{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-10-17-glm53-nvidia-fp4-ab",
  "complete": true,
  "identity": {
    "configuration": "configuration.json",
    "baseline": "EXL3 r11, TP2/DCP1, no speculation",
    "candidate": "NVIDIA FP4 r2, TP2/DCP2, MTP3",
    "hardware": "two96GB RTX PRO6000 Blackwell Max-Q"
  },
  "workload": {
    "manifest": "workload-manifest.json",
    "diagnostic_plan": "diagnostic-plan.json"
  },
  "measurement": {
    "path": "online direct; restoration also authenticated routed",
    "warm_cold_state": "Initial candidate cold startup then sequential preflight/vision/agentic/context; diagnostic separate restart and full/isolated repeats",
    "sample_count": 18,
    "sample_unit": "initial context observations per profile",
    "statistics": [
      "deterministic oracle pass counts"
    ],
    "baseline_timing_requests_completed": 32,
    "candidate_timing_requests_dispatched": 0
  },
  "gates": [
    {
      "name": "functional preflight",
      "baseline": "9/9 applicable + correctedOCR1/1; original badOCRfixture retained",
      "candidate": "9/9 + OCR1/1"
    },
    {
      "name": "vision",
      "baseline": "18/18",
      "candidate": "18/18"
    },
    {
      "name": "synthetic agentic",
      "baseline": "18/18",
      "candidate": "18/18"
    },
    {
      "name": "initial context",
      "baseline": "18/18",
      "candidate": "17/18; failed",
      "failed_case": "native-needle-318464-0.900-r0",
      "failed_input_tokens": 316249,
      "largest_successful_input_tokens": 316266
    },
    {
      "name": "first restoration",
      "status": "passed",
      "direct": "2/2",
      "authenticated_router": "2/2",
      "exact_original_container": true
    },
    {
      "name": "unchanged r2 diagnostic",
      "full_replay": "18/18",
      "original_failed_case_followups": "3/3",
      "configuration_changed": false,
      "original_failure_preserved": true,
      "mtp_off": "not run; predeclared condition not met"
    },
    {
      "name": "final restoration",
      "status": "passed",
      "direct": "2/2",
      "authenticated_router": "2/2",
      "exact_original_container": true,
      "candidate_absent": true
    }
  ],
  "headline_results": [
    "Initial exact r2 context gate:17/18 versus incumbent18/18.",
    "Unchanged r2 full replay:18/18; originally failed case passed all three follow-ups. Cause remains unproven; no configuration fix established.",
    "Exact original r11 restored; direct and authenticated routed checks2/2 each; original configs unchanged; candidate absent."
  ],
  "failures": [
    {
      "artifact": "native/candidate/0-context.json",
      "case_id": "native-needle-318464-0.900-r0",
      "visible_response": "The access marker for ORCHID is not present in the provided document.",
      "finish_reason": "stop",
      "runtime_error": false,
      "fixture_check": "context-fixture-check.json"
    }
  ],
  "limitations": [
    "One initial sampled miss does not establish failure frequency or identify quantization/MTP/runtime as its cause.",
    "Candidate timing, overlap/endurance and repository-scale coding were not run.",
    "Startup KV concurrency estimates do not prove four simultaneous327K requests.",
    "Host embeddings enabled; not strict GPU-only residency.",
    "Diagnostic results remain separate from the original failed gate; sample size and request/cache histories limit causal interpretation."
  ],
  "decision": {
    "evidence_labels": [
      "scout",
      "diagnostic"
    ],
    "decision_labels": [
      "no-promotion"
    ],
    "promotion_authorized": false,
    "next_gate": "No further diagnostic actions scheduled. Candidate timing, overlap/endurance and repository-scale coding remain untested; no qualification or promotion established."
  },
  "evidence": [
    "configuration.json",
    "workload-manifest.json",
    "native/baseline/baseline-native-index.json",
    "native/candidate/candidate-native-index.json",
    "native/restoration/restoration-native-index.json",
    "context-fixture-check.json",
    "diagnostic-plan.json",
    "diagnostic-result.json",
    "native/diagnostic/diagnostic-native-index.json",
    "restoration.json",
    "verification.json"
  ]
}