Research → synthesis: Spot vs reserved GPU capacity?

{
  "kind": "research.note",
  "query": "Spot vs reserved GPU capacity?",
  "claims": [
    {
      "claim": "Spot GPU slices cut expected spend but can be reclaimed, so long jobs pay a preemption tax.",
      "source": "doc:gpuaas-spot-reserved",
      "confidence": 0.92
    },
    {
      "claim": "Burst or exploratory jobs usually fit spot; training runs with costly checkpoints usually fit reserved.",
      "source": "doc:gpuaas-spot-reserved",
      "confidence": 0.87
    },
    {
      "claim": "Reserved GPU slices trade a commitment cost for a lower chance of interruption and more predictable start times.",
      "source": "doc:gpuaas-spot-reserved",
      "confidence": 0.89
    },
    {
      "claim": "The tradeoff is interruption risk versus commitment, not a vendor catalog.",
      "source": "doc:gpuaas-spot-reserved",
      "confidence": 0.84
    }
  ],
  "reason": "strong_topical_match",
  "privacy": "restricted",
  "sources": [
    "doc:gpuaas-spot-reserved"
  ],
  "summary": "Mapped 4 claims from 1 source: Spot versus reserved GPU tradeoffs. Spot GPU slices cut expected spend but can be reclaimed, so long jobs pay a preemption tax.",
  "corpusScope": "public",
  "evidenceStatus": "sufficient",
  "retrievalDebug": {
    "query": "Spot vs reserved GPU capacity?",
    "corpusHits": [
      "doc:gpuaas-spot-reserved",
      "doc:gpuaas-allocation-basics",
      "doc:gpuaas-bid-ranking",
      "doc:gpuaas-job-orchestration",
      "doc:gpuaas-multitenant-failure",
      "doc:gpuaas-utilization-signals",
      "doc:multi-agent-synergy"
    ],
    "corpusScope": "public",
    "fallbackUsed": false,
    "topCandidates": [
      {
        "id": "doc:gpuaas-spot-reserved",
        "tags": [
          "gpuaas",
          "compute",
          "allocation"
        ],
        "score": 45,
        "strong": true,
        "matchedTerms": [
          "spot",
          "reserved",
          "gpu"
        ]
      },
      {
        "id": "doc:gpuaas-allocation-basics",
        "tags": [
          "gpuaas",
          "compute",
          "allocation"
        ],
        "score": 19,
        "strong": true,
        "matchedTerms": [
          "gpu"
        ]
      },
      {
        "id": "doc:gpuaas-bid-ranking",
        "tags": [
          "gpuaas",
          "compute",
          "allocation",
          "research"
        ],
        "score": 19,
        "strong": true,
        "matchedTerms": [
          "gpu"
        ]
      },
      {
        "id": "doc:gpuaas-job-orchestration",
        "tags": [
          "gpuaas",
          "compute",
          "orchestration"
        ],
        "score": 19,
        "strong": true,
        "matchedTerms": [
          "gpu"
        ]
      },
      {
        "id": "doc:gpuaas-multitenant-failure",
        "tags": [
          "gpuaas",
          "compute",
          "allocation"
        ],
        "score": 19,
        "strong": true,
        "matchedTerms": [
          "gpu"
        ]
      },
      {
        "id": "doc:gpuaas-utilization-signals",
        "tags": [
          "gpuaas",
          "compute",
          "allocation"
        ],
        "score": 17,
        "strong": true,
        "matchedTerms": [
          "gpu"
        ]
      },
      {
        "id": "doc:multi-agent-synergy",
        "tags": [
          "multi-agent",
          "synergy",
          "specialization",
          "handoff"
        ],
        "score": 2,
        "strong": false,
        "matchedTerms": []
      }
    ],
    "guardTriggered": false,
    "selectedSources": [
      "doc:gpuaas-spot-reserved"
    ]
  }
}

{
  "kind": "research.synthesis",
  "brief": "Spot GPU slices cut expected spend but can be reclaimed, so long jobs pay a preemption tax. Burst or exploratory jobs usually fit spot; training runs with costly checkpoints usually fit reserved. Reserved GPU slices trade a commitment cost for a lower chance of interruption and more predictable start times. The tradeoff is interruption risk versus commitment, not a vendor catalog. Sources: doc:gpuaas-spot-reserved — Spot versus reserved GPU tradeoffs.",
  "query": "Spot vs reserved GPU capacity?",
  "reason": "strong_topical_match",
  "privacy": "restricted",
  "sources": [
    "doc:gpuaas-spot-reserved"
  ],
  "summary": "Synthesized 4 focused claims from 1 source.",
  "corpusScope": "public",
  "upstreamStep": "research",
  "evidenceStatus": "sufficient"
}