{
  "schema": "anvil-serving.benchmark-source-registry/v1",
  "campaign_id": "2026-09-23-next-model-qualification",
  "observed_at": "2026-09-23T11:05:32Z",
  "sources": [
    {"id":"lil-daily-archive-219-screen","url":"https://github.com/local-inference-lab/rtx6kpro/tree/2960b922466aa3167da0c7f6b3918b329bc36216/daily-summaries","published_or_observed":"2026-09-23","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"maintainer/community archive screening","hardware_runtime_relevance":"Mixed hardware archive; all 219 reports were screened for dual RTX PRO 6000 Max-Q, SM120, GLM-5.3-Flash, vision, cache, and failure leads.","decision_impact":"Screening identifies leads only; it does not independently verify each report claim or establish a qualified replacement."},
    {"id":"lil-daily-glm-tp2-headroom","url":"https://github.com/local-inference-lab/rtx6kpro/blob/2960b922466aa3167da0c7f6b3918b329bc36216/daily-summaries/2026-09/2026-09-22.md","published_or_observed":"2026-09-22","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"maintainer daily report","hardware_runtime_relevance":"TP2 Spark/SM120 report and recipe lead for reduced KV headroom.","decision_impact":"Nominates a no-MTP, reduced-KV GLM control; it is an external prior, not proof of local long-output or concurrency correctness."},
    {"id":"lil-vllm-pr802-flashkda-memory","url":"https://github.com/local-inference-lab/vllm/pull/802","published_or_observed":"2026-09-23","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"merged upstream pull request","hardware_runtime_relevance":"Documents an SM120 FlashKDA device-stack allocation correction relevant to startup memory headroom.","decision_impact":"Shows that runtime revision can materially change fit; it does not establish model quality, capacity, or a local allocation total."},
    {"id":"lil-vllm-pr803-kda-retention","url":"https://github.com/local-inference-lab/vllm/pull/803","published_or_observed":"2026-09-23","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"merged upstream pull request","hardware_runtime_relevance":"Documents retained synthetic KDA inputs in a Spark TP2 configuration.","decision_impact":"Supports remeasuring memory on the exact image; it does not prove admission for this campaign."},
    {"id":"lil-vllm-pr804-profiling-kv","url":"https://github.com/local-inference-lab/vllm/pull/804","published_or_observed":"2026-09-23","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"upstream pull request","hardware_runtime_relevance":"Reports retained profiling-KV storage relevant to memory accounting.","decision_impact":"Treat as an unmerged lead only; it cannot be represented as an applied runtime fix."},
    {"id":"glm-spark-nvfp4-checkpoint-a608241","url":"https://huggingface.co/local-inference-lab/GLM-5.3-Flash-NVFP4-Spark/tree/a608241037e4c2565356bff7ca293f2133888f88","published_or_observed":"2026-09-17","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"checkpoint metadata","hardware_runtime_relevance":"Immutable GLM-5.3-Flash Spark NVFP4 candidate identity.","decision_impact":"Pins the candidate weights; repository byte size is not a GPU-memory, host-RAM, context, quality, or stability measurement."},
    {"id":"lil-kk-runtime-release-fcd71e8","url":"https://github.com/local-inference-lab/blackwell-llm-docker/releases/tag/karmic-kraken-fcd71e8abd3c987b81b095fe42a331605670c6cfa39073b8599db71385e3276d","published_or_observed":"2026-09-23","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"runtime release","hardware_runtime_relevance":"SM120/CUDA 13.4 runtime identity for current GLM candidate work.","decision_impact":"Native smoke status does not qualify end-to-end GLM, Pi coding, vision, long output, or concurrent capacity."},
    {"id":"lil-vllm-issue808-beta3221","url":"https://github.com/local-inference-lab/vllm/issues/808","published_or_observed":"2026-09-23","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"maintainer integration report","hardware_runtime_relevance":"Names the beta3221 Qwen TP2/MTP3 dual RTX PRO 6000 Max-Q build and associated source reconstruction.","decision_impact":"Nominates Qwen as a hardware-matched challenger; reported results are external evidence and do not replace local qualification."},
    {"id":"lil-qwen38-tp2-benchmark-2960b92","url":"https://github.com/local-inference-lab/rtx6kpro/blob/2960b922466aa3167da0c7f6b3918b329bc36216/benchmarks/qwen38-tp2-sglang.md","published_or_observed":"2026-09-22","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"maintainer measured benchmark","hardware_runtime_relevance":"Exact Qwen QAD checkpoint, beta3221, TP2/DCP1/MTP3, dual Max-Q, 524288 configured context, 16 slots, and 8 GiB KV per rank.","decision_impact":"Provides a reproducible external throughput prior; vision, natural long output, and local Pi coding quality remain separately required."},
    {"id":"qwen38-nvfp4-checkpoint-7c4f1bc","url":"https://huggingface.co/local-inference-lab/Qwen3.8-Flash-Next-NVFP4/tree/7c4f1bc1a2d6847e0cbc01ac6b823f00251de8dd","published_or_observed":"2026-09-22","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"checkpoint metadata and template","hardware_runtime_relevance":"Immutable Qwen3.8-Flash-Next NVFP4 QAD candidate identity with multimodal template assets.","decision_impact":"Pins the candidate; checkpoint metadata alone does not prove memory fit, parser behavior, long output, vision admission, or coding correctness."},
    {"id":"lil-vllm-622912b-qwen3-registry","url":"https://github.com/local-inference-lab/vllm/blob/622912b9b2cfd04784da5b0dfc69ce8aa5aee02d/vllm/reasoning/__init__.py","published_or_observed":"2026-09-22","observed_at":"2026-09-23T11:05:32Z","age_class":"current","evidence_type":"pinned runtime source","hardware_runtime_relevance":"The exact beta3221 engine revision registers selector qwen3 to the Python Qwen3 parser adapter.","decision_impact":"No documented selector for the Rust Qwen3 reasoning parser exists in this pin; do not invent a native-parser toggle or alter router validation."},
    {"id":"lil-vllm-622912b-rust-parser-binding","url":"https://github.com/local-inference-lab/vllm/blob/622912b9b2cfd04784da5b0dfc69ce8aa5aee02d/rust/src/parser/python/src/lib.rs","published_or_observed":"2026-09-22","observed_at":"2026-09-23T11:05:32Z","age_class":"current","evidence_type":"pinned runtime source","hardware_runtime_relevance":"Rust Python bridge exports a limited tool-parser surface; it does not expose a Qwen3 reasoning-parser selector.","decision_impact":"Confirms the parser distinction is a runtime limitation, not authorization to trim visible content or relax the validator."},
    {"id":"mimo-flash-rl-official-5711b26","url":"https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL/tree/5711b268169967567844e1e560e8a3966da959b1","published_or_observed":"2026-09-22","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"official model card and checkpoint","hardware_runtime_relevance":"Multimodal MiMo Flash candidate with official larger-TP serving guidance.","decision_impact":"Candidate research lead only; no official TP2 SM120 recipe satisfying this campaign's long-output and vision gates was identified."},
    {"id":"mimo-nvfp4-tp2-34cebc5","url":"https://huggingface.co/ProCreations/MiMo-V2.6-Flash-RL-NVFP4/tree/34cebc5d75743a93ab76145ec609cfef9da8c417","published_or_observed":"2026-09-22","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"quantization maintainer model card","hardware_runtime_relevance":"Claims vLLM TP2 on two 96 GB RTX PRO 6000 Server GPUs at 12288 context and C2.","decision_impact":"A concrete feasibility lead, but the cited context is below the long-output objective and does not establish vision or local stability."},
    {"id":"mimo-dgx-spark-tool-history-issue","url":"https://github.com/tonyd2wild/MiMo-V2.6-Flash-DGX-Spark-Recipe/issues/1","published_or_observed":"2026-09-23","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"upstream issue linked by recipe author","hardware_runtime_relevance":"DGX Spark-specific tool-history and reasoning diagnosis, not a discrete SM120 TP2 result.","decision_impact":"Motivates a prior-assistant-thinking plus tool-result regression shape; it is not a portable remediation or candidate qualification."},
    {"id":"mimo-local-llama-empty-response-report","url":"https://www.reddit.com/r/LocalLLaMA/comments/1wnjbwn/mimov26flash_on_vllm_fixes_for_empty_responses/","published_or_observed":"2026-09-22","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"first-person community diagnosis","hardware_runtime_relevance":"DGX Spark and a different runtime; not the dual discrete RTX PRO 6000 Max-Q configuration.","decision_impact":"Records a template/thinking-history failure lead and reported output cap; it cannot establish a fix or hardware-fit claim."},
    {"id":"mimo-dgx-spark-community-measurement","url":"https://x.com/yume_arasaki/status/2102548506061750782","published_or_observed":"2026-09-22","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"first-person community measurement","hardware_runtime_relevance":"Two DGX Spark GB10 systems, not SM120 discrete GPUs.","decision_impact":"Reports output-shape-dependent performance and an arithmetic failure; it does not transfer to the target hardware."},
    {"id":"community-qwen-thin-decode-observation","url":"https://x.com/Klausbjarner/status/2102568631628681288","published_or_observed":"2026-09-22","observed_at":"2026-09-23T09:55:29Z","age_class":"current","evidence_type":"first-person community measurement","hardware_runtime_relevance":"Hardware and complete launch controls are not pinned.","decision_impact":"Retained as a non-transferable warning that output shape can reverse an apparent decode gain; it is not comparative benchmark evidence."}
  ]
}