{
  "session": "v13",
  "pod": "gpu-ae847611",
  "base": "checkpoints/v12_rl.pt (JevBench 60.4/35.1/48.6, ECE ~0.25)",
  "goal": "easy 80 / hard 60 (aspirational); improve calibration + bench while keeping small model",
  "experiments": {
    "v13_sft": {
      "desc": "85.4k-row recalibration anneal from v12_rl: 45% bench-adjacent replay, rlcd soft-targets, label_smooth 0.10, lr 1e-5, 1 epoch",
      "dev": {"acc": 0.7041, "ece": 0.0529, "noul_brier": 0.0311},
      "jevbench_v142": {"easy": 0.542, "hard": 0.333, "original": 0.444},
      "verdict": "calibration restored but easy eroded further - consistent SFT-dilution pattern"
    },
    "v13_rl": {
      "desc": "GRPO 800 steps from v12a_soup40 on 26k exact-answer weak-family rows, beta 0.12",
      "dev": {"acc": 0.6332, "ece": 0.1563, "noul_brier": 0.1077},
      "jevbench_v142": {"easy": 0.583, "hard": 0.342, "original": 0.472},
      "verdict": "lifted original again but collapsed calibration AND lost general capability; dominated by v12_rl"
    },
    "v12_rl_cal": {
      "desc": "temperature scaling baked into v12_rl heads: choice/score tau=2.45, noul tau=0.95, fitted on v12_dev n=4000 (held-out, no bench fitting)",
      "dev": {"acc": 0.7071, "ece": 0.0595, "noul_brier": 0.0362},
      "jevbench_v142": {"easy": 0.604, "hard": 0.351, "original": 0.486,
        "ece": {"easy": 0.247, "hard": 0.299, "original": 0.281},
        "brier": {"easy": 0.570, "hard": 0.849, "original": 0.783}},
      "verdict": "BEST OVERALL PACKAGE - same acc as v12_rl, calibration recovered from 0.25->0.06 dev"
    },
    "rlsoup_cal": {
      "desc": "3-way soup 0.35 v12_rl_cal + 0.35 v12_anneal + 0.30 v10",
      "jevbench_v142": {"easy": 0.646, "hard": 0.315, "original": 0.444},
      "verdict": "interpolates, does not dominate"
    }
  },
  "v142_reference_reruns": {"v12_rl": "60.4/35.1/48.6", "v10": "66.7/36.0/38.9"},
  "recommendation": {
    "best_overall": "checkpoints/v12_rl_cal.pt",
    "best_easy": "checkpoints/v12a_soup40.pt or v10",
    "note": "80/60 goal unreachable by hyperparameter churn; needs qualitative capability jump (state-tracking family resisted all SFT+RL; hard tier needs genuine multi-step reasoning a single-pass 0.6B may not express)"
  },
  "jevbench_note": "public tiers are frozen v1.2 items (48/111/72); v1.4 sealed decisions not locally evaluable; numbers comparable to prior sessions"
}