{
  "artifact": "sprint46_report",
  "sprint": "46",
  "claim": "Sprint 46's result stated once, in a declared orientation, with the reciprocal beside it and the pre-registration's form defect recorded.",
  "method": "Reads results/s46_apc_real_trace.json. No re-measurement. The orientation is taken from the artifact's own arm_names, which the harness fills from the config's arm order, rather than inferred from the direction of the effect -- inferring it would be assuming the answer.",
  "source_artifact": "results/s46_apc_real_trace.json",
  "arm_names": [
    "apc_off",
    "apc_on"
  ],
  "ratio_orientation": "apc_off/apc_on",
  "median_ratio_as_reported": 0.766419,
  "reciprocal_ratio": 1.304769,
  "reciprocal_orientation": "apc_on/apc_off",
  "knee_uplift_pct": 30.477,
  "ci95_as_reported": [
    0.699733,
    0.804395
  ],
  "ci95_reciprocal": [
    1.24317,
    1.429117
  ],
  "p_value": 0.000332,
  "n_replicates": 5,
  "s_y": 0.056131,
  "sign_test": {
    "n_positive": 0,
    "n_negative": 5,
    "n_ties": 0,
    "unanimous": true,
    "two_sided_p_exact": 0.0625,
    "e_value": 32.0,
    "anytime_p_value": 0.03125,
    "direction": "apc_off < apc_on",
    "note": "Distribution-free: unaffected by an outlier that would inflate s_y. Weaker than the t-test when the noise is well behaved, stronger when it is not. Report both."
  },
  "absolute_knees": {
    "apc_off": {
      "median": 7.853,
      "min": 7.4156,
      "max": 8.2904,
      "mean": 7.89674,
      "stdev": 0.391222,
      "n": 5
    },
    "apc_on": {
      "median": 10.516637,
      "min": 10.24635,
      "max": 10.786925,
      "mean": 10.516637,
      "stdev": 0.191122,
      "n": 5
    }
  },
  "knee_binding": {
    "n_arm_knees": 10,
    "n_slo_bound": 1,
    "n_harness_bound": 0,
    "ratio_is_quotable": true
  },
  "slo_budget_used_range": [
    0.55222,
    0.81244
  ],
  "absolute_knee_cv_pct": 15.22,
  "mde_table": [
    {
      "n": 3,
      "df": 2,
      "s_y": 0.056131,
      "alpha": 0.05,
      "power": 0.8,
      "t_bracket": 5.3633,
      "mde_log": 0.183214,
      "mde_ratio": 1.201071,
      "mde_pct": 20.107,
      "mde_method": "exact_noncentral_t",
      "mde_is_anti_conservative": false,
      "normal_approx_mde_pct": 18.983
    },
    {
      "n": 5,
      "df": 4,
      "s_y": 0.056131,
      "alpha": 0.05,
      "power": 0.8,
      "t_bracket": 3.7174,
      "mde_log": 0.094412,
      "mde_ratio": 1.099013,
      "mde_pct": 9.901,
      "mde_method": "exact_noncentral_t",
      "mde_is_anti_conservative": false,
      "normal_approx_mde_pct": 9.781
    },
    {
      "n": 10,
      "df": 9,
      "s_y": 0.056131,
      "alpha": 0.05,
      "power": 0.8,
      "t_bracket": 3.1456,
      "mde_log": 0.055907,
      "mde_ratio": 1.057499,
      "mde_pct": 5.75,
      "mde_method": "exact_noncentral_t",
      "mde_is_anti_conservative": false,
      "normal_approx_mde_pct": 5.742
    }
  ],
  "prereg_threshold_form_contradicts_its_own_words": true,
  "prereg_defect": "preregistered_threshold.primary reads 'paired median ratio > 1.0' but names no numerator. The harness reports arms[0]/arms[1]. The words are satisfied; the numeric form is not. Fixed structurally in bench/prereg.py, which now refuses a paired config without ratio_orientation.",
  "verdict": "PASS, and the first positive control in this program's history that works.\n  Prefix caching raises the sustainable-QPS knee on a REAL recorded trace by **30.5%** (apc_on/apc_off = 1.3048x, ci95 [1.243, 1.429], p = 0.000332), with all 5 replicates in the same direction. The bridge carries a cacheable signal to the engine, which is the precondition S50 needed and the thing a null here would have refuted.\n  THE ARTIFACT'S OWN THRESHOLD TEXT IS WRONG IN FORM. It reads 'paired median ratio > 1.0' while the harness reports apc_off/apc_on = 0.766419. The words ('the APC-on knee exceeds the APC-off knee') are satisfied decisively; the numeric form contradicts them. That is a pre-registration defect, the third orientation incident in this program, and bench/prereg.py now refuses any paired config lacking a declared `ratio_orientation` -- it refused S47's config on the first run of that check.\n  Two secondary results matter as much as the headline. **0 of 10 arm-knees are harness-bound** and the SLO budget used at the knee is 0.55-0.81 against a historical 0.035-0.090 -- so this is the first knee in the repo that is genuinely a system measurement. And the **absolute** knee CV is 15.22%, still failing Sprint 2's sealed 10% gate, while the PAIRED ratio on the identical data is significant at p = 0.000332. That is direct empirical support for SPRINTS.md 20.9's proposed R2: gate the paired ratio, not the absolute knee.",
  "does_not_prove": [
    "APC helping says NOTHING about whether OUR policy helps. It establishes that the bridge carries a cacheable signal to the engine, which is the precondition for S50 and nothing more.",
    "One model (Qwen2.5-1.5B), one GPU (L40S), one trace slice of 600 requests from a 12,031-request capture, one SLO. The knee is a property of all four and Sprint 2 measured the ABSOLUTE knee as irreproducible across containers; only the PAIRED ratio is defensible, and only on this configuration.",
    "block_tokens 64 against the trace's declared 512 is an 8x input downscale. The reuse STRUCTURE is exact but the token counts are not the trace's, so the magnitude of the uplift is specific to this downscale.",
    "n=5. The measured s_y of 0.056 puts the MDE at 9.90% at n=5, and the observed effect (30.5%) clears it with margin -- but a 12% effect, which is what S50's gate asks for, would NOT clear it at this n."
  ],
  "cost_usd": 0.0,
  "generated_utc": "2026-08-28T15:51:49.002360+00:00",
  "git_rev": "7b4f3c7",
  "git_dirty": true,
  "run_id": "abbb39df8b024208",
  "host_kind": "Darwin-arm64",
  "provenance_schema": "gpu-cert-provenance/1"
}