{
  "schema": "anvil-serving.candidate-failure.v1",
  "date": "2026-07-12",
  "candidate": "poolside/Laguna-XS-2.1-NVFP4",
  "revision": "07133fb3df1cc3111478e24ee71a823a598c8c2f",
  "hardware": "NVIDIA RTX PRO 6000 Blackwell 96 GB (sm_120)",
  "status": "rejected_tested_recipes",
  "evidence_limit": "Healthy-but-semantically-wrong response bodies and preflight JSON were not retained before container recreation. Those observations are operator-reported and incomplete; retained logs substantiate only the startup and forced-runner failures.",
  "attempts": [
    {
      "engine": "vllm 0.23.1rc1.dev531+ga65f93fb2",
      "configuration": "262144 context, five sequences, FP8 KV, FlashInfer attention",
      "health": 200,
      "preflight_default": "0/4",
      "preflight_no_thinking": "2/4; structured JSON and 20-way tool batch failed",
      "evidence_status": "operator-observed-incomplete",
      "failure": "corrupted visible output; tool batch 0/20"
    },
    {
      "engine": "vllm 0.23.1rc1.dev531+ga65f93fb2",
      "configuration": "FP8 KV skipped on layers 0-39; 131072 context",
      "health": null,
      "evidence_status": "retained-bounded-log",
      "failure": "startup did not progress beyond CUDA graph/cache profiling"
    },
    {
      "engine": "vllm 0.23.1rc1.dev531+ga65f93fb2",
      "configuration": "FP8 KV skipped on layers 0-39; 262144 context",
      "health": null,
      "evidence_status": "operator-observed-incomplete",
      "failure": "startup reportedly did not progress beyond CUDA graph/cache profiling"
    },
    {
      "engine": "sglang dev-cu13-laguna-xs-2-1",
      "configuration": "262144 context, five running requests, explicit checkpoint chat template, FlashInfer attention, SWARadixCache",
      "health": 200,
      "preflight_no_thinking": "2/4; 131072 needle empty; tool batch 0/20",
      "evidence_status": "operator-observed-incomplete",
      "failure": "repetitive or unrelated visible output despite successful service health"
    },
    {
      "engine": "sglang dev-cu13-laguna-xs-2-1",
      "configuration": "forced trtllm_mha attention",
      "health": null,
      "evidence_status": "operator-observed-incomplete",
      "failure": "ValueError: TRTLLM MHA prefill supports Blackwell SM100 only in this build"
    },
    {
      "engine": "sglang dev-cu13-laguna-xs-2-1",
      "configuration": "forced flashinfer_trtllm MoE runner; FlashInfer attention",
      "health": null,
      "evidence_status": "retained-bounded-log",
      "failure": "warmup AssertionError: layer.routing_method_type is not None"
    }
  ],
  "decision": "Do not rank or promote Laguna XS 2.1 NVFP4 from these recipes on RTX PRO 6000 sm_120."
}