{
  "artifact": "s08_o14_engine_rows",
  "sprint": "S08-O14",
  "claim": "The two rows this lane keeps under the S08 ownership split regenerate today from committed data: the engine-determinism 2x2 and the engine-counter sentence.",
  "method": "The determinism table is recomputed from the per-boot completion strings inside results/d3_determinism_2x2.json, with a planted and a scrubbed control run first; the engine-counter analysis is re-run by bench/run_s06_c1_analysis.py and compared value by value with the committed certificate, which is then restored.",
  "controls": {
    "planted_divergence": {
      "cell": "apc_off_inv0",
      "baseline": 0,
      "with_one_planted": 1,
      "control_fires": true
    },
    "scrubbed_divergence": {
      "cell": "apc_on_inv0",
      "baseline": 6,
      "with_all_removed": 0,
      "control_fires": true
    }
  },
  "controls_ok": true,
  "row1_engine_determinism": {
    "published": {
      "apc_on_inv0": 6,
      "apc_off_inv0": 0,
      "apc_on_inv1": 0,
      "apc_off_inv1": 0
    },
    "rederived": {
      "apc_on_inv0": 6,
      "apc_off_inv0": 0,
      "apc_on_inv1": 0,
      "apc_off_inv1": 0
    },
    "identical": true,
    "per_cell": {
      "apc_on_inv0": {
        "per_prompt": {
          "colours": {
            "boots": 6,
            "boots_first_differs_rest_agree": 6,
            "boots_all_identical": 0,
            "boots_rest_disagree": 0
          },
          "bridge": {
            "boots": 6,
            "boots_first_differs_rest_agree": 0,
            "boots_all_identical": 6,
            "boots_rest_disagree": 0
          }
        },
        "divergent_boot_prompts": 6,
        "prefix_caching": true,
        "batch_invariant": 0
      },
      "apc_off_inv0": {
        "per_prompt": {
          "colours": {
            "boots": 6,
            "boots_first_differs_rest_agree": 0,
            "boots_all_identical": 6,
            "boots_rest_disagree": 0
          },
          "bridge": {
            "boots": 6,
            "boots_first_differs_rest_agree": 0,
            "boots_all_identical": 6,
            "boots_rest_disagree": 0
          }
        },
        "divergent_boot_prompts": 0,
        "prefix_caching": false,
        "batch_invariant": 0
      },
      "apc_on_inv1": {
        "per_prompt": {
          "colours": {
            "boots": 6,
            "boots_first_differs_rest_agree": 0,
            "boots_all_identical": 6,
            "boots_rest_disagree": 0
          },
          "bridge": {
            "boots": 6,
            "boots_first_differs_rest_agree": 0,
            "boots_all_identical": 6,
            "boots_rest_disagree": 0
          }
        },
        "divergent_boot_prompts": 0,
        "prefix_caching": true,
        "batch_invariant": 1
      },
      "apc_off_inv1": {
        "per_prompt": {
          "colours": {
            "boots": 6,
            "boots_first_differs_rest_agree": 0,
            "boots_all_identical": 6,
            "boots_rest_disagree": 0
          },
          "bridge": {
            "boots": 6,
            "boots_first_differs_rest_agree": 0,
            "boots_all_identical": 6,
            "boots_rest_disagree": 0
          }
        },
        "divergent_boot_prompts": 0,
        "prefix_caching": false,
        "batch_invariant": 1
      }
    },
    "sentence": "On one model and one GPU class, the first request after a cold boot at temperature 0 diverges from the requests that follow it in 6 boot-prompts with prefix caching on and 0 with it off; the batch-invariant flag closes it (0 divergent).",
    "receipt": "results/d3_determinism_2x2.json"
  },
  "row2_engine_counter": {
    "rerun_exit_code": 1,
    "committed": {
      "s46c_median_ratio": 0.949884,
      "s46c_p_value": 0.62503,
      "engine_hit_rate_apc_on_max": 0.038505905648019395,
      "engine_hits_total_apc_off_max": 0,
      "s47b_median_ratio": 1.024752,
      "s47b_significant": false,
      "primary_bar_passes": false,
      "aa_bar_passes": true
    },
    "regenerated": {
      "s46c_median_ratio": 0.949884,
      "s46c_p_value": 0.62503,
      "engine_hit_rate_apc_on_max": 0.038505905648019395,
      "engine_hits_total_apc_off_max": 0,
      "s47b_median_ratio": 1.024752,
      "s47b_significant": false,
      "primary_bar_passes": false,
      "aa_bar_passes": true
    },
    "identical": true,
    "differing_keys": [],
    "note": "the committed certificate is restored after the comparison; the re-run is the receipt, not a rewrite",
    "sentence": "With the trace-to-engine bridge gated in both directions, vLLM 0.28's own prefix-cache counters read 0.0385 against this lane's hash-only model at 0.0387, with zero hits when caching is off, in eight of eight replicates.",
    "receipt": "results/s06_c1_fixed_bridge_n8.json"
  },
  "ownership": {
    "decided_by": "orchestrator S08 ownership table, row O14, accepting the two lanes' own split",
    "vllm_keeps": [
      "engine determinism (this certificate, row 1)",
      "the engine-counter sentence (row 2)"
    ],
    "f5_keeps": [
      "the theorems",
      "the cache-ceiling walls",
      "key-axis conformance",
      "the isolation tax",
      "the timing channel"
    ],
    "this_lane_cedes": "the isolation-tax claim; this lane's rows on it become receipts under f5's claim, and its own sentence retires to a cross-reference"
  },
  "verdict": "BOTH ROWS REGENERATE: controls both fire; engine determinism recomputed from the per-boot completion strings gives {'apc_on_inv0': 6, 'apc_off_inv0': 0, 'apc_on_inv1': 0, 'apc_off_inv1': 0}, matching the published {'apc_on_inv0': 6, 'apc_off_inv0': 0, 'apc_on_inv1': 0, 'apc_off_inv1': 0}; the engine-counter analysis re-runs with every headline value identical (engine hit rate with caching on 0.038505905648019395, total hits with it off 0).",
  "wall_s": 1.1,
  "cost_usd": 0.0,
  "does_not_prove": [
    "Anything beyond one model (Qwen2.5-1.5B-Instruct) and one GPU class for the determinism row: 6 boots per cell, two prompts, and no second model family.",
    "That the divergence is vLLM's fault rather than the stack beneath it: the cause is not isolated here, and the batch-invariant flag closing it is consistent with several causes.",
    "A knee uplift for the engine-counter row: the paired ratio at n=8 is not significant, and the row is the agreement between the engine's counters and the model, not an uplift.",
    "That either row regenerates the GPU run: both re-derive the analysis from committed raw data. The runs themselves are paid and are not repeated here."
  ],
  "generated_utc": "2026-09-04T19:58:43.987444+00:00",
  "git_rev": "fd54c78",
  "git_dirty": true,
  "run_id": "7f4c665b39134dd5",
  "host_kind": "Darwin-arm64",
  "provenance_schema": "gpu-cert-provenance/1"
}