{
  "artifact": "s06_c1_fixed_bridge_n8",
  "sprint": "S06-C1",
  "claim": "Prefix caching's knee uplift on Mooncake at 65,536 tokens on the fixed bridge, at n=8 with a matching A/A, against the sealed bars.",
  "method": "bench/stats over the paired analyses the driver wrote; engine counters read from the raws; every sealed clause evaluated.",
  "prereg": [
    "results/s46c_apc_fixed_bridge_n8.prereg.json",
    "results/s47b_aa_fixed_bridge_n8.prereg.json"
  ],
  "s46c": {
    "median_of_per_rep_ratios": 0.949884,
    "per_rep_ratios": [
      1.0,
      0.949884,
      0.926664,
      0.949884,
      1.176951,
      1.088476,
      0.877773,
      0.924826
    ],
    "fixed_n_test": {
      "n": 8,
      "df": 7,
      "mean_log": -0.017476,
      "s_y": 0.096718,
      "se": 0.034195,
      "t": -0.5111,
      "p_value": 0.62503,
      "significant_at_05": false,
      "median_ratio": 0.949884,
      "mean_ratio": 0.982676,
      "ci95_ratio": [
        0.906346,
        1.065434
      ],
      "ci_is_fixed_n_only": true,
      "ci_caveat": "valid only if n was fixed in advance; if replicates were added after looking, read ratio_confidence_sequence instead"
    },
    "robust_quantity": {
      "decidable": true,
      "headline": "absolutes",
      "arm_cv_pct": {
        "apc_off": 3.07,
        "apc_on": 7.34
      },
      "ratio_cv_pct": 10.43,
      "steadiest_arm": "apc_off",
      "steadiest_arm_cv_pct": 3.07,
      "why": "the paired ratio (CV 10.43%) is NOISIER than the 'apc_off' absolute (CV 3.07%), so the ratio amplifies jitter instead of cancelling it -- quote the ABSOLUTES and derive the ratio"
    },
    "knee_binding_summary": {
      "n_arm_knees": 16,
      "n_harness_bound": 0,
      "n_slo_bound": 1,
      "ratio_is_quotable": true
    },
    "engine": {
      "apc_off": {
        "hit_rate_max": null,
        "hits_total_max": 0,
        "queries_total_max": 0,
        "n_points": 64,
        "distinct_text_check": null
      },
      "apc_on": {
        "hit_rate_max": 0.038505905648019395,
        "hits_total_max": 311504.0,
        "queries_total_max": 8440056.0,
        "n_points": 58,
        "distinct_text_check": null
      }
    },
    "mde_pct_at_n8": 11.83,
    "uplift_pct_at_median": 5.28,
    "wall_s": 2083.7
  },
  "s47b": {
    "median_of_per_rep_ratios": 1.024752,
    "per_rep_ratios": [
      1.0,
      1.0,
      1.051404,
      1.025058,
      1.051404,
      1.024445,
      1.025058,
      0.871491
    ],
    "fixed_n_test": {
      "n": 8,
      "df": 7,
      "mean_log": 0.004544,
      "s_y": 0.060461,
      "se": 0.021376,
      "t": 0.2126,
      "p_value": 0.837719,
      "significant_at_05": false,
      "median_ratio": 1.024752,
      "mean_ratio": 1.004554,
      "ci95_ratio": [
        0.95504,
        1.056636
      ],
      "ci_is_fixed_n_only": true,
      "ci_caveat": "valid only if n was fixed in advance; if replicates were added after looking, read ratio_confidence_sequence instead"
    },
    "robust_quantity": {
      "decidable": true,
      "headline": "absolutes",
      "arm_cv_pct": {
        "a1": 6.27,
        "a2": 2.25
      },
      "ratio_cv_pct": 5.64,
      "steadiest_arm": "a2",
      "steadiest_arm_cv_pct": 2.25,
      "why": "the paired ratio (CV 5.64%) is NOISIER than the 'a2' absolute (CV 2.25%), so the ratio amplifies jitter instead of cancelling it -- quote the ABSOLUTES and derive the ratio"
    },
    "knee_binding_summary": {
      "n_arm_knees": 16,
      "n_harness_bound": 2,
      "n_slo_bound": 1,
      "ratio_is_quotable": false
    },
    "engine": {
      "a1": {
        "hit_rate_max": 0.037626290631247,
        "hits_total_max": 313120.0,
        "queries_total_max": 8440056.0,
        "n_points": 61,
        "distinct_text_check": null
      },
      "a2": {
        "hit_rate_max": 0.03833149922228004,
        "hits_total_max": 314208.0,
        "queries_total_max": 8440056.0,
        "n_points": 64,
        "distinct_text_check": null
      }
    },
    "mde_pct_at_n8": 7.24,
    "wall_s": 4110.0
  },
  "clauses": {
    "s46c_ratio_below_1": true,
    "s46c_significant_at_05": false,
    "s46c_ratio_cv_within_cap": false,
    "s46c_engine_apc_on_within_ceiling": true,
    "s46c_engine_apc_off_zero": true,
    "s46c_no_harness_bound_knees": true,
    "s47b_within_band": true,
    "s47b_not_significant": true,
    "s47b_ratio_quotable_by_harness_rule": false
  },
  "primary_bar_passes": false,
  "aa_bar_passes": true,
  "verdict": "CANDIDATE 1 BAR FAILS. S46c (n=8, fixed bridge): paired median knee ratio apc_off/apc_on 0.949884 (uplift 5.3%), ci95 [0.906346, 1.065434], p 0.62503, s_y 0.096718 -> MDE at n=8 11.83%; ratio CV 10.43% against the sealed cap 8.72%; engine hit rate with APC on 0.038505905648019395 (ceiling 0.169), APC off hits 0; harness-bound arm-knees 0 of 16. S47b (A/A, n=8): median 1.024752, ci95 [0.95504, 1.056636], p 0.837719, s_y 0.060461 -> MDE 7.24%; harness-bound arm-knees 2 of 16 so the A/A ratio is NOT quotable by the harness rule (clean, but two knees were client-lag-bound). The +7.9% seen at n=1 (S46b) is not resolved at n=8: the effect is smaller than the rig can see at this s_y, and the sealed CV cap is exceeded. Both outcomes were publishable; this is the negative.",
  "cost_usd": 0.0,
  "sweeps_cost_floor_usd": {
    "s46c": 1.1287,
    "s47b": 2.2262,
    "source": "results/spend_ledger.json"
  },
  "wall_s": 0.6,
  "does_not_prove": [
    "One model (Qwen2.5-1.5B), one GPU class (L40S), one trace, one capacity.",
    "The s_y on the fixed bridge (0.097 for the APC pair, 0.060 for the A/A) is what the next run must be sized on; the S05 plan sized n=8 on S47's 0.047 from the collapsed stream.",
    "Absolutes are steadier than the ratio here (apc_off CV 3.07%); the robust block says to quote absolutes."
  ],
  "generated_utc": "2026-09-03T00:33:55.589140+00:00",
  "git_rev": "093ddaf",
  "git_dirty": true,
  "run_id": "ba95e2f8a51a4b7e",
  "host_kind": "Darwin-arm64",
  "provenance_schema": "gpu-cert-provenance/1"
}