{
  "schema": "anvil-serving.runtime-workload/v1",
  "purpose": "Synthetic reproduction of long-context decode plus fresh prefill; separate configuration and hardware-envelope screens.",
  "incident_shapes": [
    {
      "anchor_computed_tokens": 174956,
      "fresh_prompt_tokens": 31963
    },
    {
      "anchor_computed_tokens": 201166,
      "fresh_prompt_tokens": 46916
    }
  ],
  "synthetic_prompt": "Three independently checkable retrieval records spread through repeated prose; anchor continues a long engineering guide. No user task content copied.",
  "prompt_count_tolerance": "0.5 percent or 8 tokens; compare tokenizer and usage",
  "output_policy": {
    "anchor_cap": 4096,
    "contender_cap": 1024,
    "thinking_enabled": true,
    "temperature": 0,
    "visible_capture_chars": 8192,
    "length_finish": "diagnostic allowed; not natural-completion qualification"
  },
  "cases": [
    "small adapter qualification",
    "serial control at each incident size",
    "overlap at each incident size",
    "three repeated-prefix rounds",
    "larger context after passing smaller cases"
  ],
  "coverage_limit": "Client overlap only; exact scheduler batch, cache hit/eviction, kernel page ID and graph state require separate owner evidence.",
  "new_failure_stop": "No following trial on identity drift, runtime failure, Xid, OOM or incomplete coverage; preserve failed artifacts before recovery."
}