{
  "reviewedOn": "2026-10-23",
  "registryArchivedOn": "2026-09-12",
  "scope": "Locally evaluated artifact/runtime combinations, including rejected and superseded candidates. Public leaderboard entries and upstream reproduction scripts are excluded.",
  "limitations": [
    "This is an evidence catalogue, not one ranking. Scores from different dates, protocols, quantizations and hardware are not interchangeable.",
    "Measured means at least one local test was recorded; it does not imply every suite, full context or production qualification passed.",
    "36 archived variants remain after consolidating duplicate original/rematch rows. Newer Studio, Aider and October evidence is added separately. Missing scores stay missing."
  ],
  "models": [
    {
      "id": "glm53-flash-exl3-abliterated-hermes",
      "name": "GLM-5.3-Flash · production control",
      "artifact": "EXL3 · abliterated serving checkpoint",
      "runtime": "TensorFold native · TP2 · primary Hermes handoff panel",
      "hardware": "DGX Spark + ASUS GX10 · two 128 GB GB10 hosts",
      "coverage": "measured",
      "evidenceWindow": "2026-10-22",
      "roles": [],
      "decision": "Previous coder; rollback preserved",
      "summary": "Matched Hermes panel: routine 1/1, hard 0/3 for both models. Correct admission and healthy-review decisions, but incorrect approval of the defective-change review. Do not pool with the earlier GLM speed artifact.",
      "metrics": {
        "routine_passed": 1,
        "routine_tasks": 1,
        "hard_passed": 0,
        "hard_tasks": 3,
        "defective_review_correct": 0
      },
      "context": "Frozen local fixture context only; not long-context qualification.",
      "suiteIds": [
        "local-repo-handoff"
      ],
      "recordIds": [
        "2026-10-22-glm-hermes-control"
      ],
      "source": "Dated exact-identity matched primary Hermes receipts · October 22–23, 2026; current DeepSeek settings independently inspected October 23.",
      "limitations": "One routine coding task and three hard coding tasks, one visit each, one harness. Not official SWE-bench Verified. Normal DeepSeek versus abliterated GLM. Different dates, tools, cache effects and production tuning prevent a general quality or speed ranking."
    },
    {
      "id": "deepseek-v41-flash-1017-tensorfold",
      "name": "DeepSeek V4.1 Flash 1017",
      "artifact": "EXL3 EAQ 2.9 bpw · normal · ABLIT=0",
      "runtime": "TensorFold native · TP2 · primary Hermes handoff panel",
      "hardware": "DGX Spark + ASUS GX10 · two 128 GB GB10 hosts",
      "coverage": "measured",
      "evidenceWindow": "2026-10-23",
      "roles": [
        "coder",
        "scout",
        "vision"
      ],
      "decision": "Active deployment; hard-coding gain unproven",
      "summary": "Matched Hermes panel: routine 1/1, hard 0/3 for both models. DeepSeek correctly rejected the defective-change review that GLM approved. Production high-reasoning tuning is separate from this thinking-off panel.",
      "metrics": {
        "routine_passed": 1,
        "routine_tasks": 1,
        "hard_passed": 0,
        "hard_tasks": 3,
        "defective_review_correct": 1
      },
      "context": "524,288 configured; local 21,875-position validation and separate 63,164-position history replay do not qualify full-window retrieval.",
      "suiteIds": [
        "local-repo-handoff",
        "local-serving-capacity"
      ],
      "recordIds": [
        "2026-10-23-deepseek-hermes-control",
        "2026-10-23-deepseek-capacity",
        "2026-10-23-deepseek-memory",
        "2026-10-23-deepseek-prefix-replay"
      ],
      "source": "Dated exact-identity matched primary Hermes receipts · October 22–23, 2026; current DeepSeek settings independently inspected October 23.",
      "limitations": "One routine coding task and three hard coding tasks, one visit each, one harness. Not official SWE-bench Verified. Normal DeepSeek versus abliterated GLM. Different dates, tools, cache effects and production tuning prevent a general quality or speed ranking."
    },
    {
      "id": "cydonia24-q6",
      "name": "Cydonia 24B v4.3",
      "artifact": "Q6_K",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical two-round study · archived Sep 12, 2026",
      "roles": [
        "creative"
      ],
      "decision": "Retired",
      "summary": "Strong prose in the original qualitative review, with dense decode around 10 tok/s.",
      "metrics": {
        "reported_decode_approx_tok_s": 10
      },
      "context": "suite context only",
      "suiteIds": [
        "longform"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "deepseek-aug3-label",
      "name": "DeepSeek-V4-Flash · unconfirmed quant",
      "artifact": "Retained report label ds4-iq2m",
      "runtime": "Runtime/artifact revision not independently frozen",
      "hardware": "Historical lab pool",
      "coverage": "measured",
      "evidenceWindow": "Aug 3, 2026",
      "roles": [],
      "decision": "Identity unresolved",
      "summary": "Keep the measured report without attributing its quality to the verified August 2 IQ2_XXS artifact.",
      "metrics": {
        "framework_pct": 84.8,
        "developer_tools_pct": 41.7,
        "c1_tok_s": 12.4,
        "c4_tok_s": 16.5
      },
      "context": "Not qualified by this gate.",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "concurrency"
      ],
      "recordIds": [
        "2026-08-03-gate-deepseek-label-unverified"
      ],
      "source": "Dated public study and retained numeric archive; reviewed October 10, 2026.",
      "limitations": "The retained label is not proof of quant identity. Scores belong to this unresolved report only."
    },
    {
      "id": "deepseek-v4-iq2xxs",
      "name": "DeepSeek-V4-Flash-0731",
      "artifact": "UD-IQ2_XXS GGUF",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Aug 2, 2026",
      "roles": [
        "code"
      ],
      "decision": "Rejected",
      "summary": "One-Spark load and 262K retrieval were demonstrated, but decode and prefill were too slow for the primary coder. Quality was not measured for this verified quant.",
      "metrics": {
        "decode_low_tok_s": 16.6,
        "decode_high_tok_s": 17.8,
        "prefill_tok_s": 530,
        "prefill_input_positions": 7000,
        "weight_gib": 85,
        "available_memory_floor_gib": 31.1,
        "cold_load_s": 376
      },
      "context": "262144 retrieval verified on one GB10 host; this quant has throughput evidence only.",
      "suiteIds": [
        "context-retrieval",
        "concurrency"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "The August 3 gate has an unresolved quant identity and is listed separately. Do not assign its quality score to this artifact."
    },
    {
      "id": "qwythos27-q4",
      "name": "Empero Qwythos-27B-v1",
      "artifact": "Q4_K_M",
      "runtime": "llama.cpp · revision not frozen in registry",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Jul 29, 2026",
      "roles": [
        "code",
        "terminal tools",
        "general agent"
      ],
      "decision": "Rejected",
      "summary": "Matched base A/B showed 84.6% versus 84.3% framework and 11.0 versus 10.8 tok/s; no useful role improvement.",
      "metrics": {
        "framework_pct": 84.6,
        "developer_tools_pct": 75,
        "broad_tools_pct": 100,
        "c1_tok_s": 11
      },
      "context": "262K configured for the gate; advertised 1M not deep-tested",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "broad-tools",
        "concurrency"
      ],
      "recordIds": [
        "2026-07-29-gate-qwythos"
      ],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "gemma4-heretic-q6",
      "name": "Gemma-4 26B-A4B Heretic / ultra-uncensored",
      "artifact": "Q6_K",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical two-round study · archived Sep 12, 2026",
      "roles": [
        "creative",
        "vision"
      ],
      "decision": "Superseded",
      "summary": "Stable, careful former writing candidate, later superseded by QAT/Q4 qualification. No numeric quality score is supplied here.",
      "metrics": {},
      "context": "suite context only",
      "suiteIds": [
        "longform"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "studio-gemma-qat-metal",
      "name": "Gemma-4 26B-A4B QAT",
      "artifact": "Q4_K_M",
      "runtime": "llama.cpp Metal · build 10964",
      "hardware": "Mac Studio · M5 Max · 48 GB",
      "coverage": "measured",
      "evidenceWindow": "Sep 22–23, 2026",
      "roles": [],
      "decision": "Writing role",
      "summary": "Highest mechanical prose constraints in the Studio comparison. HumanEval was not run for this writing role.",
      "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
      },
      "context": "Synthetic retrieval reached about 34K input positions; not 64K/262K qualified.",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "broad-tools",
        "creative-constraints"
      ],
      "recordIds": [
        "2026-09-23-studio-gemma"
      ],
      "source": "Mac Studio evaluation scorecard and retained numeric summary, 2026-09-22/23; reviewed directly 2026-10-02.",
      "limitations": "Old HumanEval can be contaminated by training data. Custom suite percentages are not autonomous coding success rates; regex criteria can penalize valid alternatives."
    },
    {
      "id": "gemma-qat-q4-gb10",
      "name": "Gemma-4 26B-A4B QAT uncensored",
      "artifact": "Q4_K_M",
      "runtime": "llama.cpp · GB10-era agent",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Jul 19, 2026",
      "roles": [
        "agent",
        "vision",
        "tools",
        "creative"
      ],
      "decision": "Former default",
      "summary": "Former multimodal agent; blind-writing mean 8.35/10 and a production context ceiling of 160K. The recorded edge test reached 119,593 input positions.",
      "metrics": {
        "reported_decode_tok_s": 60.4,
        "broad_tools_pct": 91.2,
        "vision_pct": 87.5,
        "blind_editor_mean_score": 8.35,
        "deep_edge_input_positions": 119593,
        "deep_edge_s": 78.3
      },
      "context": "Historical production ceiling 163840; retained edge probe at 119,593 input positions. No 262K qualification.",
      "suiteIds": [
        "broad-tools",
        "vision",
        "blind-editor",
        "context-retrieval",
        "concurrency"
      ],
      "recordIds": [
        "2026-07-19-agent-gemma"
      ],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "gemma4-clean-q6",
      "name": "Gemma-4 26B-A4B-it clean",
      "artifact": "Q6_K",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical two-round study · archived Sep 12, 2026",
      "roles": [
        "agent",
        "vision"
      ],
      "decision": "Rejected",
      "summary": "The retained long-form study records a benign essay refusal with a claimed platform limit. No numeric literary-quality score was retained.",
      "metrics": {},
      "context": "suite context only",
      "suiteIds": [
        "longform"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "glm45-air-q4",
      "name": "GLM-4.5-Air 106B-A12B",
      "artifact": "Q4",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical two-round study · archived Sep 12, 2026",
      "roles": [
        "general agent"
      ],
      "decision": "Rejected",
      "summary": "Stable serving but slower decode prevented a general-agent default role.",
      "metrics": {
        "reported_decode_low_tok_s": 19,
        "reported_decode_high_tok_s": 20
      },
      "context": "suite context only",
      "suiteIds": [
        "longform"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "glm53-flash-tensorfold",
      "name": "GLM-5.3-Flash",
      "artifact": "EXL3",
      "runtime": "TensorFold two-node TP2; exact probe revision not retained",
      "hardware": "DGX Spark + ASUS GX10 · two GB10 nodes",
      "coverage": "measured",
      "evidenceWindow": "2026-10-05",
      "roles": [],
      "decision": "Speed probe",
      "summary": "Historical October 5 speed probe: 97.59 tok/s code and 43.14 tok/s prose. The later abliterated production checkpoint has a separate matched Hermes entry; these speed results do not transfer to it.",
      "metrics": {
        "code_c1_tok_s": 97.59,
        "prose_c1_tok_s": 43.14,
        "prefill_ttft_s": 8.062,
        "code_four_request_mean_decode_tok_s": 78.9,
        "code_four_request_rate_sum_tok_s": 315.6
      },
      "context": "~6K input timing probe; configured 262144 is not full-window qualification.",
      "suiteIds": [
        "october-speed"
      ],
      "recordIds": [
        "2026-10-05-glm-tensorfold-probe"
      ],
      "source": "Retained lab speed report · October 5, 18:04:27–18:05:20 CDT",
      "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."
    },
    {
      "id": "goetia-gemma-q4",
      "name": "Goetia / Gemma-4 26B-A4B Heretic",
      "artifact": "Q4_K_M",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical gate · archived Sep 12, 2026",
      "roles": [
        "agent",
        "tools",
        "vision",
        "creative"
      ],
      "decision": "Rejected",
      "summary": "Near the former agent’s speed and vision accuracy but a lower blind-writing mean.",
      "metrics": {
        "reported_decode_tok_s": 59.8,
        "vision_pct": 87.5,
        "blind_editor_mean_score": 6.65
      },
      "context": "deep retrieval passed",
      "suiteIds": [
        "vision",
        "blind-editor",
        "context-retrieval"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "huihui-qwen330-q6",
      "name": "Huihui Qwen3-30B-A3B-2507 abliterated",
      "artifact": "Q6_K",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical two-round study · archived Sep 12, 2026",
      "roles": [
        "general agent"
      ],
      "decision": "Superseded",
      "summary": "Former general agent, later superseded by Gemma QAT qualification.",
      "metrics": {
        "reported_decode_approx_tok_s": 60
      },
      "context": "suite context only",
      "suiteIds": [
        "longform"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "huihui-qwen36-nvfp4",
      "name": "Huihui Qwen3.6-35B-A3B abliterated",
      "artifact": "NVFP4",
      "runtime": "Historical GB10 runtime",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical gate · archived Sep 12, 2026",
      "roles": [
        "general agent"
      ],
      "decision": "Rejected",
      "summary": "One of three long runs collapsed. This evidence applies to the tested quant, rather than every variant of this family.",
      "metrics": {
        "reported_decode_low_tok_s": 57,
        "reported_decode_high_tok_s": 60,
        "long_runs": 3,
        "collapsed_runs": 1
      },
      "context": "long-form suite",
      "suiteIds": [
        "longform"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "mage-flow",
      "name": "Mage-Flow Turbo 4B",
      "artifact": "1024×1024 local pipeline",
      "runtime": "Dated image-generation pipeline",
      "hardware": "Historical local image lane",
      "coverage": "measured",
      "evidenceWindow": "Jul 22, 2026",
      "roles": [],
      "decision": "Selected",
      "summary": "Historical preferred fit for warm latency, memory and visual quality.",
      "metrics": {
        "warm_generation_s": 3.96,
        "peak_gib": 17.03
      },
      "context": "Not a language-model context test.",
      "suiteIds": [
        "image-generation"
      ],
      "recordIds": [
        "2026-07-22-mage-flow"
      ],
      "source": "Dated public study and retained numeric archive; reviewed October 10, 2026.",
      "limitations": "Warm timings only; pipeline/model styles and quality goals differ."
    },
    {
      "id": "muse-glimmer-dynamic",
      "name": "Meta Muse Glimmer 30B",
      "artifact": "Dynamic kquant + projector + DFlash",
      "runtime": "llama.cpp b10358",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Aug 10, 2026",
      "roles": [
        "agent tools",
        "vision",
        "code",
        "creative"
      ],
      "decision": "Rejected",
      "summary": "Fast loading and complete screenshot-fact coverage, but 25% developer tools, dense-model latency and a configured 131072 ceiling failed the historical coder role.",
      "metrics": {
        "framework_pct": 81.5,
        "developer_tools_pct": 25,
        "developer_perception_facts": 30,
        "developer_perception_total": 30,
        "vision_pct": 75,
        "c1_tok_s": 10.1,
        "cold_ready_s": 17
      },
      "context": "131072 configured (Meta default); Unsloth claims up to 262K — not needle-verified here",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "developer-perception",
        "vision",
        "concurrency"
      ],
      "recordIds": [
        "2026-08-10-muse-glimmer-reject"
      ],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "miaai-flash-aider",
      "name": "MiaAI Qwen3.8-Flash-Next",
      "artifact": "September NVFP4 artifact · thinking on",
      "runtime": "vLLM · Aider whole-file editing",
      "hardware": "Two GB10 nodes",
      "coverage": "partial",
      "evidenceWindow": "Run started Sep 26, 2026",
      "roles": [],
      "decision": "Incomplete",
      "summary": "58 passes within two tries among 94 recorded exercises, out of a planned 225. No final pass rate or model winner.",
      "metrics": {
        "within_two_tries_passed": 58,
        "first_try_passed": 28,
        "tasks_recorded": 94,
        "suite_tasks": 225,
        "tasks_unrecorded": 131,
        "context_exhaustion_counter": 72,
        "request_error_output_counter": 73
      },
      "context": "Recipe specifies 65536; metadata lists 262144. Effective history cap not frozen.",
      "suiteIds": [
        "aider-polyglot"
      ],
      "recordIds": [
        "2026-09-26-aider-miaai-flash-partial"
      ],
      "source": "Dated public study and retained numeric archive; reviewed October 10, 2026.",
      "limitations": "Incomplete archive and different thinking policy. September MiaAI quality does not transfer to the current October NVIDIA artifact."
    },
    {
      "id": "mistral-small4-nvfp4",
      "name": "Mistral Small 4 119B-A6B",
      "artifact": "Official NVFP4",
      "runtime": "vLLM 0.29.0 · TP1",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Sep 12, 2026",
      "roles": [
        "agent",
        "tools",
        "code",
        "multimodal candidate"
      ],
      "decision": "Retained candidate",
      "summary": "Quality candidate left unrouted: stronger tools and prose constraints, but slower decode and long readiness.",
      "metrics": {
        "framework_pct": 81.4,
        "developer_tools_pct": 83.3,
        "broad_tools_pct": 100,
        "prose_constraints_pct": 86.9,
        "c1_tok_s": 30.7,
        "c2_tok_s": 52.2,
        "c4_tok_s": 75.9,
        "cold_ready_s": 674,
        "cached_ready_s": 585,
        "model_gib": 66.12,
        "kv_gib": 27.33
      },
      "context": "262144 configured; vLLM allocated 1,273,776 KV tokens / 4.86× capacity; retrieval not needle-verified",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "broad-tools",
        "creative-constraints",
        "concurrency"
      ],
      "recordIds": [
        "2026-09-12-mistral-quality"
      ],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "nano-omni-q4",
      "name": "Nemotron 3 Nano Omni 30B-A3B",
      "artifact": "UD-Q4_K_XL + projector",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical gate · archived Sep 12, 2026",
      "roles": [
        "developer vision",
        "OCR"
      ],
      "decision": "Former default",
      "summary": "Historical developer-vision sidecar: all 30 screenshot facts in 10 seconds. Generic fixtures were 5/8.",
      "metrics": {
        "developer_perception_facts": 30,
        "developer_perception_total": 30,
        "developer_perception_wall_s": 10,
        "vision_passed": 5,
        "vision_tasks": 8
      },
      "context": "240K verified; 262K configured",
      "suiteIds": [
        "developer-perception",
        "vision",
        "context-retrieval"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "nemotron-super-mxfp4",
      "name": "Nemotron 3 Super 120B-A12B",
      "artifact": "MXFP4",
      "runtime": "Historical GB10 runtime",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "partial",
      "evidenceWindow": "Historical partial gate · archived Sep 12, 2026",
      "roles": [
        "code"
      ],
      "decision": "Capacity rejected",
      "summary": "Answered three framework tasks, then was OOM-killed. Individual task scores are retained; no full-suite percentage is inferred.",
      "metrics": {
        "tasks_recorded": 3,
        "task_one_pct": 100,
        "task_two_pct": 100,
        "task_three_pct": 77.8,
        "reported_decode_tok_s": 16.6
      },
      "context": "262K configured; not qualified",
      "suiteIds": [
        "framework-code"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Three recorded tasks are not a completed eight-task gate. Configured context was not qualified."
    },
    {
      "id": "cascade2-q4",
      "name": "Nemotron Cascade 2 30B-A3B",
      "artifact": "Q4",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical gate · archived Sep 12, 2026",
      "roles": [
        "code",
        "long-context"
      ],
      "decision": "Specialist",
      "summary": "Context specialist with a recorded 240K edge pass in 141 seconds; did not replace the coder.",
      "metrics": {
        "framework_pct": 80.6,
        "developer_tools_pct": 66.7,
        "deep_edge_s": 141
      },
      "context": "240K verified; native up to 1M",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "context-retrieval"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "studio-lightning-mlx",
      "name": "Nemotron-3.5-Lightning 30B-A3B",
      "artifact": "MLX 4-bit",
      "runtime": "MLX / MLX-VLM",
      "hardware": "Mac Studio · M5 Max · 48 GB",
      "coverage": "measured",
      "evidenceWindow": "Sep 22–23, 2026",
      "roles": [],
      "decision": "Former default",
      "summary": "Fast September Studio agent artifact. The October agent alias serves a different model.",
      "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
      },
      "context": "Synthetic retrieval reached about 34K input positions; not 64K/262K qualified.",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "broad-tools",
        "creative-constraints",
        "human-eval"
      ],
      "recordIds": [
        "2026-09-23-studio-lightning"
      ],
      "source": "Mac Studio evaluation scorecard and retained numeric summary, 2026-09-22/23; reviewed directly 2026-10-02.",
      "limitations": "Old HumanEval can be contaminated by training data. Custom suite percentages are not autonomous coding success rates; regex criteria can penalize valid alternatives."
    },
    {
      "id": "lightning-nvfp4-dspark",
      "name": "Nemotron-3.5-Lightning 30B-A3B",
      "artifact": "Official NVFP4 + DSpark",
      "runtime": "vLLM 0.29.0",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Aug 11 / Sep 12, 2026",
      "roles": [
        "agent",
        "parallel work",
        "long context"
      ],
      "decision": "Former default",
      "summary": "September agent default. The clean rebaseline improved the framework gate to 80.3%; the long essay stopped at 3,935 of 10,000 requested words.",
      "metrics": {
        "framework_pct": 80.3,
        "developer_tools_pct": 75,
        "broad_tools_pct": 82.4,
        "prose_constraints_pct": 63.8,
        "c1_tok_s": 77.6,
        "c2_tok_s": 120.4,
        "c4_tok_s": 169,
        "longform_tok_s": 85.9,
        "longform_ttft_s": 0.13,
        "longform_words": 3935,
        "longform_requested_words": 10000
      },
      "context": "production 128K; 1M native, not re-needle-tested in this pass",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "broad-tools",
        "creative-constraints",
        "concurrency",
        "longform"
      ],
      "recordIds": [
        "2026-09-12-gate-lightning",
        "2026-08-11-gate-lightning"
      ],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "north-mini-q4",
      "name": "North Mini Code 30B-A3B",
      "artifact": "Q4",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical gate · archived Sep 12, 2026",
      "roles": [
        "code",
        "tools"
      ],
      "decision": "Rejected",
      "summary": "Thinking retest improved framework checks, but tools and deep-context/scaffold qualification still failed.",
      "metrics": {
        "framework_thinking_off_pct": 58.3,
        "framework_thinking_retest_pct": 74.4,
        "developer_tools_pct": 41.7
      },
      "context": "incomplete/failing at depth",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "context-retrieval"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "ornith9-q6",
      "name": "Ornith 1.0 9B uncensored",
      "artifact": "Q6_K",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical two-round study · archived Sep 12, 2026",
      "roles": [
        "creative"
      ],
      "decision": "Retired",
      "summary": "Long outputs at about 28 tok/s, without a current routed advantage.",
      "metrics": {
        "reported_decode_approx_tok_s": 28
      },
      "context": "suite context only",
      "suiteIds": [
        "longform"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "ornith35-q4",
      "name": "Ornith-1.0-35B",
      "artifact": "UD-Q4_K_XL",
      "runtime": "llama.cpp · tuned gate",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Jul 30, 2026",
      "roles": [
        "code",
        "tools"
      ],
      "decision": "Specialist",
      "summary": "Five framework runs: median 84.2%, standard deviation 5.55 points; Swift median 60.7%. Four tool runs: median 75%. Faster, smaller and quick to load, but not promoted over the primary coder.",
      "metrics": {
        "framework_median_pct": 84.2,
        "framework_stddev_points": 5.55,
        "swift_median_pct": 60.7,
        "developer_tools_median_pct": 75,
        "framework_runs": 5,
        "developer_tool_runs": 4,
        "c1_tok_s": 61.6,
        "cold_ready_s": 27,
        "allocated_gib": 28.2
      },
      "context": "2 x 262144 configured; native 262144",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "concurrency"
      ],
      "recordIds": [
        "2026-07-30-gate-ornith",
        "2026-07-30-swe-ornith"
      ],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "ortenzya31-q4",
      "name": "Ortenzya 31B",
      "artifact": "Q4_K_M",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical gate · archived Sep 12, 2026",
      "roles": [
        "creative specialist"
      ],
      "decision": "Specialist",
      "summary": "Optional creative specialist; blind-writing mean 8.17/10, with dense decode around 9.8 tok/s.",
      "metrics": {
        "blind_editor_mean_score": 8.17,
        "reported_decode_tok_s": 9.8
      },
      "context": "deep retrieval passed",
      "suiteIds": [
        "blind-editor",
        "context-retrieval"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "laguna-s-nvfp4",
      "name": "Poolside Laguna-S-2.1 117.6B MoE",
      "artifact": "NVFP4 · revision 0761412",
      "runtime": "vLLM · SM121",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Jul 23–26, 2026",
      "roles": [
        "code",
        "tools",
        "long-context"
      ],
      "decision": "Reversed promotion",
      "summary": "Promoted July 23 and reverted July 26. Same-day control won framework quality and wall time; Laguna retained a tool advantage but had heavy memory use and slow cold startup.",
      "metrics": {
        "framework_pct": 87.2,
        "developer_tools_pct": 75,
        "framework_wall_s": 436,
        "c1_tok_s": 17.7,
        "allocated_gib": 101,
        "cold_ready_low_s": 480,
        "cold_ready_high_s": 900
      },
      "context": "257.1K verified (middle + edge pair); 262K configured",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "context-retrieval",
        "concurrency"
      ],
      "recordIds": [
        "2026-07-26-coder-revert",
        "2026-07-21-laguna-s-reject"
      ],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "laguna-xs-nvfp4",
      "name": "Poolside Laguna-XS-2.1 33B",
      "artifact": "NVFP4",
      "runtime": "vLLM",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Jul 30, 2026",
      "roles": [
        "code",
        "tools"
      ],
      "decision": "Rejected",
      "summary": "Smaller weights did not solve the long-context KV footprint. Four-stream throughput used different slot settings than the llama.cpp incumbent; it is not an isolated model advantage.",
      "metrics": {
        "framework_pct": 83.9,
        "developer_tools_pct": 50,
        "c1_tok_s": 35.3,
        "c4_tok_s": 127.7,
        "cold_ready_s": 200
      },
      "context": "262144 configured",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "concurrency"
      ],
      "recordIds": [
        "2026-07-30-gate-laguna-xs"
      ],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "qwen235-q2",
      "name": "Qwen3-235B-A22B-2507",
      "artifact": "Q2_K_XL; separate Q3 load failed",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical study · archived Sep 12, 2026",
      "roles": [
        "general agent"
      ],
      "decision": "Rejected",
      "summary": "Q3 load failed with OOM. The tested Q2 ran around 14 tok/s with qualitative quantization damage; legacy refusal counts are separate from quality.",
      "metrics": {
        "reported_decode_approx_tok_s": 14,
        "legacy_refusal_cases": 6,
        "legacy_refusals": 1
      },
      "context": "suite context only",
      "suiteIds": [
        "longform"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "qwen330-clean-q6",
      "name": "Qwen3-30B-A3B-Instruct-2507 clean",
      "artifact": "Q6_K",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical two-round study · archived Sep 12, 2026",
      "roles": [
        "general agent"
      ],
      "decision": "Rejected",
      "summary": "Stable serving, but outline and long-form decay prevented a default role.",
      "metrics": {
        "reported_decode_approx_tok_s": 62
      },
      "context": "suite context only",
      "suiteIds": [
        "longform"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "coder-next-q5",
      "name": "Qwen3-Coder-Next 80B-A3B",
      "artifact": "Q5_K_M",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical gate · archived Sep 12, 2026",
      "roles": [
        "code",
        "Git",
        "Bash",
        "tools"
      ],
      "decision": "Rejected",
      "summary": "Framework score near the Q4 baseline, but Swift and Tailwind regressed and code decode was slower.",
      "metrics": {
        "framework_pct": 86.2,
        "developer_tools_pct": 66.7,
        "code_decode_tok_s": 45.3
      },
      "context": "same native 262K; deep test reserved for best Q5",
      "suiteIds": [
        "framework-code",
        "developer-tools"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "coder-next-q4",
      "name": "Qwen3-Coder-Next 80B-A3B",
      "artifact": "UD-Q4_K_XL",
      "runtime": "llama.cpp · dated runtime/tuning variants",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Jul 21–30, 2026",
      "roles": [
        "code",
        "Git",
        "Bash",
        "tools"
      ],
      "decision": "Former default",
      "summary": "Restored after the Laguna reversal. Six-run calibration median 86.4%; a separate July 26 run reached 90.9%. These are different summaries, not two contradictory scores.",
      "metrics": {
        "framework_median_pct": 86.4,
        "framework_stddev_points": 2.9,
        "jul26_framework_pct": 90.9,
        "jul26_developer_tools_pct": 66.7,
        "jul26_c1_tok_s": 52.7,
        "jul26_c4_tok_s": 76,
        "allocated_gib": 53.9,
        "cold_ready_approx_s": 90
      },
      "context": "2 x 262K configured on GX10; 2 x ~241K concurrently verified",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "context-retrieval",
        "concurrency"
      ],
      "recordIds": [
        "2026-07-27-gate-coder-next",
        "2026-07-26-coder-revert",
        "2026-07-21-coder-2x256k"
      ],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "coder-next-udq5m",
      "name": "Qwen3-Coder-Next 80B-A3B",
      "artifact": "UD-Q5_K_M",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical gate · archived Sep 12, 2026",
      "roles": [
        "code",
        "Git",
        "Bash",
        "tools"
      ],
      "decision": "Rejected",
      "summary": "No assigned-role win over the Q4 incumbent.",
      "metrics": {
        "framework_pct": 87.4,
        "developer_tools_pct": 58.3,
        "code_decode_tok_s": 41.6
      },
      "context": "same native 262K; deep test reserved for best Q5",
      "suiteIds": [
        "framework-code",
        "developer-tools"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "coder-next-udq5xl",
      "name": "Qwen3-Coder-Next 80B-A3B",
      "artifact": "UD-Q5_K_XL",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical gate · archived Sep 12, 2026",
      "roles": [
        "code",
        "Git",
        "Bash",
        "tools"
      ],
      "decision": "Watch",
      "summary": "88.0% overall and 74.1% Swift, with slower decode and Tailwind/Bash regressions. Retained as a Swift watch candidate.",
      "metrics": {
        "framework_pct": 88,
        "swift_pct": 74.1
      },
      "context": "231.7K + edge pair verified",
      "suiteIds": [
        "framework-code",
        "context-retrieval"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "qwen35-122-q4",
      "name": "Qwen3.5-122B-A10B",
      "artifact": "Q4 + projector",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical gate · archived Sep 12, 2026",
      "roles": [
        "code",
        "vision",
        "tools"
      ],
      "decision": "Rejected",
      "summary": "Good developer tools, but weaker framework quality and long deep-context wall time.",
      "metrics": {
        "framework_pct": 79.5,
        "developer_tools_pct": 83.3,
        "deep_edge_s": 480.6
      },
      "context": "231.7K verified; native 262K",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "context-retrieval"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "qwen35-27-base-q4",
      "name": "Qwen3.5-27B",
      "artifact": "Q4_K_M",
      "runtime": "llama.cpp",
      "hardware": "Historical lab pool",
      "coverage": "measured",
      "evidenceWindow": "Jul 29, 2026",
      "roles": [],
      "decision": "Control",
      "summary": "Matched Qwythos control. Framework 84.3%, tools 75%, broad tools 100% and single-wave throughput 10.8 tok/s.",
      "metrics": {
        "framework_pct": 84.3,
        "developer_tools_pct": 75,
        "broad_tools_pct": 100,
        "c1_tok_s": 10.8
      },
      "context": "262K configured; no deep qualification in this matched gate.",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "broad-tools",
        "concurrency"
      ],
      "recordIds": [
        "2026-07-29-gate-qwen35-base"
      ],
      "source": "Dated public study and retained numeric archive; reviewed October 10, 2026.",
      "limitations": "Applies to the tested artifact, runtime, hardware and dated protocol. It does not transfer to other variants."
    },
    {
      "id": "qwen35-9-q6-vision",
      "name": "Qwen3.5-9B",
      "artifact": "Q6_K + BF16 projector",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical gate · archived Sep 12, 2026",
      "roles": [
        "small vision fallback"
      ],
      "decision": "Fallback",
      "summary": "Complete screenshot-fact coverage, but the six images took 80.7 seconds versus Omni’s 10 seconds.",
      "metrics": {
        "developer_perception_facts": 30,
        "developer_perception_total": 30,
        "developer_perception_wall_s": 80.7,
        "vision_passed": 6,
        "vision_tasks": 8
      },
      "context": "231.7K verified; native 262K",
      "suiteIds": [
        "developer-perception",
        "vision",
        "context-retrieval"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "qwen36-27-q4",
      "name": "Qwen3.6-27B",
      "artifact": "Q4",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical gate · archived Sep 12, 2026",
      "roles": [
        "code",
        "vision"
      ],
      "decision": "Fallback",
      "summary": "A fallback with slower deep retrieval, rather than a primary upgrade.",
      "metrics": {
        "framework_pct": 81.8,
        "developer_tools_pct": 75,
        "vision_pct": 75,
        "deep_edge_s": 535
      },
      "context": "231.7K verified; native 262K",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "vision",
        "context-retrieval"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "fable-fusion27-q4",
      "name": "Qwen3.6-27B Fable Fusion",
      "artifact": "Q4_K_M",
      "runtime": "llama.cpp",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Historical gate · archived Sep 12, 2026",
      "roles": [
        "creative",
        "tools",
        "vision"
      ],
      "decision": "Rejected",
      "summary": "Long-form repetition collapse. Vision reached 75% only with tuned settings.",
      "metrics": {
        "blind_editor_mean_score": 4.77,
        "reported_decode_tok_s": 10.8,
        "tuned_vision_pct": 75
      },
      "context": "deep retrieval passed",
      "suiteIds": [
        "blind-editor",
        "longform",
        "vision",
        "context-retrieval"
      ],
      "recordIds": [],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "qwen36-35-q4",
      "name": "Qwen3.6-35B-A3B",
      "artifact": "UD-Q4_K_XL",
      "runtime": "llama.cpp · original and tuned rematch",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Jul 21 / Jul 30, 2026",
      "roles": [
        "code",
        "tools"
      ],
      "decision": "Rejected",
      "summary": "Original 75.2% framework rose to 81.7% in the tuned rematch, still below the historical primary baseline. Same artifact appears twice in the archive and is consolidated here.",
      "metrics": {
        "original_framework_pct": 75.2,
        "rematch_framework_pct": 81.7,
        "developer_tools_pct": 75,
        "vision_pct": 75,
        "c1_tok_s": 62.4,
        "c4_tok_s": 95.7
      },
      "context": "231.7K retrieval recorded in the original gate; two 262144 lanes configured for the later rematch, without a new two-lane qualification.",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "vision",
        "context-retrieval",
        "concurrency"
      ],
      "recordIds": [
        "2026-07-30-gate-qwen36"
      ],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "qwen36-heretic-q6-mtp",
      "name": "Qwen3.6-35B-A3B Heretic",
      "artifact": "Q6_K + native MTP",
      "runtime": "llama.cpp b9967 in July speed A/B",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Jul 11, 2026",
      "roles": [
        "agent",
        "code",
        "vision"
      ],
      "decision": "Superseded",
      "summary": "Six draft positions sped up code and slowed creative writing in the same quiet single-slot comparison.",
      "metrics": {
        "code_mtp_off_tok_s": 64.1,
        "code_mtp_on_tok_s": 100.9,
        "creative_mtp_off_tok_s": 64.6,
        "creative_mtp_on_tok_s": 53.6
      },
      "context": "~255K verified in candidate config",
      "suiteIds": [
        "concurrency",
        "context-retrieval"
      ],
      "recordIds": [
        "2026-07-11-mtp-workloads"
      ],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "studio-qwen38-mlx",
      "name": "Qwen3.8-27B",
      "artifact": "MLX 4-bit",
      "runtime": "MLX / MLX-VLM",
      "hardware": "Mac Studio · M5 Max · 48 GB",
      "coverage": "measured",
      "evidenceWindow": "Sep 22–23, 2026",
      "roles": [],
      "decision": "Coding candidate",
      "summary": "Higher HumanEval and framework checks than Lightning in the same Studio study.",
      "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
      },
      "context": "Synthetic retrieval reached about 34K input positions; not 64K/262K qualified.",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "broad-tools",
        "creative-constraints",
        "human-eval"
      ],
      "recordIds": [
        "2026-09-23-studio-qwen38"
      ],
      "source": "Mac Studio evaluation scorecard and retained numeric summary, 2026-09-22/23; reviewed directly 2026-10-02.",
      "limitations": "Old HumanEval can be contaminated by training data. Custom suite percentages are not autonomous coding success rates; regex criteria can penalize valid alternatives."
    },
    {
      "id": "qwen38-27b-nvfp4-dflash",
      "name": "Qwen3.8-27B",
      "artifact": "NVFP4 target + DFlash2 draft",
      "runtime": "SGLang · dated draft/runtime variants",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Aug 19 / Sep 12 / Sep 26, 2026",
      "roles": [
        "code"
      ],
      "decision": "Former default",
      "summary": "September coder default; isolated August gate 92.0%, September gate 85.9%. Completed Aider run: 112/225 passed within two tries. Draft/runtime and workloads vary across these dates.",
      "metrics": {
        "framework_pct": 85.9,
        "developer_tools_pct": 75,
        "broad_tools_pct": 94.1,
        "c1_tok_s": 19.6,
        "c2_tok_s": 31.5,
        "c4_tok_s": 42.7,
        "aider_first_try_passed": 21,
        "aider_within_two_tries_passed": 112,
        "aider_tasks_recorded": 225,
        "aider_suite_tasks": 225,
        "aider_within_two_tries_pct": 49.8
      },
      "context": "262144 configured/native; not needle-verified in this engine pass",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "broad-tools",
        "concurrency",
        "aider-polyglot"
      ],
      "recordIds": [
        "2026-09-12-gate-qwen38-dflash",
        "2026-08-19-gate-qwen38-dflash",
        "2026-09-26-aider-qwen38-27b"
      ],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "qwen38-27b-fp8",
      "name": "Qwen3.8-27B",
      "artifact": "Official fine-grained FP8",
      "runtime": "vLLM 0.27.1",
      "hardware": "Historical GB10 / lab pool; use dated record for exact host",
      "coverage": "measured",
      "evidenceWindow": "Aug 14, 2026",
      "roles": [
        "code",
        "agent",
        "native vision"
      ],
      "decision": "Rejected",
      "summary": "Strong broad tool and screenshot checks did not offset 8 tok/s decode, a 1,072-second framework run and 433-second cold readiness.",
      "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,
        "prose_constraints_pct": 77.8,
        "c1_tok_s": 8,
        "c4_tok_s": 31.2,
        "framework_wall_s": 1072,
        "cold_ready_s": 433,
        "allocated_gib": 91.8
      },
      "context": "262144 configured; advertised native — not needle-verified (speed already failed)",
      "suiteIds": [
        "framework-code",
        "developer-tools",
        "broad-tools",
        "developer-perception",
        "vision",
        "creative-constraints",
        "concurrency"
      ],
      "recordIds": [
        "2026-08-14-gate-qwen38-fp8"
      ],
      "source": "Archived local variant registry, last updated September 12, 2026; retained summaries reviewed October 10. Exact dates are shown only where retained.",
      "limitations": "Historical protocols and revisions vary. Registry-only values retain their original rounding; repeat spread, timing denominator or blind-judge task count may be unspecified. See each linked study for stronger provenance."
    },
    {
      "id": "qwen38-flash-tensorfold",
      "name": "Qwen3.8-Flash-Next",
      "artifact": "NVFP4 · October artifact",
      "runtime": "TensorFold two-node TP2; exact probe revision not retained",
      "hardware": "DGX Spark + ASUS GX10 · two GB10 nodes",
      "coverage": "measured",
      "evidenceWindow": "2026-10-09",
      "roles": [],
      "decision": "Historical runtime evidence",
      "summary": "October TensorFold speed evidence; runtime and workload conditions remain attached to the dated report. Superseded as the active coder; current DeepSeek has separate evidence.",
      "metrics": {
        "code_c1_tok_s": 80.09,
        "prose_c1_tok_s": 50.26,
        "prefill_ttft_s": 2.108,
        "code_four_request_mean_decode_tok_s": 53.7,
        "code_four_request_rate_sum_tok_s": 214.78
      },
      "context": "~6K input timing probe; configured 262144 is not full-window qualification.",
      "suiteIds": [
        "october-speed"
      ],
      "recordIds": [
        "2026-10-09-qwen-tensorfold-probe"
      ],
      "source": "Retained lab speed report · October 9, 22:56:32–22:57:14 CDT",
      "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."
    },
    {
      "id": "qwen38-flash-vllm",
      "name": "Qwen3.8-Flash-Next",
      "artifact": "NVFP4 · October NVIDIA serving artifact",
      "runtime": "vLLM two-node TP2; exact probe revision not retained",
      "hardware": "DGX Spark + ASUS GX10 · two GB10 nodes",
      "coverage": "measured",
      "evidenceWindow": "2026-10-05",
      "roles": [],
      "decision": "Historical runtime evidence",
      "summary": "Historical October vLLM speed evidence and October 10 configuration inspection. The short probe does not qualify quality or full context. Superseded as the active coder; current DeepSeek has separate evidence.",
      "metrics": {
        "code_c1_tok_s": 67.25,
        "prose_c1_tok_s": 44.79,
        "prefill_ttft_s": 1.851,
        "code_four_request_mean_decode_tok_s": 53.48,
        "code_four_request_rate_sum_tok_s": 213.91
      },
      "context": "~6K input timing probe; configured 262144 is not full-window qualification.",
      "suiteIds": [
        "october-speed"
      ],
      "recordIds": [
        "2026-10-05-qwen-vllm-probe"
      ],
      "source": "Retained lab speed report · October 5, 16:41:01–16:41:49 CDT",
      "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."
    },
    {
      "id": "sana-sprint",
      "name": "Sana Sprint 1.6B",
      "artifact": "1024×1024 local pipeline",
      "runtime": "Dated image-generation pipeline",
      "hardware": "Historical local image lane",
      "coverage": "measured",
      "evidenceWindow": "Jul 22, 2026",
      "roles": [],
      "decision": "Alternative",
      "summary": "Fastest warm timing of these three; the study did not establish a shared quantitative image-quality score.",
      "metrics": {
        "warm_generation_s": 0.79,
        "peak_gib": 10.19
      },
      "context": "Not a language-model context test.",
      "suiteIds": [
        "image-generation"
      ],
      "recordIds": [
        "2026-07-22-mage-flow"
      ],
      "source": "Dated public study and retained numeric archive; reviewed October 10, 2026.",
      "limitations": "Warm timings only; pipeline/model styles and quality goals differ."
    },
    {
      "id": "superqwen-studio-unscored",
      "name": "SuperQwen · current Studio agent",
      "artifact": "MLX 4-bit · October alias",
      "runtime": "MLX",
      "hardware": "Mac Studio · M5 Max · 48 GB",
      "coverage": "unscored",
      "evidenceWindow": "Configured artifact reviewed Oct 10, 2026",
      "roles": [],
      "decision": "Configured",
      "summary": "The current agent artifact has no local score in the reviewed archive. September Nemotron results remain attached to Nemotron.",
      "metrics": {},
      "context": "No independent full-context qualification in the reviewed evidence.",
      "suiteIds": [],
      "recordIds": [],
      "source": "Direct October 10 configuration review.",
      "limitations": "Packaged upstream reproduction scripts are not measurements from this lab."
    },
    {
      "id": "zimage-turbo",
      "name": "Z-Image-Turbo",
      "artifact": "1024×1024 local pipeline",
      "runtime": "Dated image-generation pipeline",
      "hardware": "Historical local image lane",
      "coverage": "measured",
      "evidenceWindow": "Jul 22, 2026",
      "roles": [],
      "decision": "Alternative",
      "summary": "Historical warm budget exceeded the under-20-GiB target.",
      "metrics": {
        "warm_generation_s": 5.7,
        "peak_gib": 21.69
      },
      "context": "Not a language-model context test.",
      "suiteIds": [
        "image-generation"
      ],
      "recordIds": [
        "2026-07-22-mage-flow"
      ],
      "source": "Dated public study and retained numeric archive; reviewed October 10, 2026.",
      "limitations": "Warm timings only; pipeline/model styles and quality goals differ."
    }
  ]
}
