{
 "_doc": "The lab's current claim and limits for one result, copied word for word from the entry named here. This result has no row in the portfolio's current-claims list. Its claim and limits are copied here word for word from the lab's own file named under source, at the commit and sha256 given there.",
 "id": "vllm:NIII-1",
 "title": "Part III One result that depends on which trace you point it at",
 "claim": "At 65,536 tokens the offload headroom is 13.004 pp on Mooncake (floor 0.8313, LRU 0.0387, ceiling 0.1687, LRU gets 22.9%) and 0.642 pp on WildChat-derived (floor 0.3267, LRU 0.6669, 99.0%); granularity explains 0.5% of the difference (0.573 -> 0.642 pp).",
 "limits": "block-hit ceiling, not GPU-hours; WildChat stream is derived_from_real_text",
 "source": {
  "lab": "vLLM lab",
  "commit": "eff30fb87201cb1690872532e5585697168543bf",
  "file": "top40.json",
  "sha256": "239790e974c7fe94f4a2a34628816bef732ae5ff6adead873b1cfa6295b1a9fa",
  "bytes": 380332,
  "entry_id": "NIII-1"
 },
 "result_receipt": "/receipts/vllm/results/offload_ceiling_generalization.json",
 "result_sha256": "a2488027d42be234f649ad67e0f94f8e1b13aeae1755c7813bcbd07e9b086784",
 "result_bytes": 6422,
 "entry": {
  "id": "NIII-1",
  "block": "B",
  "title": "Part III One result that depends on which trace you point it at",
  "claim": "At 65,536 tokens the offload headroom is 13.004 pp on Mooncake (floor 0.8313, LRU 0.0387, ceiling 0.1687, LRU gets 22.9%) and 0.642 pp on WildChat-derived (floor 0.3267, LRU 0.6669, 99.0%); granularity explains 0.5% of the difference (0.573 -> 0.642 pp).",
  "source": "NEGATIVE_RESULTS.md:388-443",
  "artifact_path": "results/offload_ceiling_generalization.json",
  "reproduce_command": "python3 bench/run_offload_ceiling_generalization.py",
  "run_status": "RAN",
  "cost_estimate": "$0.00",
  "third_party_graded": {
   "system": "Mooncake FAST'25 public trace vs WildChat-derived stream against a Lean-bounded ceiling",
   "bar": "s50_gate_pct 12.0 (inline in results/sprint50_offload_oracle.json)",
   "prereg_receipt": "inline",
   "bar_sealed_before_run": false
  },
  "negative_control": "granularity control (16 -> 64 tokens)",
  "limits": "block-hit ceiling, not GPU-hours; WildChat stream is derived_from_real_text",
  "scope": "",
  "rank": 60,
  "site": "inference",
  "ip_class": "UNKNOWN",
  "axis": "third_party_graded",
  "verdict": "CORRECT AS WRITTEN",
  "verdict_receipt": "results/offload_ceiling_generalization.json; results/sprint50_offload_oracle.json (s50_authorised_to_spend true, s50_spend_needs_owner_approval true); results/s46b_engine_hit_rate.json",
  "largest_supported_sentence": "as written, and the oracle's own precondition ('until a run reports the engine's counters') is met by a measurement that is not a bridge collision",
  "evidence_tier": "NUMBERS_IDENTICAL",
  "runtime_estimate": "seconds-minutes CPU",
  "output_excerpt": "",
  "improvable": "UNDERCLAIMED: corroborated by an engine",
  "pin_sweep": {
   "n_numbers": 13,
   "n_unfound": 0,
   "unfound": []
  },
  "paths": [
   "results/offload_ceiling_generalization.json",
   "results/sprint50_offload_oracle.json",
   "results/s46b_engine_hit_rate.json"
  ],
  "reverdicted_after_merge": "d5340ad",
  "gate4": true,
  "gate4_verdict": "REPRODUCED"
 }
}