools-specialist slot; coder-laguna retired. One gate run is not a result — Ornith five-run spread 73.2–87.3. Not live Oct 1 telemetry.",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-07-30-swebench-review",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ]
  },
  {
    "id": "2026-08-03-gate-deepseek-label-unverified",
    "date": "2026-08-03",
    "machineId": "lab-pool",
    "model": "DeepSeek-V4-Flash · quant identity unconfirmed",
    "engine": "Retained gate label ds4-iq2m; exact artifact not frozen",
    "task": "Archived coder gate · ds4-iq2m",
    "metrics": {
      "framework_pct": 84.8,
      "developer_tools_pct": 41.7,
      "c1_tok_s": 12.4,
      "c2_tok_s": 12.7,
      "c4_tok_s": 16.5,
      "four_request_mean_decode_tok_s": 17,
      "framework_wall_s": 371,
      "gate_first_response_s": 1,
      "bash_framework_pct": 88.9,
      "nuxt_framework_pct": 85.7,
      "repo_framework_pct": 100,
      "swift_framework_pct": 68.8,
      "tailwind_framework_pct": 66.7,
      "vue_framework_pct": 100
    },
    "source": "Retained numeric coder-gate summary, 2026-08-03; reviewed October 10, 2026.",
    "notes": "Measured historical gate. Measured report retained under its historical label. The label alone does not establish the quant. Do not transfer these quality scores to the verified August 2 UD-IQ2_XXS throughput-only artifact.",
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ],
    "methodology": "Eight framework prompts scored by regex, twelve developer tool scenarios, and warm lighthouse-prose concurrency waves. Defaults: framework temperature 0.2 / output 4096; tools 0.1 / 1024; concurrency 0.7 / 512; thinking off. c1/c2/c4 divide total completion positions by whole-wave wall time. Mean per-request decode excludes each request’s TTFT. Gate first response is the short ping, not an independently measured model load time.",
    "limitations": "Quant identity is unresolved. Score attribution stays with the retained report label. No code compilation, tool execution or full-context qualification in this gate; one report does not establish repeat spread."
  },
  {
    "id": "2026-08-10-muse-glimmer-reject",
    "date": "2026-08-10",
    "machineId": "lab-pool",
    "model": "Meta Muse Glimmer 30B vs Coder-Next / Gemma agent",
    "engine": "llama.cpp side image · dated bench page",
    "task": "Aug 10 Muse Glimmer dual-role reject",
    "metrics": {
      "muse_c1_tok_s": 10.1,
      "champion_c1_tok_s": 50.9,
      "gemma_c1_tok_s": 60.4,
      "muse_devcode_pct": 81.5,
      "muse_devtool_pct": 25,
      "muse_toolbench_pct": 91.2
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical. Rejected for both fleet roles. Tools and sidecar quality were real; decode 0.20× the coder champion and 0.17× Gemma. Native context 131K fails the verified-256K coder gate. Not added to llama-swap. Not live Oct 1 telemetry.",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-08-10-muse-glimmer-reject",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "developer-perception",
      "vision",
      "concurrency"
    ]
  },
  {
    "id": "2026-08-11-gate-lightning",
    "date": "2026-08-11",
    "machineId": "lab-pool",
    "model": "Nemotron-3.5-Lightning 30B-A3B · NVFP4 + DSpark",
    "engine": "vLLM · August gate revision not frozen",
    "task": "Archived coder gate · lightning-gate",
    "metrics": {
      "framework_pct": 70,
      "developer_tools_pct": 66.7,
      "c1_tok_s": 82.3,
      "c2_tok_s": 108.5,
      "c4_tok_s": 161.5,
      "four_request_mean_decode_tok_s": 45.7,
      "framework_wall_s": 44,
      "gate_first_response_s": 1,
      "bash_framework_pct": 66.7,
      "nuxt_framework_pct": 100,
      "repo_framework_pct": 50,
      "swift_framework_pct": 60.7,
      "tailwind_framework_pct": 55.6,
      "vue_framework_pct": 83.3
    },
    "source": "Retained numeric coder-gate summary, 2026-08-11; reviewed October 10, 2026.",
    "notes": "Measured historical gate. Earlier isolated gate. September rebaseline is a separate dated measurement, not a replacement for this result.",
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ],
    "methodology": "Eight framework prompts scored by regex, twelve developer tool scenarios, and warm lighthouse-prose concurrency waves. Defaults: framework temperature 0.2 / output 4096; tools 0.1 / 1024; concurrency 0.7 / 512; thinking off. c1/c2/c4 divide total completion positions by whole-wave wall time. Mean per-request decode excludes each request’s TTFT. Gate first response is the short ping, not an independently measured model load time.",
    "limitations": "No code compilation or actual tool execution. One report does not establish repeat spread or context qualification. Different slot counts, runtime revisions and dates limit cross-engine comparisons."
  },
  {
    "id": "2026-08-11-lightning-coder-gate",
    "date": "2026-08-11",
    "machineId": "lab-pool",
    "model": "Nemotron-3.5-Lightning 30B-A3B NVFP4 + DSpark vs Coder-Next",
    "engine": "vLLM 0.27.1 · dated bench page",
    "task": "Aug 11 Lightning coder gate",
    "metrics": {
      "lightning_devcode_pct": 70,
      "champion_devcode_pct": 86.4,
      "lightning_c1_tok_s": 82.3,
      "champion_c1_tok_s": 50.9,
      "lightning_c4_agg_tok_s": 161.5,
      "champion_c4_agg_tok_s": 76
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical. Speed win (1.62× c1) and concurrency scales; quality −16.4 outside ±4 noise. Rejected as primary coder; kept as on-demand lightning. Not live Oct 1 telemetry.",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-08-11-lightning-gate",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ]
  },
  {
    "id": "2026-08-14-gate-qwen38-fp8",
    "date": "2026-08-14",
    "machineId": "lab-pool",
    "model": "Qwen3.8-27B · official FP8",
    "engine": "vLLM 0.27.1",
    "task": "Archived dual-role gate · Qwen3.8 FP8",
    "metrics": {
      "framework_pct": 87.5,
      "developer_tools_pct": 50,
      "broad_tools_thinking_off_pct": 100,
      "broad_tools_thinking_on_pct": 100,
      "developer_perception_facts": 30,
      "developer_perception_total": 30,
      "vision_pct": 62.5,
      "vision_tools_pct": 50,
      "prose_constraints_pct": 77.8,
      "c1_tok_s": 8,
      "c2_tok_s": 16,
      "c4_tok_s": 31.2,
      "four_request_mean_decode_tok_s": 7.8,
      "framework_wall_s": 1072,
      "cold_ready_s": 433,
      "allocated_gib": 91.8
    },
    "source": "Retained August 14 dual-role numeric summary and archived model registry; reviewed October 10, 2026.",
    "notes": "Measured historical gate; rejected for latency and readiness despite strong broad tools and developer screenshot facts. Configured 262144 context was not needle-verified.",
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "broad-tools",
      "developer-perception",
      "vision",
      "creative-constraints",
      "concurrency"
    ],
    "methodology": "Custom framework/tool/vision/writing suites plus warm prose concurrency waves. Broad tools were separately checked with thinking off and on. Context configuration and measured cold readiness are separate fields.",
    "limitations": "No executable framework validation. No full-window retrieval test. Different protocol axes cannot be combined into one universal model score."
  },
  {
    "id": "2026-08-19-gate-qwen38-dflash",
    "date": "2026-08-19",
    "machineId": "lab-pool",
    "model": "Qwen3.8-27B · NVFP4 + DFlash2",
    "engine": "SGLang · retained DFlash gate variant",
    "task": "Archived coder gate · qwen38-dflash-mf082",
    "metrics": {
      "framework_pct": 92,
      "developer_tools_pct": 50,
      "c1_tok_s": 22.9,
      "c2_tok_s": 42.8,
      "c4_tok_s": 43.3,
      "four_request_mean_decode_tok_s": 21.2,
      "framework_wall_s": 106,
      "gate_first_response_s": 2,
      "bash_framework_pct": 100,
      "nuxt_framework_pct": 100,
      "repo_framework_pct": 100,
      "swift_framework_pct": 87.5,
      "tailwind_framework_pct": 77.8,
      "vue_framework_pct": 91.7
    },
    "source": "Retained numeric coder-gate summary, 2026-08-19; reviewed October 10, 2026.",
    "notes": "Measured historical gate. Single August gate; 92% is not a repeat median or a new universal champion score.",
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ],
    "methodology": "Eight framework prompts scored by regex, twelve developer tool scenarios, and warm lighthouse-prose concurrency waves. Defaults: framework temperature 0.2 / output 4096; tools 0.1 / 1024; concurrency 0.7 / 512; thinking off. c1/c2/c4 divide total completion positions by whole-wave wall time. Mean per-request decode excludes each request’s TTFT. Gate first response is the short ping, not an independently measured model load time.",
    "limitations": "No code compilation or actual tool execution. One report does not establish repeat spread or context qualification. Different slot counts, runtime revisions and dates limit cross-engine comparisons."
  },
  {
    "id": "2026-09-12-catalog-agent-lightning",
    "date": "2026-09-12",
    "machineId": "lab-pool",
    "model": "agent / lightning · Nemotron-3.5-Lightning 30B-A3B",
    "engine": "UI catalog copy",
    "task": "roster catalog · resident agent",
    "metrics": {
      "catalog_tok_s_low": 80,
      "catalog_tok_s_high": 82
    },
    "source": "Coordinator roster + benches catalog text, page last-modified 2026-09-12",
    "notes": "UI catalog copy only. Coordinator roster still prints ~80–82 tok/s for resident agent / lightning. Oct 1 live snapshot saw this selector unloaded. Do not read these figures as live tok/s.",
    "kind": "catalog",
    "pageLastModified": "2026-09-12"
  },
  {
    "id": "2026-09-12-coder-ops",
    "date": "2026-07-26",
    "machineId": "lab-pool",
    "model": "coder ops note (post Coder-Next revert)",
    "engine": "dated bench page",
    "task": "coder request log (ops summary)",
    "metrics": {
      "requests": 507,
      "errors": 0,
      "avg_tok_s": 59.3,
      "longest_turn_s": 123
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical coder ops note from the July 26 revert write-up on the benches page: 507 requests, 0 errors, 59.3 tok/s average, longest turn 123 s. Not a live counter.",
    "kind": "ops",
    "pageLastModified": "2026-09-12"
  },
  {
    "id": "2026-09-12-concurrent-streams",
    "date": "2026-09-12",
    "machineId": "lab-pool",
    "model": "Lightning 30B-A3B · Qwen3.8-27B + DFlash2 · Mistral Small 4",
    "engine": "vLLM 0.29 / SGLang · dated bench page",
    "task": "Sep 12 four concurrent streams (c4 aggregate)",
    "metrics": {
      "lightning_c4_tok_s": 169,
      "qwen38_27b_c4_tok_s": 42.7,
      "mistral_c4_tok_s": 75.9
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical Sep 12 table. Lightning 2.2× Mistral at c4. Four concurrent streams. Not re-measured on 2026-10-01.",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-09-12-production-rebaseline",
    "suiteIds": [
      "concurrency"
    ]
  },
  {
    "id": "2026-09-12-gate-lightning",
    "date": "2026-09-12",
    "machineId": "lab-pool",
    "model": "Nemotron-3.5-Lightning 30B-A3B · NVFP4 + DSpark",
    "engine": "vLLM 0.29.0",
    "task": "Archived coder gate · baseline-lightning-v029",
    "metrics": {
      "framework_pct": 80.3,
      "developer_tools_pct": 75,
      "c1_tok_s": 77.6,
      "c2_tok_s": 120.4,
      "c4_tok_s": 169,
      "four_request_mean_decode_tok_s": 45.8,
      "framework_wall_s": 42,
      "gate_first_response_s": 0,
      "bash_framework_pct": 88.9,
      "nuxt_framework_pct": 85.7,
      "repo_framework_pct": 75,
      "swift_framework_pct": 74.1,
      "tailwind_framework_pct": 44.4,
      "vue_framework_pct": 100
    },
    "source": "Retained numeric coder-gate summary, 2026-09-12; reviewed October 10, 2026.",
    "notes": "Measured historical gate. Clean September rebaseline. Historical agent, superseded later.",
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ],
    "methodology": "Eight framework prompts scored by regex, twelve developer tool scenarios, and warm lighthouse-prose concurrency waves. Defaults: framework temperature 0.2 / output 4096; tools 0.1 / 1024; concurrency 0.7 / 512; thinking off. c1/c2/c4 divide total completion positions by whole-wave wall time. Mean per-request decode excludes each request’s TTFT. Gate first response is the short ping, not an independently measured model load time.",
    "limitations": "No code compilation or actual tool execution. One report does not establish repeat spread or context qualification. Different slot counts, runtime revisions and dates limit cross-engine comparisons."
  },
  {
    "id": "2026-09-12-gate-qwen38-dflash",
    "date": "2026-09-12",
    "machineId": "lab-pool",
    "model": "Qwen3.8-27B · NVFP4 + DFlash2",
    "engine": "SGLang Qwen3.8 image · retained gate revision",
    "task": "Archived coder gate · baseline-qwen38-sglang-v255",
    "metrics": {
      "framework_pct": 85.9,
      "developer_tools_pct": 75,
      "c1_tok_s": 19.6,
      "c2_tok_s": 31.5,
      "c4_tok_s": 42.7,
      "four_request_mean_decode_tok_s": 22.6,
      "framework_wall_s": 154,
      "gate_first_response_s": 1,
      "bash_framework_pct": 77.8,
      "nuxt_framework_pct": 100,
      "repo_framework_pct": 100,
      "swift_framework_pct": 74.1,
      "tailwind_framework_pct": 77.8,
      "vue_framework_pct": 91.7
    },
    "source": "Retained numeric coder-gate summary, 2026-09-12; reviewed October 10, 2026.",
    "notes": "Measured historical gate. Clean September rebaseline. Historical coder, superseded later.",
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "concurrency"
    ],
    "methodology": "Eight framework prompts scored by regex, twelve developer tool scenarios, and warm lighthouse-prose concurrency waves. Defaults: framework temperature 0.2 / output 4096; tools 0.1 / 1024; concurrency 0.7 / 512; thinking off. c1/c2/c4 divide total completion positions by whole-wave wall time. Mean per-request decode excludes each request’s TTFT. Gate first response is the short ping, not an independently measured model load time.",
    "limitations": "No code compilation or actual tool execution. One report does not establish repeat spread or context qualification. Different slot counts, runtime revisions and dates limit cross-engine comparisons."
  },
  {
    "id": "2026-09-12-generation-champions",
    "date": "2026-09-12",
    "machineId": "lab-pool",
    "model": "Lightning 30B-A3B · Qwen3.8-27B + DFlash2 · Mistral Small 4",
    "engine": "vLLM 0.29 / SGLang · dated bench page",
    "task": "Sep 12 generic generation (c1 decode)",
    "metrics": {
      "lightning_c1_tok_s": 77.6,
      "qwen38_27b_c1_tok_s": 19.6,
      "mistral_c1_tok_s": 30.7
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical Sep 12 2026 production rebaseline. Lightning 2.5× Mistral at c1. Not live telemetry. Live generation on 2026-10-01 was idle (0 tok/s).",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-09-12-production-rebaseline",
    "suiteIds": [
      "concurrency"
    ]
  },
  {
    "id": "2026-09-12-longform-pass",
    "date": "2026-09-12",
    "machineId": "lab-pool",
    "model": "Lightning 30B-A3B · long-form pass",
    "engine": "dated bench page",
    "task": "Sep 12 Lightning long-form generation",
    "metrics": {
      "tok_s": 85.9,
      "ttft_s": 0.13,
      "essay_words": 3935
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical long-form pass cited on the Sep 12 2026 rebaseline. A requested 10K-word essay stopped at 3,935 words. Not live telemetry.",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-09-12-production-rebaseline",
    "suiteIds": [
      "longform",
      "concurrency"
    ]
  },
  {
    "id": "2026-09-12-mistral-quality",
    "date": "2026-09-12",
    "machineId": "lab-pool",
    "model": "Mistral Small 4 119B-A6B NVFP4",
    "engine": "vLLM 0.29 · one Spark · 256K",
    "task": "Sep 12 Mistral quality watch (unrouted)",
    "metrics": {
      "broad_tools_pct": 100,
      "prose_constraints_pct": 86.9,
      "devcode_pct": 81.4,
      "devtool_pct": 83.3,
      "cold_ready_s": 674,
      "cached_ready_s": 585
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical fast gate. Perfect broad tools and stronger prose than Lightning (63.8%). Fits TP1 at 256K (66.12 GiB weights + 27.33 GiB KV) but ten-minute reload keeps it an unrouted specialist. Retrieval was not needle-tested. Not live telemetry.",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-09-12-production-rebaseline",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "broad-tools",
      "creative-constraints",
      "concurrency"
    ]
  },
  {
    "id": "2026-09-12-role-split-quality",
    "date": "2026-09-12",
    "machineId": "lab-pool",
    "model": "Lightning 30B-A3B · Qwen3.8-27B + DFlash2 · Mistral Small 4",
    "engine": "vLLM 0.29 / SGLang · non-reasoning · dated bench page",
    "task": "Sep 12 role-split quality gate",
    "metrics": {
      "lightning_devcode_pct": 80.3,
      "qwen_devcode_pct": 85.9,
      "mistral_devcode_pct": 81.4,
      "lightning_broad_tools_pct": 82.4,
      "qwen_broad_tools_pct": 94.1,
      "mistral_broad_tools_pct": 100
    },
    "source": "Internal benches page HTML, last-modified 2026-09-12",
    "notes": "Historical. Keep the split: Lightning owns latency, Qwen owns code (perfect multi-turn and restraint), Mistral stays an unrouted quality watch. Not live Oct 1 telemetry.",
    "kind": "historical",
    "pageLastModified": "2026-09-12",
    "journalSlug": "2026-09-12-production-rebaseline",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "broad-tools",
      "creative-constraints"
    ]
  },
  {
    "id": "2026-09-23-studio-gemma",
    "date": "2026-09-23",
    "machineId": "mac-studio",
    "model": "Gemma-4 26B-A4B QAT · Q4_K_M",
    "engine": "llama.cpp Metal · build 10964",
    "task": "Mac Studio · writing-specialist evaluation",
    "metrics": {
      "code_decode_tok_s": 78.1,
      "prose_decode_tok_s": 76.2,
      "framework_pct": 84.3,
      "broad_tools_pct": 94.1,
      "developer_tools_pct": 66.7,
      "prose_constraints_pct": 97
    },
    "source": "Mac Studio evaluation scorecard and retained numeric summary, 2026-09-22/23; reviewed directly 2026-10-02.",
    "notes": "Measured on M5 Max, 40 GPU cores, 48 GB unified memory. Single-stream decode is the median of three short 512-output probes on a loaded backend, with warm/fresh prefixes retained. Framework scores are regex criteria, not compilation. Broad/developer tools and prose constraints are custom suites. HumanEval was not run for this writing role; no coding benchmark score is inferred. Synthetic retrieval checks reached about 34K input positions, not 64K/262K qualification.",
    "kind": "historical",
    "pageLastModified": "2026-09-23",
    "journalSlug": "2026-09-23-studio-evaluation",
    "methodology": "Sequential large-model tests; thinking off; repaired proxy path. Quality suites and timing probes use different prompts.",
    "limitations": "Old HumanEval can be contaminated by training data. Custom suite percentages are not autonomous coding success rates; regex criteria can penalize valid alternatives.",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "broad-tools",
      "creative-constraints"
    ]
  },
  {
    "id": "2026-09-23-studio-lightning",
    "date": "2026-09-23",
    "machineId": "mac-studio",
    "model": "Nemotron-3.5-Lightning 30B-A3B · MLX 4-bit",
    "engine": "MLX / MLX-VLM",
    "task": "Mac Studio · fast-agent evaluation",
    "metrics": {
      "code_decode_tok_s": 150.1,
      "prose_decode_tok_s": 150.4,
      "framework_pct": 71.5,
      "broad_tools_pct": 88.2,
      "developer_tools_pct": 66.7,
      "prose_constraints_pct": 73.5,
      "humaneval_passed": 140,
      "humaneval_tasks": 164,
      "humaneval_pct": 85.4
    },
    "source": "Mac Studio evaluation scorecard and retained numeric summary, 2026-09-22/23; reviewed directly 2026-10-02.",
    "notes": "Measured on M5 Max, 40 GPU cores, 48 GB unified memory. Single-stream decode is the median of three short 512-output probes on a loaded backend, with warm/fresh prefixes retained. Framework scores are regex criteria, not compilation. Broad/developer tools and prose constraints are custom suites. HumanEval: all 164 original tasks; one greedy chat sample; 2048-output cap; thinking off; original sandboxed tests; no repair or retries. Chat-adapted pass@1, not a leaderboard reproduction. Synthetic retrieval checks reached about 34K input positions, not 64K/262K qualification.",
    "kind": "historical",
    "pageLastModified": "2026-09-23",
    "journalSlug": "2026-09-23-studio-evaluation",
    "methodology": "Sequential large-model tests; thinking off; repaired proxy path. Quality suites and timing probes use different prompts.",
    "limitations": "Old HumanEval can be contaminated by training data. Custom suite percentages are not autonomous coding success rates; regex criteria can penalize valid alternatives.",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "broad-tools",
      "creative-constraints",
      "human-eval"
    ]
  },
  {
    "id": "2026-09-23-studio-qwen38",
    "date": "2026-09-23",
    "machineId": "mac-studio",
    "model": "Qwen3.8-27B · MLX 4-bit",
    "engine": "MLX / MLX-VLM",
    "task": "Mac Studio · coding evaluation",
    "metrics": {
      "code_decode_tok_s": 33,
      "prose_decode_tok_s": 33,
      "framework_pct": 89.5,
      "broad_tools_pct": 100,
      "developer_tools_pct": 50,
      "prose_constraints_pct": 87.2,
      "humaneval_passed": 156,
      "humaneval_tasks": 164,
      "humaneval_pct": 95.1
    },
    "source": "Mac Studio evaluation scorecard and retained numeric summary, 2026-09-22/23; reviewed directly 2026-10-02.",
    "notes": "Measured on M5 Max, 40 GPU cores, 48 GB unified memory. Single-stream decode is the median of three short 512-output probes on a loaded backend, with warm/fresh prefixes retained. Framework scores are regex criteria, not compilation. Broad/developer tools and prose constraints are custom suites. HumanEval: all 164 original tasks; one greedy chat sample; 2048-output cap; thinking off; original sandboxed tests; no repair or retries. Chat-adapted pass@1, not a leaderboard reproduction. Synthetic retrieval checks reached about 34K input positions, not 64K/262K qualification.",
    "kind": "historical",
    "pageLastModified": "2026-09-23",
    "journalSlug": "2026-09-23-studio-evaluation",
    "methodology": "Sequential large-model tests; thinking off; repaired proxy path. Quality suites and timing probes use different prompts.",
    "limitations": "Old HumanEval can be contaminated by training data. Custom suite percentages are not autonomous coding success rates; regex criteria can penalize valid alternatives.",
    "suiteIds": [
      "framework-code",
      "developer-tools",
      "broad-tools",
      "creative-constraints",
      "human-eval"
    ]
  },
  {
    "id": "2026-09-26-aider-miaai-flash-partial",
    "date": "2026-09-26",
    "machineId": "lab-pool",
    "model": "MiaAI Qwen3.8-Flash-Next · NVFP4 (September artifact)",
    "engine": "vLLM two-GB10 · Aider whole-file editing",
    "task": "Aider polyglot · incomplete Flash-Next run",
    "metrics": {
      "within_two_tries_passed": 58,
      "tasks_recorded": 94,
      "suite_tasks": 225,
      "tasks_unrecorded": 131,
      "first_try_passed": 28,
      "test_timeouts": 0,
      "malformed_response_counter": 1,
      "context_exhaustion_counter": 72,
      "request_error_output_counter": 73,
      "cpp_tasks_recorded": 12,
      "cpp_first_try_passed": 1,
      "cpp_within_two_tries_passed": 5,
      "go_tasks_recorded": 11,
      "go_first_try_passed": 4,
      "go_within_two_tries_passed": 6,
      "java_tasks_recorded": 22,
      "java_first_try_passed": 6,
      "java_within_two_tries_passed": 15,
      "javascript_tasks_recorded": 24,
      "javascript_first_try_passed": 8,
      "javascript_within_two_tries_passed": 16,
      "python_tasks_recorded": 14,
      "python_first_try_passed": 4,
      "python_within_two_tries_passed": 10,
      "rust_tasks_recorded": 11,
      "rust_first_try_passed": 5,
      "rust_within_two_tries_passed": 6
    },
    "source": "Aggregated retained per-exercise Aider result files and local run recipe. Run start September 26; archive reviewed October 10, 2026.",
    "notes": "Measured partial archive: 94/225 task result files, 28 first-try passes and 58 passes within two tries among the recorded tasks. No final suite percentage or winner is reported. The missing 131 tasks are unrecorded, not imputed failures. This is the September MiaAI artifact, not the October NVIDIA serving artifact.",
    "kind": "historical",
    "completionStatus": "partial",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "aider-polyglot"
    ],
    "methodology": "225 Exercism tasks across six languages; whole-file editing; temperature 0; one worker; up to two tries with test feedback. Thinking on. Run recipe specifies 65536 context; metadata lists 262144 input and 32768 output. Per-language pass counts are preserved.",
    "limitations": "Effective chat-history cap is not independently frozen; earlier notes differ. Harness context-exhaustion, malformed-response and request-error-output counters are event counters, not counts of distinct failed tasks or HTTP requests. Incomplete run with different artifact, thinking policy and infrastructure; observed pass counts cannot support a final 225-task rate."
  },
  {
    "id": "2026-09-26-aider-qwen38-27b",
    "date": "2026-09-26",
    "machineId": "dgx-spark-a",
    "model": "Qwen3.8-27B · NVFP4 + DFlash2",
    "engine": "SGLang + DFlash2 · Aider whole-file editing",
    "task": "Aider polyglot · complete 225-task run",
    "metrics": {
      "within_two_tries_pct": 49.8,
      "first_try_pct": 9.3,
      "tasks_recorded": 225,
      "suite_tasks": 225,
      "first_try_passed": 21,
      "within_two_tries_passed": 112,
      "test_timeouts": 2,
      "malformed_response_counter": 0,
      "context_exhaustion_counter": 0,
      "request_error_output_counter": 0,
      "cpp_tasks_recorded": 26,
      "cpp_first_try_passed": 2,
      "cpp_within_two_tries_passed": 19,
      "go_tasks_recorded": 39,
      "go_first_try_passed": 5,
      "go_within_two_tries_passed": 19,
      "java_tasks_recorded": 47,
      "java_first_try_passed": 1,
      "java_within_two_tries_passed": 21,
      "javascript_tasks_recorded": 49,
      "javascript_first_try_passed": 3,
      "javascript_within_two_tries_passed": 26,
      "python_tasks_recorded": 34,
      "python_first_try_passed": 4,
      "python_within_two_tries_passed": 14,
      "rust_tasks_recorded": 30,
      "rust_first_try_passed": 6,
      "rust_within_two_tries_passed": 13
    },
    "source": "Aggregated retained per-exercise Aider result files and local run recipe. Run start September 26; archive reviewed October 10, 2026.",
    "notes": "Measured full recorded run: 21 first-try passes and 112 passes within two tries across all 225 exercises. EQ-Bench long-form generation for this arm was also recorded as completed with judging skipped; it provides no writing-quality score.",
    "kind": "historical",
    "completionStatus": "complete",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-benchmark-tests-and-models",
    "suiteIds": [
      "aider-polyglot"
    ],
    "methodology": "225 Exercism tasks across six languages; whole-file editing; temperature 0; one worker; up to two tries with test feedback. Thinking off. Run recipe specifies 65536 context; metadata lists 262144 input and 32768 output. Per-language pass counts are preserved.",
    "limitations": "Effective chat-history cap is not independently frozen; earlier notes differ. Harness context-exhaustion, malformed-response and request-error-output counters are event counters, not counts of distinct failed tasks or HTTP requests. One local run, without repeat spread or a matched completed Flash result."
  },
  {
    "id": "2026-10-05-glm-tensorfold-probe",
    "date": "2026-10-05",
    "machineId": "dgx-spark-a",
    "model": "GLM-5.3-Flash · EXL3",
    "engine": "TensorFold two-node TP2; exact probe revision not retained",
    "task": "October short-prompt runtime probe · GLM / TensorFold",
    "metrics": {
      "code_c1_tok_s": 97.59,
      "prose_c1_tok_s": 43.14,
      "prefill_ttft_s": 8.062,
      "code_c4_mean_tok_s": 78.9,
      "code_c8_mean_tok_s": 46.2,
      "prose_c4_mean_tok_s": 28.06,
      "prose_c8_mean_tok_s": 23.49,
      "code_c4_rate_sum_tok_s": 315.6,
      "code_c8_rate_sum_tok_s": 369.61,
      "prose_c4_rate_sum_tok_s": 112.25,
      "prose_c8_rate_sum_tok_s": 187.92,
      "code_c1_ttft_s": 0.181,
      "code_c1_completion_tokens": 207,
      "prefill_input_tokens": 6032,
      "wave_requests": 26,
      "wave_successes": 26,
      "prefill_requests": 1,
      "http_errors": 0
    },
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-two-spark-review",
    "source": "Retained lab speed report · October 5, 18:04:27–18:05:20 CDT",
    "notes": "Measured historical speed probe on the DGX Spark + ASUS GX10 pair. Every wave request and the separate prefill probe returned HTTP 200. Code was not executed and prose was not scored. A speed probe is not a model promotion or full-context qualification.",
    "methodology": "One prose prompt and one Python merge-function prompt; one wave each at concurrency 1, 4 and 8. Temperature 0, requested reasoning_effort medium, streaming with usage, wave output cap 320. TTFT includes first content or reasoning. Per-request decode = reported completion positions / elapsed time after first output. Concurrent means average these request rates; rate sums add them and are not wave-wall throughput. The separate ~6K input probe used output cap 16.",
    "limitations": "One wave per workload/concurrency, no repeat spread. Dates, model tokenizers and output lengths differ; exact runtime revision, effective thinking policy, cache state and external contention were not independently frozen. The prefill probe does not qualify 262K retrieval. No executed correctness, tool suite, or writing-quality result.",
    "suiteIds": [
      "october-speed"
    ]
  },
  {
    "id": "2026-10-05-qwen-vllm-probe",
    "date": "2026-10-05",
    "machineId": "dgx-spark-a",
    "model": "Qwen3.8-Flash-Next · NVFP4",
    "engine": "vLLM two-node TP2; exact probe revision not retained",
    "task": "October short-prompt runtime probe · Qwen / vLLM",
    "metrics": {
      "code_c1_tok_s": 67.25,
      "prose_c1_tok_s": 44.79,
      "prefill_ttft_s": 1.851,
      "code_c4_mean_tok_s": 53.48,
      "code_c8_mean_tok_s": 45.47,
      "prose_c4_mean_tok_s": 31.33,
      "prose_c8_mean_tok_s": 26.62,
      "code_c4_rate_sum_tok_s": 213.91,
      "code_c8_rate_sum_tok_s": 363.78,
      "prose_c4_rate_sum_tok_s": 125.32,
      "prose_c8_rate_sum_tok_s": 213,
      "code_c1_ttft_s": 0.317,
      "code_c1_completion_tokens": 244,
      "prefill_input_tokens": 6031,
      "wave_requests": 26,
      "wave_successes": 26,
      "prefill_requests": 1,
      "http_errors": 0
    },
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-two-spark-review",
    "source": "Retained lab speed report · October 5, 16:41:01–16:41:49 CDT",
    "notes": "Measured historical speed probe on the DGX Spark + ASUS GX10 pair. Every wave request and the separate prefill probe returned HTTP 200. Code was not executed and prose was not scored. A speed probe is not a model promotion or full-context qualification.",
    "methodology": "One prose prompt and one Python merge-function prompt; one wave each at concurrency 1, 4 and 8. Temperature 0, requested reasoning_effort medium, streaming with usage, wave output cap 320. TTFT includes first content or reasoning. Per-request decode = reported completion positions / elapsed time after first output. Concurrent means average these request rates; rate sums add them and are not wave-wall throughput. The separate ~6K input probe used output cap 16.",
    "limitations": "One wave per workload/concurrency, no repeat spread. Dates, model tokenizers and output lengths differ; exact runtime revision, effective thinking policy, cache state and external contention were not independently frozen. The prefill probe does not qualify 262K retrieval. No executed correctness, tool suite, or writing-quality result.",
    "suiteIds": [
      "october-speed"
    ]
  },
  {
    "id": "2026-10-09-qwen-tensorfold-probe",
    "date": "2026-10-09",
    "machineId": "dgx-spark-a",
    "model": "Qwen3.8-Flash-Next · NVFP4",
    "engine": "TensorFold two-node TP2; exact probe revision not retained",
    "task": "October short-prompt runtime probe · Qwen / TensorFold",
    "metrics": {
      "code_c1_tok_s": 80.09,
      "prose_c1_tok_s": 50.26,
      "prefill_ttft_s": 2.108,
      "code_c4_mean_tok_s": 53.7,
      "code_c8_mean_tok_s": 36.51,
      "prose_c4_mean_tok_s": 39.36,
      "prose_c8_mean_tok_s": 31.8,
      "code_c4_rate_sum_tok_s": 214.78,
      "code_c8_rate_sum_tok_s": 292.04,
      "prose_c4_rate_sum_tok_s": 157.46,
      "prose_c8_rate_sum_tok_s": 254.43,
      "code_c1_ttft_s": 0.119,
      "code_c1_completion_tokens": 228,
      "prefill_input_tokens": 6033,
      "wave_requests": 26,
      "wave_successes": 26,
      "prefill_requests": 1,
      "http_errors": 0
    },
    "kind": "historical",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-two-spark-review",
    "source": "Retained lab speed report · October 9, 22:56:32–22:57:14 CDT",
    "notes": "Measured historical speed probe on the DGX Spark + ASUS GX10 pair. Every wave request and the separate prefill probe returned HTTP 200. Code was not executed and prose was not scored. A speed probe is not a model promotion or full-context qualification.",
    "methodology": "One prose prompt and one Python merge-function prompt; one wave each at concurrency 1, 4 and 8. Temperature 0, requested reasoning_effort medium, streaming with usage, wave output cap 320. TTFT includes first content or reasoning. Per-request decode = reported completion positions / elapsed time after first output. Concurrent means average these request rates; rate sums add them and are not wave-wall throughput. The separate ~6K input probe used output cap 16.",
    "limitations": "One wave per workload/concurrency, no repeat spread. Dates, model tokenizers and output lengths differ; exact runtime revision, effective thinking policy, cache state and external contention were not independently frozen. The prefill probe does not qualify 262K retrieval. No executed correctness, tool suite, or writing-quality result.",
    "suiteIds": [
      "october-speed"
    ]
  },
  {
    "id": "2026-10-10-coder-operations",
    "date": "2026-10-10",
    "machineId": "dgx-spark-a",
    "model": "Qwen3.8-Flash-Next · NVFP4",
    "engine": "vLLM 0.30.0 · two-node TP2 + expert parallelism + MTP3",
    "task": "One hour of mixed coder traffic · October 10",
    "metrics": {
      "nonzero_sample_median_tok_s": 47.9,
      "peak_sample_tok_s": 103.4,
      "throughput_samples": 319,
      "nonzero_generation_samples": 310,
      "min_nonzero_sample_tok_s": 0.1,
      "max_sampled_running": 4,
      "max_sampled_waiting": 1,
      "peak_kv_cache_pct": 14.6,
      "first_prefix_hit_gauge_pct": 74.5,
      "last_prefix_hit_gauge_pct": 76.8,
      "head_logged_error_lines": 0,
      "head_logged_oom_lines": 0,
      "head_logged_preemption_lines": 0,
      "head_kv_cache_gib": 20.95,
      "startup_kv_tokens": 1014349,
      "startup_full_context_equivalent": 3.87,
      "api_log_to_ready_s": 869.18
    },
    "kind": "ops",
    "pageLastModified": "2026-10-10",
    "journalSlug": "2026-10-10-two-spark-review",
    "source": "Direct read-only head-container log summary · October 10, 20:00–21:00 CDT; separate October 10 startup log",
    "notes": "Dated operations observation, not a controlled decode benchmark or task-success study. Generation rates are engine-wide mixed-traffic samples. Startup cache capacity is the engine's calculation, not a full-context concurrency test.",
    "methodology": "Fixed interval 2026-10-11 01:00–02:00 UTC = October 10 20:00–21:00 America/Chicago. Parsed 319 head log rows; median over 310 rows with generation > 0. Peak running/waiting/cache values are sampled gauges. Prefix-hit values are counter-derived gauges, not an isolated interval hit rate. Counted ERROR/Traceback, OutOfMemory/out of memory and preempted lines in the same head log. Startup timing is first API argument log to Application startup complete on the same date.",
    "limitations": "No prompts or responses published. Irregular log intervals and mixed workloads; sampled maxima may miss peaks. Zero matching head log lines does not certify request correctness, worker or kernel health, or long-context reliability. The startup memory calculation and timing describe one boot. No current throughput test, retrieval qualification, or full-window soak was run."
  },
  {
    "id": "2026-10-22-glm-hermes-control",
    "date": "2026-10-22",
    "pageLastModified": "2026-10-23",
    "machineId": "dgx-spark-a",
    "model": "GLM-5.3-Flash · EXL3 · abliterated",
    "engine": "TensorFold TP2 · primary Hermes harness · frozen completion-v11 protocol",
    "task": "Matched local repository handoff panel · coding and review controls",
    "metrics": {
      "routine_passed": 1,
      "routine_tasks": 1,
      "hard_passed": 0,
      "hard_tasks": 3,
      "defective_review_correct": 0,
      "scope_admission_correct": 1,
      "healthy_review_correct": 1,
      "routine_wall_s": 63.487,
      "menu_behavior_wall_s": 376.562,
      "strict_types_wall_s": 190.034,
      "socket_repair_wall_s": 580.339
    },
    "kind": "historical",
    "suiteIds": [
      "local-repo-handoff"
    ],
    "completionStatus": "complete",
    "journalSlug": "2026-10-23-deepseek-two-gb10",
    "source": "Dated frozen primary Hermes comparison receipts · October 22–23, 2026; native identities and source fingerprints checked",
    "notes": "Measured local portfolio panel. Routine coding and hard coding have separate denominators; admission, defective-change review and healthy-change review are separate diagnostic controls. Both models tied on coding; no difficult-coding advantage was demonstrated.",
    "methodology": "Same frozen tasks, sources, sampling, budgets and independent acceptance gates. Temperature 0.2, thinking disabled, 4,096 output allowance per request. Routine: 360 seconds / 32 requests; hard visits: 1,440 seconds / 96 requests including repairs and reporting. Acceptance requires behavior, allowed scope, structured report, exact native identity and verified cleanup. Each control is one decision, encoded 1 for correct and 0 for incorrect. One visit per case; the healthy-menu positive control is reported separately from the six-case panel.",
    "limitations": "One primary harness, one routine coding task, three hard coding tasks, and no statistical ranking. This is not an official SWE-bench Verified result. Wall time includes tools and cache effects. Normal DeepSeek is compared with the previously serving abliterated GLM checkpoint. Later high-reasoning production tuning and historical mixed-revision controls are not pooled into this comparison. Public commands passing alone do not override behavior or scope failures."
  },
  {
    "id": "2026-10-23-deepseek-capacity",
    "date": "2026-10-23",
    "pageLastModified": "2026-10-23",
    "machineId": "dgx-spark-a",
    "model": "DeepSeek V4.1 Flash 1017 · EXL3 · normal",
    "engine": "TensorFold native · two-node TP2 · four shared slots",
    "task": "Short native capacity probe · 1, 2 and 4 concurrent requests",
    "metrics": {
      "c1_tok_s": 41.160227,
      "c1_output_tokens": 160,
      "c1_wave_wall_s": 3.887248,
      "c2_tok_s": 75.777256,
      "c2_output_tokens": 307,
      "c2_wave_wall_s": 4.051348,
      "c4_tok_s": 91.567748,
      "c4_output_tokens": 598,
      "c4_wave_wall_s": 6.530684,
      "prompt_tokens": 2787,
      "later_cached_tokens_per_request": 2048,
      "requests_completed": 7,
      "requests_attempted": 7
    },
    "kind": "historical",
    "suiteIds": [
      "local-serving-capacity"
    ],
    "journalSlug": "2026-10-23-deepseek-two-gb10",
    "source": "Dated native capacity receipt · October 23, 2026, 03:09 CDT",
    "notes": "Measured aggregate output divided by complete wave wall time: about 41.16, 75.78 and 91.57 tok/s at 1, 2 and 4 requests. Includes prefill; not pure decode speed or a quality score.",
    "methodology": "One sequential wave at each concurrency. Temperature zero, thinking disabled, output cap 160. Each request had 2,787 prompt positions. The first wave was uncached; all requests in the later waves reported 2,048 cached positions. Sum each wave’s completion positions and divide by the measured wave wall seconds. Record stop versus length termination and verify the exact native model identity for all seven requests.",
    "limitations": "Tiny synthetic workload, seven requests and no repeat spread. Warm/cache state differs between waves; several responses hit the output cap. Scaling is not a cold-cache concurrency comparison. No correctness grading, long-context retrieval or concurrent full-window soak. The earlier October harness’s sums of individual decode rates use a different denominator."
  },
  {
    "id": "2026-10-23-deepseek-hermes-control",
    "date": "2026-10-23",
    "pageLastModified": "2026-10-23",
    "machineId": "dgx-spark-a",
    "model": "DeepSeek V4.1 Flash 1017 · EXL3 · normal",
    "engine": "TensorFold TP2 · primary Hermes harness · frozen completion-v11 protocol",
    "task": "Matched local repository handoff panel · coding and review controls",
    "metrics": {
      "routine_passed": 1,
      "routine_tasks": 1,
      "hard_passed": 0,
      "hard_tasks": 3,
      "defective_review_correct": 1,
      "scope_admission_correct": 1,
      "healthy_review_correct": 1,
      "routine_wall_s": 74.698,
      "menu_behavior_wall_s": 350.188,
      "strict_types_wall_s": 412.824,
      "socket_repair_wall_s": 814.588
    },
    "kind": "historical",
    "suiteIds": [
      "local-repo-handoff"
    ],
    "completionStatus": "complete",
    "journalSlug": "2026-10-23-deepseek-two-gb10",
    "source": "Dated frozen primary Hermes comparison receipts · October 22–23, 2026; native identities and source fingerprints checked",
    "notes": "Measured local portfolio panel. Routine coding and hard coding have separate denominators; admission, defective-change review and healthy-change review are separate diagnostic controls. Both models tied on coding; no difficult-coding advantage was demonstrated.",
    "methodology": "Same frozen tasks, sources, sampling, budgets and independent acceptance gates. Temperature 0.2, thinking disabled, 4,096 output allowance per request. Routine: 360 seconds / 32 requests; hard visits: 1,440 seconds / 96 requests including repairs and reporting. Acceptance requires behavior, allowed scope, structured report, exact native identity and verified cleanup. Each control is one decision, encoded 1 for correct and 0 for incorrect. One visit per case; the healthy-menu positive control is reported separately from the six-case panel.",
    "limitations": "One primary harness, one routine coding task, three hard coding tasks, and no statistical ranking. This is not an official SWE-bench Verified result. Wall time includes tools and cache effects. Normal DeepSeek is compared with the previously serving abliterated GLM checkpoint. Later high-reasoning production tuning and historical mixed-revision controls are not pooled into this comparison. Public commands passing alone do not override behavior or scope failures."
  },
  {
    "id": "2026-10-23-deepseek-memory",
    "date": "2026-10-23",
    "pageLastModified": "2026-10-23",
    "machineId": "dgx-spark-a",
    "model": "DeepSeek V4.1 Flash 1017 · EXL3 · normal",
    "engine": "TensorFold native · two-node TP2",
    "task": "One-hour initial memory observation · both GB10 ranks",
    "metrics": {
      "head_min_available_gib": 4.92712,
      "worker_min_available_gib": 6.545525,
      "head_samples": 1800,
      "worker_samples": 1800,
      "runtime_guard_floor_gib": 2.5,
      "head_recorded_supervisor_restarts": 0
    },
    "kind": "ops",
    "journalSlug": "2026-10-23-deepseek-two-gb10",
    "source": "Dated bounded memory sampler and runtime receipts · October 23, 03:02–04:02 CDT",
    "notes": "Operations observation covering initial validation, cache replay, benchmark visits and early live work. Minimum sampled MemAvailable: head 4.93 GiB, worker 6.55 GiB. No recorded OOM or supervisor restart during this bounded observation.",
    "methodology": "1,800 samples on each host over roughly one hour: head 08:02:22–09:02:22 UTC and worker 08:02:24–09:02:23 UTC. Report minimum host MemAvailable separately, retaining unchanged recipe admission, reserve and runtime guards. Samplers stopped after this window.",
    "limitations": "Not a controlled capacity experiment, a leak diagnosis, days of uptime, or a heavy concurrent long-context soak. Sampled minima can miss brief peaks; head and worker minima occurred at different instants. Swap occupancy is not an active paging rate."
  },
  {
    "id": "2026-10-23-deepseek-prefix-replay",
    "date": "2026-10-23",
    "pageLastModified": "2026-10-23",
    "machineId": "dgx-spark-a",
    "model": "DeepSeek V4.1 Flash 1017 · EXL3 · normal",
    "engine": "TensorFold native · TP2 · high reasoning",
    "task": "Held-history compatibility replay · cold and warm requests",
    "metrics": {
      "prompt_tokens": 63164,
      "history_messages": 220,
      "complete_tool_pairs": 110,
      "cold_wall_s": 43.9,
      "warm_wall_s": 8.8,
      "warm_cached_tokens": 61440,
      "structured_responses_passed": 2,
      "requests_attempted": 2
    },
    "kind": "historical",
    "suiteIds": [
      "local-serving-capacity"
    ],
    "journalSlug": "2026-10-23-deepseek-two-gb10",
    "source": "Dated copied-history compatibility receipts · October 23, before the 03:57 CDT loop resume",
    "notes": "Measured compatibility/cache observation. Two high-reasoning requests used a copy of an existing 220-message history with 110 complete tool-call/result pairs. Both returned valid structured responses; tools were not executed.",
    "methodology": "Replay the same copied history twice against the new native model without changing the original held session or executing proposed tools. Record total prompt length, response structure, request wall time and second-request cache usage. The warm request reported 61,440 cached positions out of 63,164 input positions.",
    "limitations": "One copied conversation and two requests, with no repeat spread, task-success grading or retrieval test. These wall times are not TTFT measurements. Cache reuse does not qualify the configured 524,288-position window, other conversations, or simultaneous full-context requests. Private history text is not published."
  },
  {
    "id": "2026-catalog-agent-gemma",
    "date": "2026-10-01",
    "machineId": "lab-pool",
    "model": "agent-gemma · Gemma-4-26B-A4B QAT",
    "engine": "UI catalog copy",
    "task": "roster catalog · on-demand agent-gemma",
    "metrics": {
      "catalog_tok_s": 65
    },
    "source": "Coordinator roster catalog (text extract). Not the Oct 1 idle probe.",
    "notes": "UI catalog copy only. On-demand llama-swap id, ttl 600, 160k. Do not read as live tok/s. Dated / historical / not-live.",
    "kind": "catalog",
    "pageLastModified": "2026-09-12"
  },
  {
    "id": "2026-catalog-omni",
    "date": "2026-10-01",
    "machineId": "lab-pool",
    "model": "omni · Nemotron 3 Nano Omni 30B-A3B",
    "engine": "UI catalog copy",
    "task": "roster catalog · on-demand omni",
    "metrics": {
      "catalog_tok_s": 60
    },
    "source": "Coordinator roster catalog (text extract). Not the Oct 1 idle probe.",
    "notes": "UI catalog copy only. On-demand multimodal OCR/GUI lane, ttl 600, 256k. Do not read as live tok/s. Dated / historical / not-live.",
    "kind": "catalog",
    "pageLastModified": "2026-09-12"
  }
]
