{
 "id": "qwen3.arch_axis",
 "kind": "comparison",
 "unit": "ΔBPB vs faithful baseline (wikitext-2)",
 "entries": [
  {
   "name": "arch (bundle: vr+ln+hg)",
   "value": 0.11792615800601025,
   "ci": [
    0.10045055390862463,
    0.13540176210339586
   ],
   "seeds": [
    0.12333850495449372,
    0.10699158813774812,
    0.12344838092578803
   ],
   "significant": true
  },
  {
   "name": "wsd schedule",
   "value": 0.02445100663943567,
   "ci": [
    -0.017262911764197773,
    0.06616492504306912
   ],
   "seeds": [
    0.010673019714336318,
    0.03023596362253178,
    0.03244403658143846
   ],
   "significant": false
  },
  {
   "name": "z-loss",
   "value": -0.003374986704193006,
   "ci": [
    -0.017536542115922118,
    0.010786568707536106
   ],
   "seeds": [
    -0.004562343878277586,
    -0.006294069745714781,
    0.000731453511412683
   ],
   "significant": false
  }
 ],
 "direction": "higher_is_better",
 "generated_by": "scripts/export_site_results.py",
 "generated_at": "2026-07-07T19:00:23Z",
 "notes": "Phase-1 de-confound: 2000 steps / 131,072,000 tokens per cell, 3 paired seeds, iso-FLOP (ratio 1.00042), text-lm-v2, Welch-t 95% CI. seeds[] are paired per-seed deltas (baseline - treatment). Only arch survives; drivers=['arch']. code_py deltas: arch +0.3049 (sig), wsd +0.0361 (n.s.), zloss -0.0007 (n.s.).",
 "derived_from": [
  "Qwen3-0.6B/experiments/2026-06-18_qwen3-0.6b_imu1-deconfound-p1/verdict.json",
  "Qwen3-0.6B/experiments/2026-06-18_qwen3-0.6b_imu1-deconfound-p1/cohort_bpb.json",
  "Qwen3-0.6B/experiments/2026-06-18_qwen3-0.6b_imu1-deconfound-p1/c5_evidence.json"
 ]
}