{
  "version": 1,
  "architecture": {
    "representation": "gqa",
    "layers": 32,
    "query_heads": 32,
    "kv_heads": 8,
    "key_dimension": 128,
    "value_dimension": 128,
    "bytes_per_element": 2
  },
  "runtime": {
    "cache_pool_bytes": 68719476736,
    "page_tokens": 16,
    "prefix_cache": true,
    "retention": "lru"
  },
  "workload": {
    "context_lengths": [8192, 32768, 131072],
    "concurrency": [1, 4, 16],
    "shared_prefix_fraction": [0, 0.5, 0.9],
    "retrieval_k": [0, 4, 8, 16]
  },
  "fixtures": {
    "single_retrieval": 25,
    "distractor_retrieval": 25,
    "multi_hop": 25,
    "position_sensitive": 25
  },
  "metrics": {
    "serving": [
      "physical_cache_bytes",
      "page_waste_bytes",
      "cached_tokens",
      "evicted_blocks",
      "ttft_ms",
      "tpot_ms"
    ],
    "retrieval": ["recall_at_k", "gold_passage_rank", "retrieved_tokens"],
    "task": [
      "exact_match",
      "citation_precision",
      "citation_recall",
      "canary_retention"
    ]
  },
  "required_demonstrations": [
    "change one early token and record the prefix miss",
    "compare TTFT and TPOT with a warm prefix",
    "increase page size and measure rounding waste",
    "force LRU thrashing with interleaved prefixes",
    "compact history and record lost facts",
    "sweep retrieval_k and report recall, rank, token cost, and task quality",
    "compare fixed recurrent-state bytes with token-growing KV under matched concurrency"
  ]
}