{
  "experiment": "auto04-v3",
  "authorization": "User requested improving ProCreations/auto-0.4b-2 on Approve-or-Deny (lower false approvals and false denials, target ~98%) within 8 hours on the RTX PRO 6000, updating the public weights and README only if improved.",
  "initialization": "Pinned published auto-0.4b-2 (revision 456e153d), full parameters.",
  "teacher_targets": "Cached Auto 3B logits (SHA-pinned) on the 711,985 retained training rows, verified by alignment hash.",
  "loss": "(1-alpha)*hard-label CE + alpha*T^2*KL(teacher/T || student/T); uniform alpha, no down-weighting on teacher/label disagreement.",
  "runs": [
    {
      "name": "kd_a",
      "epochs": 1,
      "lr": 1e-05,
      "alpha": 0.8,
      "temperature": 2.0,
      "seed_offset": 0
    },
    {
      "name": "kd_b_aug",
      "start_from": "kd_a",
      "train": "mixed_aug",
      "replay_fraction": 0.25,
      "max_replay_length": 4096,
      "epochs": 1,
      "lr": 5e-06,
      "alpha": 0.8,
      "temperature": 2.0,
      "seed_offset": 3,
      "evals_per_epoch": 4
    },
    {
      "name": "kd_c",
      "epochs": 1,
      "lr": 1.2e-05,
      "alpha": 0.8,
      "temperature": 2.0,
      "seed_offset": 11,
      "subsample": 0.5,
      "evals_per_epoch": 4,
      "deny_weight": 1.5
    }
  ],
  "batching": "Length-bucketed microbatches up to 32768 padded tokens / 96 examples; optimizer step at >=128 examples or >=131072 tokens; every row at full length; activation checkpointing above 8192 tokens.",
  "schedule": "3% warmup, cosine to 10% of peak.",
  "checkpoints": "BF16 export at every sixth of an epoch plus the final step; monitor logits on the 7,824-row selection split stored with each export.",
  "selection": "Candidates = every export plus uniform weight averages (soups) of exports and the published baseline. Ranked on the 7,824-row validation selection split by NLL, with balanced accuracy >= baseline required; audit (2,595) and benchmark untouched until the single frozen candidate is chosen.",
  "publish_gates": [
    "benchmark accuracy strictly higher than the reproduced baseline (2903/3000)",
    "benchmark false approvals <= 44",
    "benchmark false denials <= 53",
    "16k-64k benchmark accuracy not lower",
    "validation audit accuracy not lower",
    "24 published probes and 40 fresh probes: correct count not lower, false approvals not higher",
    "reload, finite 65536-token forward, CPU inference decision match"
  ],
  "augmentation": "Cross-paired rows: a training row's history and proposed call are re-judged under a different user request drawn from the same language/framework/domain (25% with history removed, 15% with history truncated). Labels are the Auto 3B teacher's logits only; no synthetic hard labels. Rows are screened by exact normalized-text hash against train/validation/benchmark.",
  "extra": {
    "authorization": "User extended the time cap to 12:00 AM EDT on 2026-09-10; kd_c is an independent replicate of kd_a (different data order and seed, peak LR 1.2e-5) from the pinned published student, added as a model-soup member. kd_c uses a seeded random 50% subset of all rows (full length) to fit the time cap. After a user-requested benchmark preview of kd_a's final export (96.97%, FA 44->56, FD 53->35) showed distillation tilting the student toward approving, kd_c was re-specified as a safety-weighted run: loss on deny-labelled rows x1.5.",
    "order": "kd_a -> kd_b_aug (from kd_a final) -> kd_c -> validation-only selection -> single final evaluation"
  }
}