{
 "generated": "2026-11-06",
 "author": "technique-designer",
 "status": "PENDING",
 "md": "01_technique_proposals.md",
 "key_finding": {
  "text": "att_col==0 for 100% of samples in 100% of trials of AVGC/DTU/KUL/UHD/USTC; synthetic LODO schedule starts unswapped with first segment >=30 s and gaps U(30,90) s. Stem-family LODO 0.87-0.884 exploits this (abs pos, zero pad, bias, y=(att==0)); Carta Acc 49-51 vs mean-4 58.7.",
  "honest_lodo_baseline": {
   "H": 0.7981,
   "Acc": 73.3,
   "config": "4 dec + 4 clf sw2, a=0.5, p_stay 0.998"
  },
  "external_best": {
   "AASD": {
    "H": 0.6453,
    "Acc": 54.05
   },
   "Carta2026": {
    "H": 0.652,
    "Acc": 60.54
   }
  },
  "implication": "valid LODO H>=0.85 needs ~+5-6 pt frame accuracy in evidence; decision-layer space is closed"
 },
 "acceptance_rule": "paired by seed: AASD and Carta H +>=0.005 each AND honest LODO drop <=0.01; all gates pass; AASD/Carta labels never used for selection; user decides keep/drop",
 "gates": [
  "column_swap |f(x)+f(-x)|<1e-5",
  "position_invariance prepend 60 s -> >=99% decision agreement after 60 s",
  "leakage_test <=0.969 s"
 ],
 "candidates": [
  {
   "rank": 1,
   "id": "T1",
   "name": "LagBank-Sel: latency audit + causal within-trial stimulus lag-bank",
   "track": "evidence/inference-only",
   "hypothesis": "classifier is lag-sensitive (lag jitter +-60 ms cost -0.03 H); AASD aligned by assuming epoch sample 0 = audio onset; a fixed offset would push Acc to chance (AASD acc1 0.50-0.54); hidden test sets use different amplifiers",
   "not_repeat_of": [
    "lag2 jitter aug",
    "A3 label shift",
    "TRF lag-bank Path A"
   ],
   "expected": {
    "AASD_Acc": "+0..+6",
    "Carta_Acc": "+0..+3",
    "LODO_H": "+-0.003"
   },
   "risk": "no offset -> ~0 gain; selection noise (gamma prior limits it)",
   "effort": "low: 0.5 d code, ~3 GPU h",
   "hyperparams": {
    "lags_diag_samples_32Hz": [
     -16,
     16,
     1
    ],
    "lags_select": [
     -16,
     16,
     2
    ],
    "tau_s": [
     30,
     60,
     120
    ],
    "beta": [
     2,
     5,
     10
    ],
    "gamma": [
     0,
     0.05,
     0.2
    ],
    "hmm": [
     [
      0.994,
      0.8
     ],
     [
      0.998,
      1.0
     ]
    ]
   },
   "files": {
    "new": [
     "/root/disk_3tb/eeg-thao/lagbank_sel.py"
    ],
    "reads": [
     "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/models/all5_direct_clf_w10b3sw2P_s{0..3}.pt",
     "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/models/fold_direct_clf__eaP_n5_t60_w10b3_ptSPK[_sN]_aug-sw2_wts__{DS}.pt"
    ],
    "outputs": [
     "/root/disk_3tb/eeg-thao/results/research/lagbank_diag.csv",
     "/root/disk_3tb/eeg-thao/results/research/lagbank_sel.json",
     "/root/disk_3tb/eeg-thao/multi_agent_opt/results_lagbank_sel.md"
    ]
   },
   "steps": [
    "A diag acc1 vs lag per dataset (labels read only)",
    "if AASD/Carta peak != 0 by >2 SE: fix converter offset, rescore baselines",
    "B causal soft lag selection, select on LODO inner"
   ]
  },
  {
   "rank": 2,
   "id": "T2",
   "name": "ESR-stem: exactly odd, position-free, schedule-resynthesised logit stem",
   "track": "second-stage",
   "hypothesis": "stem gain = legit temporal integration + col-0 shortcut; odd net + no pos/zero-pad + resynthesised schedules (random start sign, dwell 0.5*LogU(4,40)+0.5*U(30,90) s, ramps U(0.5,6) s) keeps the legit part",
   "not_repeat_of": [
    "stem_ms96 + fine-tunes",
    "stem-synth-switch",
    "antisym_transformer_experiment (8 s, raw 4 logits, abs pos)",
    "B3 learned rule",
    "zip sym stem"
   ],
   "expected": {
    "LODO_H_honest": "0.82-0.85",
    "AASD_H": "+0.005..+0.02 vs mean-4",
    "Carta_H": "+0.005..+0.02"
   },
   "risk": "may not beat 0.827 symmetrised stem by much; ramp is a model of lag",
   "effort": "medium: 1 d code, ~2 GPU h",
   "hyperparams": {
    "window": [
     96,
     192
    ],
    "d_model": 64,
    "layers": 2,
    "lr": 0.0008,
    "steps": 1500,
    "batch": 512,
    "ema": 0.995,
    "switch_weight": 2.5,
    "tv": 0.05,
    "K_resynth": 8,
    "hmm_grid": {
     "p_stay": [
      0.95,
      0.99,
      0.994,
      0.998
     ],
     "scale": [
      0.8,
      1.0
     ]
    },
    "selector": "acc74 inner LODO"
   },
   "files": {
    "new": [
     "/root/disk_3tb/eeg-thao/stem_esr.py"
    ],
    "template": "/root/disk_3tb/eeg-thao/stem_ms96_group_dro.py",
    "reuses": [
     "/root/disk_3tb/eeg-thao/ensemble_clf_transformer_v2.py (LogitTransformer, DilatedResidual, switch_weights)",
     "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/hmm_learned.py (ema_features)"
    ],
    "outputs": [
     "/root/disk_3tb/eeg-thao/results/research/stem_esr.json",
     "/root/disk_3tb/eeg-thao/models/stem_esr/",
     "/root/disk_3tb/eeg-thao/multi_agent_opt/results_stem_esr.md"
    ]
   },
   "ablations": [
    "odd only, original schedule",
    "ramp r=0",
    "short dwell only"
   ]
  },
  {
   "rank": 3,
   "id": "T3",
   "name": "CF-Null: counterfactual-null regulariser + causal null-debias",
   "track": "evidence/train+inference",
   "hypothesis": "antisymmetry does not remove stimulus-only (acoustic) evidence; its variance differs on Mandarin/babble external audio; existing shift_eeg control measured accuracy, not null-logit variance",
   "not_repeat_of": [
    "contrast-aad InfoNCE",
    "aux-env",
    "DANN",
    "noise aug FT",
    "shift_eeg control"
   ],
   "expected": {
    "AASD_Acc": "+0.5..+2",
    "Carta_Acc": "+0.5..+2",
    "LODO_H": "+-0.005"
   },
   "risk": "lambda too high -> under-confident logits; skip if diag rho<0.25",
   "effort": "medium: diag 1 GPU h, FT ~8 GPU h",
   "hyperparams": {
    "null_lag_s": 20,
    "lambda": [
     0.1,
     0.3
    ],
    "ft_steps": "4x250",
    "debias_tau_s": 60,
    "gate_rho": 0.25
   },
   "files": {
    "new": [
     "/root/disk_3tb/eeg-thao/cfnull_diag.py",
     "/root/disk_3tb/eeg-thao/cfnull_debias.py"
    ],
    "edit": [
     "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/direct_classifier.py fit(): flag clf_null_lambda (default 0)"
    ],
    "outputs": [
     "/root/disk_3tb/eeg-thao/multi_agent_opt/results_cfnull.md"
    ]
   }
  },
  {
   "rank": 4,
   "id": "T4",
   "name": "Lapse-aware label cleaning from offline bidirectional OOF teacher",
   "track": "evidence/train",
   "hypothesis": "nominal labels include attention lapses; forward-backward HMM on OOF logits flags lapses; reweight windows",
   "not_repeat_of": [
    "C3 lapse aug",
    "KD ens_all"
   ],
   "expected": {
    "LODO_Acc": "+0.5..+1.5",
    "real_Acc": "+0..+1"
   },
   "risk": "teacher self-confirmation; nested OOF bookkeeping",
   "effort": "medium-high",
   "hyperparams": {
    "posterior_thr": [
     0.2,
     0.3
    ],
    "min_lapse_s": 5,
    "weight_floor": 0.1,
    "seeds_first": 2
   }
  },
  {
   "rank": 5,
   "id": "T5",
   "name": "In-trial distractor ranking with past-envelope reference",
   "track": "evidence/inference-only",
   "hypothesis": "shared reference reduces degenerate pairwise comparisons when one stream is quiet",
   "expected": {
    "real_Acc": "+0..+1"
   },
   "risk": "net not trained on real-vs-distractor",
   "effort": "low",
   "hyperparams": {
    "K": 4,
    "min_shift_s": 8
   }
  },
  {
   "rank": 6,
   "id": "T6",
   "name": "Lookahead reallocation: classifier-only stack at 0.96 s",
   "track": "evidence/retrain",
   "hypothesis": "external leaders are classifier-only at 0.75 s; extra 0.21 s EEG covers neural response to last stimulus samples",
   "expected": {
    "H_both": "+0.003..+0.01"
   },
   "risk": "small effect",
   "effort": "medium",
   "hyperparams": {
    "causal_future_s": 0.96,
    "seeds": 4
   }
  },
  {
   "rank": 7,
   "id": "T7",
   "name": "Weight soup across seeds x folds (SPK-initialised classifiers)",
   "track": "evidence/inference-only",
   "hypothesis": "shared init -> mode connected; soups generalise OOD better than logit ensembles",
   "expected": {
    "real_Acc": "+0..+1",
    "LODO_H": "+-0.005"
   },
   "risk": "soup collapses if not mode-connected",
   "effort": "low",
   "hyperparams": {
    "lodo": "4 held-X seeds",
    "external": "24 ckpts uniform then greedy by LODO inner H"
   }
  }
 ],
 "top3": [
  "T1",
  "T2",
  "T3"
 ]
}