{
 "id": "qwen3.grpo_null",
 "kind": "comparison",
 "unit": "pass@1, gsm8k (n=100 items, T=0.8, 8 samples)",
 "entries": [
  {
   "name": "SFT floor",
   "value": 0.0112,
   "ci": [
    0.0059,
    0.0212
   ]
  },
  {
   "name": "GRPO",
   "value": 0.01,
   "ci": [
    0.0051,
    0.0196
   ]
  },
  {
   "name": "random-reward gate",
   "value": 0.01
  },
  {
   "name": "RFT (iso-generation control)",
   "value": 0.015
  }
 ],
 "direction": "higher_is_better",
 "generated_by": "scripts/export_site_results.py",
 "generated_at": "2026-07-07T19:00:23Z",
 "notes": "Pre-registered null, confirmed: GRPO beats neither the SFT floor nor the random-reward gate (both false on gsm8k AND math500_l13). pass@8 gsm8k: sft 0.07, grpo 0.07, random 0.06, rft 0.11; math500_l13 pass@1: sft 0.015, grpo 0.0225, random 0.015, rft 0.0175. Verdict 'directional' (n=1 seed, capped by pre-registration); proceed_to_phase3=False. At ~1% reward the learning signal is too sparse at 0.6B. RFT sits nominally above the floor but CIs overlap (n.s.).",
 "derived_from": [
  "Qwen3-0.6B/experiments/2026-07-02_qwen3-0.6b_grpo-phase2/verdict.json",
  "Qwen3-0.6B/experiments/2026-07-02_qwen3-0.6b_grpo-phase2/passk_grpo.json",
  "Qwen3-0.6B/experiments/2026-07-01_qwen3-0.6b_rlvr-phase1-passk/phase1_passk.json"
 ]
}