{
  "session": "v16 latent-action ablation",
  "date": "2026-09-26",
  "pod": {
    "instance": "gpu-31479ece",
    "name": "alea-latent-v17",
    "gpu": "A100-SXM4-80GB",
    "region": "us-east",
    "rate_usd_hr": 1.25,
    "auto_terminate_at": "2026-09-26T23:33:02Z",
    "note": "3 cheaper offerings failed placement; this was the only placeable capacity"
  },
  "data": {
    "train": "data/latent_action_mix.jsonl (56,000 rows; adj traces + 12 tasksource procedural families + RLCD soft labels + filtered v15 replay; agentic corpus excluded for streaming cost)",
    "dev_eval": "data/latent_action_dev_eval.jsonl (8,000 family-stratified rows)",
    "leak_checks": "exact public-JevBench state overlap: 0; train/dev group overlap: 0",
    "jev_distill": "data/jev_distill_20k.jsonl on HF: 19,115 rows of jev-latest soft distributions on our own states; 885 API errors; teacher disagreed with programmatic argmax on 5,580/22,417 comparable labels (24.9%)"
  },
  "arms": {
    "control_no_latent": {
      "config": "continued SFT from v15, no latent module, aux kept",
      "dev": {"acc": 0.5913, "ece10": 0.0442},
      "jevbench_official_6k": {"easy": 31, "hard": 34, "original": 28, "total": 93}
    },
    "latent_keep_aux": {
      "config": "4-step latent action workspace (8 ops, action supervision 0.5), aux kept",
      "dev": {"acc": 0.6003, "ece10": 0.0379},
      "jevbench_official_6k": {"easy": 30, "hard": 34, "original": 28, "total": 92},
      "action_trace": "learned perfectly or near-perfectly on generator families (adj_ambig exact trace 1.0 over 163 seqs)"
    },
    "latent_reset_aux": {
      "config": "latent + aux banks reinitialized",
      "dev": {"acc": 0.5129, "ece10": 0.0361},
      "jevbench_official_6k": {"easy": 26, "hard": 35, "original": 23, "total": 84},
      "note": "reset destroys aux-stored knowledge; not re-learnable in 1 epoch at lr_aux 5e-6"
    },
    "latent_no_aux": {
      "config": "latent, aux banks removed entirely",
      "dev": {"acc": 0.4754, "ece10": 0.0359},
      "jevbench_official_6k": {"easy": 23, "hard": 39, "original": 21, "total": 83},
      "jevbench_4k": {"easy": 23, "hard": 39, "original": 21},
      "hard_family_detail": "multi_hop 38.9% (v15: 11.1%), temporal_numeric 46.7%, trap 62.5%, judge_hard 47.1%; long_policy 15.8% (v15: 26.3%)",
      "churn_vs_control": "hard tier: +21 gained / -16 lost of 111 items -- broad boundary shift, not uniform gain",
      "memory": "peak 9.9GB vs ~20.3GB with aux -> ngram banks ~half of trainable footprint"
    },
    "soup_keep_noaux": {
      "config": "0.5*latent_no_aux + 0.5*latent_keep_aux shared tensors, aux banks from keep",
      "jevbench_official_6k": {"easy": 32, "hard": 36, "original": 26, "total": 94}
    },
    "latent_no_aux_s68": {
      "config": "seed-68 replicate of latent_no_aux",
      "jevbench_official_6k": {"easy": 25, "hard": 38, "original": 21, "total": 84},
      "note": "replicates seed-67 (39 hard): hard-tier gain over aux arms (34-35) is real, not eval noise; easy/original cost also replicates"
    }
  },
  "baseline_v15_official": {"easy": 32, "hard": 35, "original": 29, "total": 96},
  "interpretation": [
    "Latent action supervision is learnable (trace acc ~1.0) and improves dev acc (+0.9pp) and calibration (ECE 0.038 vs 0.044) at matched steps, but does not transfer to authored JevBench by itself.",
    "Aux ngram banks carry real knowledge (reset collapses broadly) yet appear to anchor surface-pattern matching: removing them shifts hard-tier gains into reasoning families (multi_hop +6, temporal +4) while costing long_policy/tradeoff and collapsing easy/original.",
    "Soup recovers easy fully and keeps part of the hard gain -> the no_aux hard shift is partially weight-space real, not pure eval noise.",
    "No arm beats v15 total (96). Best arm product: soup 94. Hard-tier best: latent_no_aux 38-39 across two seeds (with unacceptable easy/original cost).",
    "seed-68 replicate confirms latent_no_aux hard-tier effect is real: 38 vs 39 across seeds while aux arms sit at 34-35.",
    "Aux banks ~= half of model size: 1.40GB ckpt vs 4.04GB; they are the dominant parameter cost with negative-authored-transfer value."
  ],
  "artifacts_on_hf": [
    "runs/latent_action_v16/{control_no_latent,latent_keep_aux,latent_reset_aux,latent_no_aux}/ (final.pt + logs + evals + manifest)",
    "runs/latent_action_v16/latent_no_aux_s68/ (final.pt + train.log + jev eval)",
    "runs/latent_action_v16/soup_keep_noaux.pt",
    "data/jev_distill_20k.jsonl"
  ]
}