{
  "name": "ao.iolens.rl.final.ddp600.s0",
  "kind": "RL (GRPO) AO checkpoint \u2014 activation->text inverter, on-policy",
  "warm_start": "ckpts/ao/distill/final.s0/step105 (iolens.final SFT bullet student)",
  "steps": 600,
  "iter": "iter_000600",
  "reward": "per-bullet JOINT FVE (whitened, unit concept dirs), k_max=4, on-policy T=1",
  "injection": {
    "transform": "unit",
    "alpha": 16000,
    "layers": "20..60 step4",
    "marker": "\u321c/158983"
  },
  "hparams": {
    "lr": 5e-05,
    "kl_beta": 0.02,
    "batch_prompts": 14,
    "group_size": 64,
    "max_new_tokens": 128
  },
  "ar": "/workspace/sc/hf/ckpts/ar/chat/mlayer.lc.s0/ex16014240",
  "eval": {
    "metric": "joint FVE on rl_gate_0 (128)",
    "sft_baseline": 0.12,
    "rl_mean": 0.155,
    "rl_peak": 0.159,
    "gt_text_ceiling": 0.129,
    "note": "RL greedy exceeds literal-source-text k=1 recon; eval set is hard (median GT 0.075). Oracle-16 nnomp shows further policy headroom."
  },
  "degeneracy": "none \u2014 bullet_diversity 0.15->0.34, truncation ->0; plateau is advantage convergence (88% groups saturated), not reward hacking",
  "adapter": "PEFT LoRA (default adapter). reference/ (KL anchor = the SFT warm start) omitted; identical to ckpts/ao/distill/final.s0/step105."
}