{
 "generated": "2026-11-06",
 "task": "EEG-sAAD auditory attention decoding: inventory of applied methods (blacklist)",
 "best_metrics": {
  "conventions": {
   "H": "official harmonic mean 2*Acc*L/(Acc+L), fraction (0.85 = 85)",
   "Acc": "subject-mean 128 Hz sample accuracy after HMM, PERCENT in project logs",
   "L": "mean over subjects of median-over-switches exp(-SDT/60)",
   "scorer": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/reports/eda_full/scripts/scorer.py (fast_score_subject, aggregate); LODO wrapper /root/disk_3tb/eeg-thao/score.py; external wrapper eval_external.score",
   "protocols": {
    "LODO_synthetic": "train on 4 of {AVGC,DTU,KUL,UHD,USTC}, hold out the 5th, score with SYNTHETIC switch schedule (e1_switch.synth_trial); mean of the 5 held-out folds = 'train-5-datasets' headline",
    "all5_external": "train on all 5 (+Fuglsang, +SparrKULee init for the classifier family), run causally on AASD and Carta2026 REAL switches; HMM emissions from the matching LODO folds only; AASD/Carta labels never used for fitting",
    "caveat": "No file named 'train-together' exists. 'Train 5 datasets together' = the all5 checkpoints (models/*/all5.pt, eval_external_clf.py trains on all 5) used only for AASD/Carta; LODO is the only held-out-by-dataset metric."
   },
   "targets_in_ledger": {
    "train_mean": "H>=0.85 and Acc>=78 (met only by stem family, see integrity flag)",
    "each_train_dataset": "H>=0.85 and Acc>=78 on each of 5 (AVGC open)",
    "external_reading1": "AASD and Carta each H>0.63 and Acc>60",
    "external_reading2": "AASD >63 and Carta >60 on both H and Acc"
   }
  },
  "lodo_5_dataset_synthetic": {
   "integrity_flag": "All stem_ms96-family numbers (H 0.86-0.884) come from a non-antisymmetric second-stage net (target y=(att==0), zero-init bias head, zero-padded window start, absolute position embedding). The sibling repo audit of the identical zip net found a column-0/start-of-trial prior: LODO 0.863 -> 0.827 when symmetrised, and AASD 0.613-0.618 / Carta 0.546-0.561 on real switches (below mean-of-4-logits). No column-swap / symmetrisation / position-invariance test was ever run on the workspace stem variants (grep: none). Treat 0.87-0.884 as UNVERIFIED upper bounds, not as honest best.",
   "best_numeric_tip_unverified": {
    "name": "focal-IRM <-> stem-synth-switch fuse w=0.5 + held_T_nll_med (causal_ds_temp.py)",
    "mean_H": 0.88438,
    "mean_Acc": 83.378,
    "per_dataset": {
     "AVGC": {
      "H": 0.8245,
      "Acc": 75.64,
      "SDT": 7.07
     },
     "DTU": {
      "H": 0.9221,
      "Acc": 90.33,
      "SDT": 2.01
     },
     "KUL": {
      "H": 0.8825,
      "Acc": 82.94,
      "SDT": 3.28
     },
     "UHD": {
      "H": 0.863,
      "Acc": 80.82,
      "SDT": 1.25
     },
     "USTC": {
      "H": 0.9298,
      "Acc": 87.15,
      "SDT": 0.19
     }
    },
    "hmm": {
     "p_stay": 0.994,
     "scale": 0.8
    },
    "source": "/root/disk_3tb/eeg-thao/results/research/causal_ds_temp.json (variants.held_T_nll_med), /root/disk_3tb/eeg-thao/multi_agent_opt/results_causal_ds_temp.md"
   },
   "nested_decision_layer_unverified": {
    "name": "causal_eeg_nested_lodo.py (joint p_stay/scale/tau/threshold by inner LODO, objective 0.5H+0.5Acc/100) on the same frozen fused evidence",
    "mean_H": 0.8834,
    "mean_Acc": 83.45,
    "mean_L": 0.939,
    "mean_SDT_s": 2.81,
    "per_dataset": {
     "AVGC": {
      "H": 0.82,
      "Acc": 75.75
     },
     "DTU": {
      "H": 0.9225,
      "Acc": 90.38
     },
     "KUL": {
      "H": 0.8828,
      "Acc": 83.08
     },
     "UHD": {
      "H": 0.8621,
      "Acc": 80.92
     },
     "USTC": {
      "H": 0.9297,
      "Acc": 87.14
     }
    },
    "tau_selected": 0.0,
    "threshold_selected": 0.65,
    "frozen_external_cfg": {
     "emission_scale": 0.65,
     "p_stay": 0.985,
     "tau_s": 0.0,
     "threshold": 0.65
    },
    "source": "/root/disk_3tb/eeg-thao/results/research/causal_eeg_nested_lodo.md/.json (finished 2026-11-06 08:07 UTC, elapsed 98376 s)"
   },
   "stem_family_other": {
    "stem_ms96_nested_published": {
     "mean_H": 0.8677,
     "mean_Acc": 82.42,
     "AVGC": [
      0.7945,
      74.06
     ],
     "DTU": [
      0.9061,
      89.85
     ],
     "KUL": [
      0.8755,
      82.33
     ],
     "UHD": [
      0.8407,
      79.47
     ],
     "USTC": [
      0.9219,
      86.36
     ]
    },
    "IRM_SWA_fuse_w0.5": {
     "mean_H": 0.8725,
     "mean_Acc": 82.58,
     "AVGC": [
      0.81,
      75.14
     ],
     "UHD": [
      0.858,
      null
     ]
    },
    "IRM_synth_fuse_w0.5": {
     "mean_H": 0.879,
     "mean_Acc": 83.11,
     "AVGC": [
      0.8192,
      75.52
     ]
    },
    "focal_synth_fuse_w0.5_raw": {
     "mean_H": 0.8821,
     "mean_Acc": 83.36,
     "AVGC": [
      0.8226,
      75.54
     ]
    },
    "group_dro_fuse_w0.5": {
     "mean_H": 0.8743,
     "mean_Acc": 82.73,
     "AVGC": [
      0.8218,
      75.58
     ],
     "pure_group_dro": [
      0.8609,
      81.54
     ]
    }
   },
   "best_honest_classifier_family": {
    "name": "4 decoders (vlaai_ws 5 s, 4 blocks) + 4 direct classifiers (vlaai_ws 10 s, 3 blocks, sw2 aug, SparrKULee init, +Fuglsang), prep eaP_n5_t60, LDA evidence weights, classifier weight a=0.5, DEC_OFFSET=3, 2-state HMM p_stay 0.998 (antisymmetric by construction)",
    "mean_H": 0.7981,
    "mean_Acc": 73.3,
    "per_dataset_H": {
     "AVGC": 0.726,
     "DTU": 0.842,
     "KUL": 0.809,
     "UHD": 0.791,
     "USTC": 0.823
    },
    "variants": {
     "p_stay_0.9997_a0.7": 0.8003,
     "classifier_only_4seed": 0.7951,
     "single_seed_range": [
      0.7761,
      0.7944
     ],
     "schedule_noise_std": 0.0015,
     "fuse_general_a_grid_best_a0.75_H_Acc": [
      0.7972,
      72.8
     ]
    },
    "source": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/README_BEST.md sec 9; /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/results/research/baseline_recipe_4plus4.csv (sw2 all4 synthetic mean H 0.7981 Acc 73.30)"
   },
   "other_classifier_tracks": {
    "recipe_aug_4plus4_all4": {
     "synthetic_H": 0.7937,
     "synthetic_Acc": 73.04,
     "real_mean_H": 0.6393,
     "real_mean_Acc": 55.91,
     "note": "sw2 reference in same file: 0.7981/73.30 and real 0.6384/56.02 -> no gain"
    },
    "bmms3enh_single_clf_seed0": {
     "synthetic_mean_H_AVGC_DTU_KUL_UHD_USTC": [
      0.7329,
      0.818,
      0.7794,
      0.7604,
      0.7982
     ],
     "vs_direct_sw2": "worse on DTU/KUL/UHD, +0.01 AVGC; real AASD 0.6363/52.25, Carta 0.6447/60.11"
    },
    "rate_transformer_75_25_mix_mean4": {
     "mean_H_over_5": 0.7535,
     "mean_Acc": 68.57,
     "note": "recipe_transformer_ensemble.csv weight 0.25, p0.998; below baseline"
    }
   }
  },
  "external_real_switch": {
   "note": "AASD 18 subj/900 trials/4600 switches (key-press); Carta2026 24 subj/480 trials/2880 switches (cued). ~52-54 % AASD frame accuracy is the bottleneck; acc1 0.50-0.54.",
   "best_H_and_Acc_pre_specified_operating_point": {
    "linear_hp30 (causal tau30 s one-pole high-pass of 4 classifier seed logits, then mean)": {
     "AASD": {
      "H": 0.6422,
      "Acc": 53.12,
      "L": 0.812,
      "SDT": 13.0
     },
     "Carta2026": {
      "H": 0.652,
      "Acc": 60.54,
      "SDT": 19.5
     },
     "hmm": "p_stay 0.994 scale 0.8",
     "train_LODO": [
      0.7899,
      69.21
     ],
     "src": "/root/disk_3tb/eeg-thao/results/research/ensemble_clf_transformer_v4_external.csv"
    },
    "tip_fuse_a0.50@SW2emis (speech-xattn<->hard_abs_swmix, best Acc)": {
     "AASD": {
      "H": 0.6453,
      "Acc": 54.05
     },
     "Carta2026": {
      "H": 0.6488,
      "Acc": 60.6
     },
     "hmm": "0.994/0.8 under published SW2 emissions",
     "src": "/root/disk_3tb/eeg-thao/multi_agent_opt/results_xattn_sw2_emis_fuse.md"
    },
    "sw2_multi_tau soft_hp50": {
     "AASD": {
      "H": 0.6447,
      "Acc": 53.51
     },
     "Carta2026": {
      "H": 0.6498,
      "Acc": 60.17
     },
     "src": "/root/disk_3tb/eeg-thao/multi_agent_opt/results_sw2_multi_tau.md"
    },
    "mean4 logits (published best classifier family, p 0.998)": {
     "AASD": {
      "H": 0.6329,
      "Acc": 52.11
     },
     "Carta2026": {
      "H": 0.6427,
      "Acc": 58.73
     },
     "src": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/results/research/stem_external.csv variant mean4"
    }
   },
   "best_AASD_H_on_disk_illegal_test_label_sweep": {
    "H": 0.6525,
    "Acc": 54.03,
    "p_stay": 0.95,
    "src": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/results/research/fuse_external.csv name RB",
    "note": "tuned on AASD labels; do not use as headline"
   },
   "kept_decision_options_not_chosen": {
    "B1_semi_Markov_m5_k1.5": {
     "AASD_H": 0.654,
     "Carta_H": 0.651,
     "LODO_synth": 0.782
    },
    "B3_learned_rule_V2": {
     "AASD_H": 0.6371,
     "Carta_H": 0.586,
     "LODO_synth": 0.8109
    }
   },
   "stem_family_on_real_switches_BAD": {
    "causal_eeg_nested_lodo_frozen_cfg": {
     "AASD": {
      "H": 0.6224,
      "Acc": 51.18,
      "L": 0.794,
      "SDT": 13.0
     },
     "Carta2026": {
      "H": 0.5556,
      "Acc": 49.51,
      "L": 0.633,
      "SDT": 26.94
     }
    },
    "held_T_nll_med_tip": {
     "AASD": {
      "H": 0.6227,
      "Acc": 51.17
     },
     "Carta2026": {
      "H": 0.5534,
      "Acc": 49.41
     }
    },
    "stem_ms96_published": {
     "AASD": {
      "H": 0.6158,
      "Acc": 51.54
     },
     "Carta2026": {
      "H": 0.5695,
      "Acc": 51.03
     }
    },
    "group_dro_pure": {
     "AASD": {
      "H": 0.6217,
      "Acc": 51.23
     },
     "Carta2026": {
      "H": 0.5459,
      "Acc": 48.25
     }
    }
   },
   "targets_status": {
    "AASD_H>0.63": "met (classifier family)",
    "AASD_Acc>60": "NOT met (max 54.05)",
    "Carta_H>0.63_and_Acc>60": "met only by linear_hp30 / xattn tip (H 0.652, Acc 60.5-60.6)",
    "no_single_model_meets_train_mean_and_both_external": true
   },
   "hmm_caution": "p_stay 0.9997 is worst on real switches; p_stay 0.95-0.998 flat; do not retune p_stay on AASD/Carta labels."
  }
 },
 "forbidden_methods": [
  {
   "id": "PRE-01",
   "family": "preprocessing",
   "method": "organiser pre_*.mat 1-9 Hz Butterworth-2 filtfilt, 58-ch union (zeros for missing), CAR, drop 1 s ends, one full-trial z-score",
   "status": "in_stack",
   "evidence": "base of every model; do not re-derive",
   "source": "config.yaml preps.mat"
  },
  {
   "id": "PRE-02",
   "family": "preprocessing",
   "method": "eaP_n5_t60: causal Euclidean alignment restarted each trial, Fuglsang population prior weight 5 s, tau 60 s, shrink 0.1",
   "status": "in_stack",
   "evidence": "current legal prep",
   "source": "prepare_data.py, build_ea_prior.py"
  },
  {
   "id": "PRE-10",
   "family": "preprocessing",
   "method": "causal running z-score (_rz)",
   "status": "closed_negative",
   "evidence": "0.694 vs 0.715 mat",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-11",
   "family": "preprocessing",
   "method": "FFT whole-trial bandpass 1-9 / 1-15 Hz",
   "status": "closed_negative",
   "evidence": "0.700-0.710; illegal at test",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-12",
   "family": "preprocessing",
   "method": "3-band causal spectral equalisation (beq)",
   "status": "closed_negative",
   "evidence": "0.673-0.694",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-13",
   "family": "preprocessing",
   "method": "gamma envelope channels 30-40 / 35-48 Hz",
   "status": "closed_negative",
   "evidence": "0.727-0.729 or chance DTU/KUL",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-14",
   "family": "preprocessing",
   "method": "delta/theta phase channels (FIR-Hilbert)",
   "status": "closed_negative",
   "evidence": "0.692",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-15",
   "family": "preprocessing",
   "method": "causal AR(6) temporal whitening tw / tw05 / tws recolour",
   "status": "closed_negative",
   "evidence": "0.751-0.755 vs 0.754",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-16",
   "family": "preprocessing",
   "method": "EA grid: forget 30/120 s, n0 5/15/30/60 s, shrink 0.3, identity init, mat_reap",
   "status": "closed_negative",
   "evidence": "tune_ea.py 16 configs, no gain",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-17",
   "family": "preprocessing",
   "method": "mat_reac (EA state carried across trials)",
   "status": "closed_negative",
   "evidence": "+0.007..0.010 but ILLEGAL (rule 6.3)",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-18",
   "family": "preprocessing",
   "method": "nearest-neighbour surface Laplacian lap_eaP",
   "status": "closed_negative",
   "evidence": "train H 0.771 Acc 70.74; AASD 0.629/51.72; Carta 0.634/57.78",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-19",
   "family": "preprocessing",
   "method": "causal 7 Hz low-pass lp7_eaP",
   "status": "closed_negative",
   "evidence": "train H 0.788 Acc 72.48; AASD 0.624/51.11; Carta 0.634/58.02 (~= eaP 0.785)",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-20",
   "family": "preprocessing",
   "method": "spherical-spline CSD Laplacian sph_lap_eaP -> vlaai_ws",
   "status": "closed_negative",
   "evidence": "chance at 50 steps (loss 0.7122)",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-21",
   "family": "preprocessing",
   "method": "nearest-neighbour bipolar montage bipolar_eaP -> vlaai_ws",
   "status": "closed_negative",
   "evidence": "chance (loss 0.6886)",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-22",
   "family": "preprocessing",
   "method": "causal Hjorth activity/mobility/complexity C=232 stack",
   "status": "closed_negative",
   "evidence": "chance (loss 0.6932)",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-23",
   "family": "preprocessing",
   "method": "glim/mask/soft mask/orel/odiff/ediff/delta-theta speech features",
   "status": "closed_negative",
   "evidence": "no gain; env+onset+dom stay",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "PRE-24",
   "family": "preprocessing",
   "method": "preprocessing-agent real-switch event-lock + causal posterior EMA (tau grid 0..8 s)",
   "status": "closed_negative",
   "evidence": "tau=0 selected on all folds; no H/Acc gain",
   "source": "ALREADY_TRIED.md C2; multi_agent_opt/results_preprocessing.md"
  },
  {
   "id": "ARC-01",
   "family": "architecture",
   "method": "vlaai_ws 2/4/6 blocks, 32/128 filters, 3 s/5 s/20 s/15 s windows",
   "status": "closed_negative",
   "evidence": "2 blk 0.775; 4 blk 0.766; 15 s alone LODO 0.7693/71.82; AASD 0.6290/51.48",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-02",
   "family": "architecture",
   "method": "deformer, deformer_ws, lsm_cnn, ssmamba, simple_cnn, original VLAAI, CCA/FBCCA/WF/Riemann/fwd-WF, ADT, HappyQuokka decoders",
   "status": "closed_negative",
   "evidence": "0.50-0.756 < vlaai_ws",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-03",
   "family": "architecture",
   "method": "FBCSP / Riemann left-right spatial decode",
   "status": "closed_negative",
   "evidence": "49-53 % held-out",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-04",
   "family": "architecture",
   "method": "ATCNet 10 s direct classifier",
   "status": "closed_negative",
   "evidence": "LODO 0.7425/68.06; AASD 0.6290/52.05",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-05",
   "family": "architecture",
   "method": "EEG Conformer 10 s direct classifier",
   "status": "closed_negative",
   "evidence": "LODO 0.7471/67.93; AASD 0.6398/52.65",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-06",
   "family": "architecture",
   "method": "ShallowConvNet 10 s",
   "status": "closed_negative",
   "evidence": "LODO 0.7537/68.64; AASD 0.6365/52.39",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-07",
   "family": "architecture",
   "method": "gru_ffn, conv_attn, env_xattn direct-classifier backbones (w10, random init)",
   "status": "closed_negative",
   "evidence": "below vlaai_ws (results/direct_clf__eaP_n5_t60_w10_{gru_ffn,conv_attn,env_xattn}_s0.json)",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-08",
   "family": "architecture",
   "method": "EEG-TCNet, DeepConvNet, WaveNet, Mamba DirectNet, Transformer DirectNet (TFDirect), TSception, FBCNet-SPK, SpecViT(Mel-STFT), xcorr-matched, EEGNet DWS, C=232 filterbank-only",
   "status": "closed_negative",
   "evidence": "all abandoned at AVGC chance probe (loss >=0.685 after 50 steps); no LODO/external",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-09",
   "family": "architecture",
   "method": "channel-mask + XY FiLM, learned channel-adjacency (graph) mixer, SE channel-gate, MixStyle, multi-rate envelope bank",
   "status": "closed_negative",
   "evidence": "LODO 0.7669-0.7815; AASD 0.6307-0.6405; none beat linear_hp30",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-10",
   "family": "architecture",
   "method": "in-net FIR filterbank residual, complex-Morlet residual, FBCNet parallel bands -> vlaai_ws",
   "status": "closed_negative",
   "evidence": "LODO 0.7569-0.7639; AASD 0.6372-0.6405 (FBCNet: chance)",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-11",
   "family": "architecture",
   "method": "zero-init residual heads on SPK vlaai_ws: speech-FiLM (speech_resid), temporal MHSA (temp_attn), speech cross-attn (speech_xattn), aux-env MSE, contrastive InfoNCE, TRF Path B, frozen-trunk GRU (sw2_feat_gru)",
   "status": "closed_negative",
   "evidence": "LODO 0.7629-0.7848; AASD 0.6343-0.6423 (xattn feeds Acc tip 54.05 only through fuse)",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-12",
   "family": "architecture",
   "method": "kd_ensemble_front: KD of stem ens_all into vlaai_ws 10 s",
   "status": "closed_negative",
   "evidence": "LODO 0.7595/68.97; AASD 0.6408/52.71",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-13",
   "family": "architecture",
   "method": "BMMSNet-style bidirectional multilevel correspondence classifier (bmms_env, bmms_enhanced, bmms_3feature), 2-4 seeds",
   "status": "closed_negative",
   "evidence": "synthetic <= direct sw2; real AASD 0.636/52.3, Carta 0.645-0.659/60.1-60.6 ~ equal; mixing weight 0.1/0.25 lowers H; DECOMMISSIONED as gain",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-14",
   "family": "architecture",
   "method": "Foundation backbones/embeddings: CBraMod (linear read-out, LoRA D2, scratch D3, LP-FT, stimulus-contrastive pretrain H1), BrainOmni, BIOT, LaBraM, EEGPT; fm_front H5",
   "status": "closed_negative",
   "evidence": "d' 0.09-0.12 LODO, 0.006 AASD; fine-tune from step 0 chance; LP-FT ~= orgbase at 15x cost; H1 stopped; D2/D3 held",
   "source": "ALREADY_TRIED.md C1/E; decoder_lodo git log; CONTEXT_MEMORY_real-switch-findings.md"
  },
  {
   "id": "ARC-20",
   "family": "architecture",
   "method": "vlaai_ws shared-weight conv, 3 blocks x 64 filters, 10 s window, antisymmetric g(E,A,B)-g(E,B,A), SparrKULee match/mismatch init, +Fuglsang",
   "status": "in_stack",
   "evidence": "published classifier: single seed 0.776-0.794; keep as-is",
   "source": "direct_classifier.py"
  },
  {
   "id": "ARC-21",
   "family": "architecture",
   "method": "vlaai_ws 5 s 4-block reconstruction decoder env+onset+dom, Pearson loss, cluster aug, SparrKULee init",
   "status": "in_stack",
   "evidence": "single seed 0.759-0.766",
   "source": "run_experiment.py --decoder vlaai_ws"
  },
  {
   "id": "TRN-01",
   "family": "training",
   "method": "contrastive Pearson loss (in decoder) / BCE on antisymmetric logit (in classifier)",
   "status": "in_stack",
   "evidence": "in stack; no re-tune",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "TRN-02",
   "family": "training",
   "method": "end-to-end CRF/sequence loss 40 s with fake switches; learned fusion weights by sequence loss",
   "status": "closed_negative",
   "evidence": "0.757-0.762 / tied",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "TRN-03",
   "family": "training",
   "method": "dataset- and subject-adversarial (GRL) loss on decoder/classifier; DANN on stem features",
   "status": "closed_negative",
   "evidence": "-0.004 / tip unchanged (0.8620 pure)",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "TRN-04",
   "family": "training",
   "method": "label smoothing, mixup (feature/label), group-DRO (v3 gate_robust), CORAL (train-time), TV smoothness, focal BCE, soft-H surrogate, AVGC loss upweight 2.5x, AVGC hard curriculum, PCGrad, IRMv1, SAM, RSC, SWA/SWA2k, layer-wise LR, stochastic depth",
   "status": "closed_negative",
   "evidence": "each <= 0.0035 H on AVGC tip or negative; AVGC never reached 0.85/78 (best 0.8245)",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "TRN-05",
   "family": "training",
   "method": "Group-DRO / CVaR worst-dataset stem fine-tune (stem_ms96_group_dro.py, 800 steps from IRM)",
   "status": "closed_negative",
   "evidence": "pure LODO 0.8609/81.54; fuse 0.8743/82.73; AVGC 0.8218/75.58 (tip -0.0027); AASD 0.6217/51.23; Carta 0.5459/48.25",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "TRN-06",
   "family": "training",
   "method": "SSL pretrain on stem windows: VICReg, Barlow, temporal CPC",
   "status": "closed_negative",
   "evidence": "pure 0.8556-0.8561; fuse tip -0.015..-0.020",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "TRN-07",
   "family": "training",
   "method": "augment: noise02, pclus05 (50 % cluster drop), covmix, spectral tilt, lag jitter +-60 ms (-0.03), featdrop, sw2+featdrop",
   "status": "closed_negative",
   "evidence": "none helps; lag2 -0.03",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "TRN-08",
   "family": "training",
   "method": "sw2 switch aug (p_switch 0.5, second switch 0.5)",
   "status": "in_stack",
   "evidence": "in_stack: +0.001..0.008; fusion 0.798",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "TRN-09",
   "family": "training",
   "method": "denser/longer/shorter switch schedules FT (hardswitch p_switch 0.75/0.6, switch-dur lognormal, shorthold, decoder-sw3 seg_switch 0.75, stem synth_switch 0.7, cross-fade C1, label delay C2, lapse C3, babble C4, montage C5, time-warp C6, straddle-real C7, recipe variable-rate/rate-aux/masked-rate/global-gain/time-mask)",
   "status": "closed_negative",
   "evidence": "AASD +-0.002..0.013, Carta mostly down, LODO synthetic -0.02 or flat; clf noise/SpecAug FT flat; ensemble 4+4 no better than sw2",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "TRN-10",
   "family": "training",
   "method": "hyper-parameter search lr/wd/batch/dropout/noise/switch-rate on 10 s classifier",
   "status": "closed_negative",
   "evidence": "-0.002..-0.011; published point AdamW 5e-4 wd 5e-4 batch 64 is best",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "TRN-11",
   "family": "training",
   "method": "SparrKULee mixed 25 % into batches; Broderick2018; extra SparrKULee pretrain schedules",
   "status": "closed_negative",
   "evidence": "no gain / 0.740 hurts KUL-AV-GC",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "TRN-12",
   "family": "training",
   "method": "AASD or Carta added to training (decoder, classifier A1-A5, LODO xreal)",
   "status": "closed_negative",
   "evidence": "LODO 0.734-0.742 or synthetic -0.015..-0.018; real not better than baseline. AASD/Carta are EVALUATION sets",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "TRN-13",
   "family": "training",
   "method": "within-trial self-training on pseudo-labels",
   "status": "closed_negative",
   "evidence": "-0.018..-0.001",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "TRN-14",
   "family": "training",
   "method": "learning-curve subsampling 25-100 %",
   "status": "closed_negative",
   "evidence": "0.719..0.742: saturated",
   "source": "ALREADY_TRIED.md C3-C5/E; REAL_SWITCH_STRATEGY.md sec 8"
  },
  {
   "id": "LOG-01",
   "family": "second_stage",
   "method": "v1 causal logit transformer tf_raw_w16/w48, tf_hp_w32/w48",
   "status": "closed_negative",
   "evidence": "best tf_hp_w48 0.8208/78.09 (acc74); external not scored",
   "source": "ensemble_clf_transformer.py"
  },
  {
   "id": "LOG-02",
   "family": "second_stage",
   "method": "v2 recipes recipe_hp48, ms_w96, ms_w192, stem_ms96 (dilated causal conv stem + 24-d multi-scale features, window 96, d 64, EMA, balanced batches), ens_all",
   "status": "integrity_suspect",
   "evidence": "stem_ms96 LODO 0.8677/82.42 but AASD 0.6158/51.54, Carta 0.5695/51.03; non-antisymmetric, column-0 prior (see integrity_flag)",
   "source": "ensemble_clf_transformer_v2.py"
  },
  {
   "id": "LOG-03",
   "family": "second_stage",
   "method": "stem_ms96_image_split (subject-level split instead of dataset-inner)",
   "status": "closed_no_gain",
   "evidence": "0.8627/82.50",
   "source": "stem_ms96_image_split.py"
  },
  {
   "id": "LOG-04",
   "family": "second_stage",
   "method": "v3 gated local attention (gate_local, gate_robust), ens_gate_local_stem",
   "status": "closed_no_gain",
   "evidence": "0.8631/81.97; ens 0.8676/82.75",
   "source": "ensemble_clf_transformer_v3.py"
  },
  {
   "id": "LOG-05",
   "family": "second_stage",
   "method": "v4 rel-bias local attention rel_ms/rel_hp, linear_hp30 (tau 30 s high-pass)",
   "status": "closed_no_gain",
   "evidence": "rel_ms AASD 0.6214/51.25; linear_hp30 external winner but LODO 0.7899/69.21",
   "source": "ensemble_clf_transformer_v4.py"
  },
  {
   "id": "LOG-10",
   "family": "second_stage",
   "method": "ALiBi on stem",
   "status": "integrity_suspect",
   "evidence": "LODO 0.8320/75.04",
   "source": "multi_agent_opt/results_*.md"
  },
  {
   "id": "LOG-11",
   "family": "second_stage",
   "method": "RoPE on stem",
   "status": "integrity_suspect",
   "evidence": "0.8711/82.68; AVGC 0.8124/75.66",
   "source": "multi_agent_opt/results_*.md"
  },
  {
   "id": "LOG-12",
   "family": "second_stage",
   "method": "stem fine-tunes: SWA, SWA2k, IRM, AVGC-upweight, soft-H, SAM, PCGrad, RSC, DropPath, layer-wise LR, frozen-GRU residual, ms-feature AdaLN/FiLM, denser synth-switch, SW2-as-ms-channel, label-free TTA entropy-min",
   "status": "integrity_suspect",
   "evidence": "0.8590-0.8772 each; AASD 0.617-0.624 flat",
   "source": "multi_agent_opt/results_*.md"
  },
  {
   "id": "LOG-13",
   "family": "second_stage",
   "method": "logit blends IRM<->SWA / RoPE / synth / stochdepth / focal<->synth / mixup<->synth / dann<->synth / vicreg<->synth / cpc<->synth (fixed w 0.5 and global_w)",
   "status": "integrity_suspect",
   "evidence": "tip 0.8821/83.36 at best; later blends lower",
   "source": "multi_agent_opt/results_*.md"
  },
  {
   "id": "LOG-14",
   "family": "second_stage",
   "method": "multi-stem soft-attn pool, evidential/NNLS/logistic stack, agree-gate, conf-resid affine/MLP/GRU on tip fuse",
   "status": "integrity_suspect",
   "evidence": "AVGC tip -0.0005..-0.0169",
   "source": "multi_agent_opt/results_*.md"
  },
  {
   "id": "DEC-01",
   "family": "decision_layer",
   "method": "2-state HMM, Gaussian emissions fit on other datasets, p_stay 0.998 / 0.994 + scale 0.8",
   "status": "in_stack",
   "evidence": "in stack",
   "source": "ALREADY_TRIED.md B/C6/E; REAL_SWITCH_STRATEGY.md"
  },
  {
   "id": "DEC-02",
   "family": "decision_layer",
   "method": "p_stay sweeps 0.9..0.9997, emission scale, tempering (scale>1), Student-t emissions, reliability weights, CUSUM, multi-rate HMM, MSM bank, running z-score of evidence",
   "status": "closed_no_gain",
   "evidence": "<= +0.003; p 0.9997 worst on real",
   "source": "ALREADY_TRIED.md B/C6/E; REAL_SWITCH_STRATEGY.md"
  },
  {
   "id": "DEC-03",
   "family": "decision_layer",
   "method": "semi-Markov HMM (Weibull dwell) B1",
   "status": "kept_open",
   "evidence": "real +0.005..0.016, synthetic -0.002..-0.016; user KEPT, parameters open (do not hard-code m=5 s)",
   "source": "ALREADY_TRIED.md B/C6/E; REAL_SWITCH_STRATEGY.md"
  },
  {
   "id": "DEC-04",
   "family": "decision_layer",
   "method": "BOCPD B2 (hazard grid, bma, learned, robust)",
   "status": "closed_no_gain",
   "evidence": "real +0.003..0.012, synthetic -0.008..-0.011; user STOPPED",
   "source": "ALREADY_TRIED.md B/C6/E; REAL_SWITCH_STRATEGY.md"
  },
  {
   "id": "DEC-05",
   "family": "decision_layer",
   "method": "learned antisymmetric decision rule B3 (V1 real only, V2 real+synth)",
   "status": "kept_open",
   "evidence": "V2 synthetic 0.8109 but Carta 0.5860; user KEPT as research",
   "source": "ALREADY_TRIED.md B/C6/E; REAL_SWITCH_STRATEGY.md"
  },
  {
   "id": "DEC-06",
   "family": "decision_layer",
   "method": "A4 aux 'switch in last 5 s' head + gated HMM",
   "status": "closed_no_gain",
   "evidence": "~= baseline (AASD 0.636, Carta 0.643)",
   "source": "ALREADY_TRIED.md B/C6/E; REAL_SWITCH_STRATEGY.md"
  },
  {
   "id": "DEC-07",
   "family": "decision_layer",
   "method": "hmm posterior EMA (tau 0..8 s), decision threshold, nested joint selection (causal_eeg_nested_lodo)",
   "status": "integrity_suspect",
   "evidence": "tau=0 always; LODO 0.8834 on suspect evidence; external 0.6224/0.5556",
   "source": "ALREADY_TRIED.md B/C6/E; REAL_SWITCH_STRATEGY.md"
  },
  {
   "id": "DEC-08",
   "family": "decision_layer",
   "method": "logit temperature / Platt / isotonic (global, held-only, per-dataset, per-subject), causal-ds-temp held_T_nll_med, adapt-readout, multi-tau HP bank, subject-conditional calibration",
   "status": "closed_no_gain",
   "evidence": "T global is HMM-invariant; held_T tiny +0.0019 AVGC; iso Acc+1 H-0.01",
   "source": "ALREADY_TRIED.md B/C6/E; REAL_SWITCH_STRATEGY.md"
  },
  {
   "id": "DEC-09",
   "family": "decision_layer",
   "method": "MoE/gates fuse<->hp30, domain emission-LL, trial MMD/Mahalanobis, conf-gated cross-stream, seed-disagreement, seed-reliability, power-mean/trim, rank-Gauss, SW2<->hp30/swmix/tri blend, SW2 5th seed, prep-diverse bank, fuse_general a-grid, dec<->clf disagree-gate, tip-sdur gate",
   "status": "closed_no_gain",
   "evidence": "AASD Acc stays 50.3-54.0; selectors collapse to const_*",
   "source": "ALREADY_TRIED.md B/C6/E; REAL_SWITCH_STRATEGY.md"
  },
  {
   "id": "DEC-10",
   "family": "decision_layer",
   "method": "emission-shift / adaptive emission (subj-adapt-emis, proto-mem-emis, online-moment hp30/fuse/SW2, ext-coral logits, ms-feat CORAL, partial CORAL, cross-stream scale)",
   "status": "closed_no_gain",
   "evidence": "AASD Acc <= 53.17; LODO regress",
   "source": "ALREADY_TRIED.md B/C6/E; REAL_SWITCH_STRATEGY.md"
  },
  {
   "id": "DEC-11",
   "family": "decision_layer",
   "method": "4 decoder + 4 classifier fusion LDA weights a=0.5; 5 s / 3 s / 20 s extra classifiers, 2nd decoder family, multi-scale decoder evidence",
   "status": "in_stack",
   "evidence": "in stack / 0..-0.01",
   "source": "ALREADY_TRIED.md B/C6/E; REAL_SWITCH_STRATEGY.md"
  },
  {
   "id": "PTA-01",
   "family": "classical_pathA",
   "method": "TRF lag-bank ridge Path A (broadband, band 1-8, dual 1-4+4-8, early past-EEG, CCA-spatial)",
   "status": "closed_negative",
   "evidence": "LODO 0.6622-0.6816; AASD 0.6219-0.6252/50.9-51.1",
   "source": "multi_agent_opt/results_trf_*.md"
  },
  {
   "id": "PTA-02",
   "family": "classical_pathA",
   "method": "DSS/SSD spatial filter -> TRF",
   "status": "closed_negative",
   "evidence": "LODO 0.6891/62.75; AASD 0.6342/52.09",
   "source": "multi_agent_opt/results_trf_*.md"
  },
  {
   "id": "PTA-03",
   "family": "classical_pathA",
   "method": "CSP->Riemann MDM, joint SPD; Hilbert PLV coherence",
   "status": "closed_negative",
   "evidence": "chance, abandoned",
   "source": "multi_agent_opt/results_trf_*.md"
  },
  {
   "id": "PTA-04",
   "family": "classical_pathA",
   "method": "TRF Path A <-> SW2/swmix Acc-leader blend; TRF Path B into vlaai",
   "status": "closed_negative",
   "evidence": "AASD 52.68-52.88",
   "source": "multi_agent_opt/results_trf_*.md"
  },
  {
   "id": "ENS-01",
   "family": "ensemble",
   "method": "4-seed average of classifier + 4-seed decoder, 5th seed s4 (newseed-hp30)",
   "status": "in_stack",
   "evidence": "published; s4 hp30 Acc 53.27 (+0.15 flat)",
   "source": "ALREADY_TRIED.md B/E"
  },
  {
   "id": "ENS-02",
   "family": "ensemble",
   "method": "LDA/inv-var/median/logreg stacks on 4 logits; nested lda4/dyn; fractional HP hpfrac",
   "status": "closed_no_gain",
   "evidence": "0.7947-0.8089; Acc <= 73",
   "source": "ALREADY_TRIED.md B/E"
  },
  {
   "id": "ENS-03",
   "family": "ensemble",
   "method": "SW2<->swmix hard_abs gate (Acc tip 53.53)",
   "status": "closed_no_gain",
   "evidence": "AASD 0.6424/53.53; Carta 0.6460/59.71",
   "source": "ALREADY_TRIED.md B/E"
  },
  {
   "id": "EXT-01",
   "family": "data",
   "method": "Fuglsang2020 as extra train + EA prior",
   "status": "in_stack",
   "evidence": "+0.004",
   "source": ""
  },
  {
   "id": "EXT-02",
   "family": "data",
   "method": "SparrKULee match/mismatch pretraining (checkpoints models/pretrain_spk_*)",
   "status": "in_stack",
   "evidence": "+0.005",
   "source": ""
  },
  {
   "id": "EXT-03",
   "family": "data",
   "method": "analysis only (no model): attention transition event-locking, dataset profile",
   "status": "closed_negative",
   "evidence": "tracking margin post-pre ~1e-3; AASD keypress artifact likely",
   "source": "results/research/attention_transition_*.csv"
  },
  {
   "id": "INT-01",
   "family": "integrity",
   "method": "zip stem_ms96 (eeg-thao.zip) & sibling audit: first 24 s hard column 0, symmetrised 0.827, real AASD 0.613-0.618 Carta 0.546-0.561",
   "status": "closed_negative",
   "evidence": "sibling repo stopped it 2026-10-28",
   "source": "REAL_SWITCH_STRATEGY.md sec 1, 8"
  }
 ],
 "forbidden_note": "status in_stack = already part of the base recipe; re-using as a frozen component is allowed but re-tuning/re-running is not. integrity_suspect = numbers likely inflated by column/position prior; do not rerun; any new second-stage net must be exactly antisymmetric f(-x)=-f(x) without bias/pos-embedding/zero pad.",
 "train_entrypoints": [
  {
   "name": "LODO + all5 classifier (direct) per seed S",
   "where": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo",
   "cmd": "cd /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo; GLIMPSE_THR=onorm /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python direct_classifier.py --prep eaP_n5_t60 --window-s 10 --set model_blocks=3 --init-from models/pretrain_spk_direct_clf_w10_b3_sS.pt --min-epochs 10 --train-seed S --p-switch 0.5 --aug-set clf_p_switch2=0.5 --tag _w10b3_ptSPK_sS_aug-sw2",
   "produces": "5 LODO folds folds/<exp>/<ds>.pkl + results/<exp>.json; mirrored copy of code also in /root/disk_3tb/eeg-thao",
   "note": "jobs go through runners/jobqueue.py (24 GB RAM shared)"
  },
  {
   "name": "SparrKULee pretrain of classifier",
   "where": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo",
   "cmd": "GLIMPSE_THR=onorm /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python direct_classifier.py --window-s 10 --pretrain-spk --tag _w10_b3 --train-seed S --min-epochs 25 --max-epochs 80 --set model_blocks=3"
  },
  {
   "name": "Reconstruction decoder LODO",
   "where": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo",
   "cmd": "GLIMPSE_THR=norm /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python run_experiment.py --decoder vlaai_ws --prep eaP_n5_t60 --aug-variant cluster --extra-train Fuglsang2020 --targets env,onset,dom --init-from models/pretrain_spk_vlaai_ws_sS.pt --evidence-weights lda --tag _aug_cluster_onset_dom_xFUG_ptSPK_sS --train-seed S"
  },
  {
   "name": "Prep + EA prior",
   "where": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo",
   "cmd": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python build_ea_prior.py && /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python prepare_data.py eaP_n5_t60"
  },
  {
   "name": "all5 classifier for AASD/Carta (trains on all 5 + Fuglsang, runs on external)",
   "where": "/root/disk_3tb/eeg-thao (also /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo)",
   "cmd": "GLIMPSE_THR=onorm /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python eval_external_clf.py --seed S --window-s 10 --init-from models/pretrain_spk_direct_clf_w10_b3_sS.pt --ref direct_clf__eaP_n5_t60_w10b3_ptSPK_sS_aug-sw2 --tag _w10b3sw2P --prep eaP_n5_t60 --set model_blocks=3 --p-switch 0.5 --aug-set clf_p_switch2=0.5"
  },
  {
   "name": "all5 decoder for AASD/Carta",
   "where": "/root/disk_3tb/eeg-thao",
   "cmd": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python eval_external.py --seed S --prep eaP_n5_t60 ... (see docstring; reads external_data/<DS>/cache/eaP_n5_t60/*.npz)"
  },
  {
   "name": "stem_ms96 LODO (nested inner selection) - integrity suspect",
   "where": "/root/disk_3tb/eeg-thao",
   "cmd": "cd /root/disk_3tb/eeg-thao; export PYTHONDONTWRITEBYTECODE=1 PYTHONUNBUFFERED=1 OMP_NUM_THREADS=8; STEPS=1000 BATCH=512 /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python -u ensemble_clf_transformer_v2.py",
   "produces": "results/research/ensemble_clf_transformer_v2.csv, models/logit_tf_v2/"
  },
  {
   "name": "stem fine-tune family (all train LODO folds + all5.pt, e.g. IRM/focal/synth/group-dro)",
   "where": "/root/disk_3tb/eeg-thao",
   "cmd": "STEPS=800 flock /root/disk_3tb/eeg-thao/multi_agent_opt/gpu.lock /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python -u stem_ms96_group_dro.py  (also stem_ms96_irm.py, stem_ms96_focal_avgc.py, stem_ms96_synth_switch.py)",
   "produces": "models/stem_ms96_*/{held-*,all5}.pt; results/research/<name>.json"
  },
  {
   "name": "external winner linear_hp30 / v4",
   "where": "/root/disk_3tb/eeg-thao",
   "cmd": "MODE=external STEPS=800 /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python -u ensemble_clf_transformer_v4.py ; MODE=lodo_hp /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python -u ensemble_clf_transformer_v4.py"
  }
 ],
 "eval_entrypoints": [
  {
   "name": "LODO synthetic-switch scorer",
   "file": "/root/disk_3tb/eeg-thao/score.py",
   "use": "score_dataset / score_experiment (HMM emissions from other datasets; official scorer port)"
  },
  {
   "name": "Official scorer port",
   "file": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/reports/eda_full/scripts/scorer.py",
   "use": "fast_score_subject, aggregate"
  },
  {
   "name": "Fusion LODO 4dec+4clf",
   "file": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/fuse_general.py",
   "cmd": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python fuse_general.py --dec D D_s1 D_s2 D_s3 --clf W_aug-sw2 W_s1_aug-sw2 W_s2_aug-sw2 W_s3_aug-sw2 --a 0.5 --pstay 0.997 0.998 0.9997"
  },
  {
   "name": "Fusion on AASD/Carta",
   "file": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/fuse_external.py",
   "cmd": "python fuse_external.py --dec CORPUS_EXP:SEED ... --clf CORPUS_EXP:EXT_TAG:SEED ... [--a ...] [--pstay ...] [--name label]"
  },
  {
   "name": "External scorer (stem family)",
   "file": "/root/disk_3tb/eeg-thao/eval_external.py",
   "use": "E.score(...) used by stem_ms96_*.py, ensemble_clf_transformer_v4.py"
  },
  {
   "name": "stem on real switches with symmetrisation (use this to audit any second stage)",
   "file": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/eval_stem_external.py",
   "cmd": "python eval_stem_external.py --ckpt-dir DIR [--device cuda]  -> results/research/stem_external.csv (variants mean4/orig/sym)"
  },
  {
   "name": "Lookahead / leakage test (<=0.969 s)",
   "file": "/root/disk_3tb/eeg-thao/leakage_test.py; /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/leakage_test_b1.py (column swap + position invariance)",
   "cmd": "GLIMPSE_THR=norm python leakage_test.py"
  },
  {
   "name": "Decision-layer nested LODO + external frozen cfg",
   "file": "/root/disk_3tb/eeg-thao/causal_eeg_nested_lodo.py",
   "cmd": "CUDA_VISIBLE_DEVICES= OMP_NUM_THREADS=8 /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python -u causal_eeg_nested_lodo.py (98 k s; reads models/focal_synth_adapt_readout/cache)"
  }
 ],
 "data_paths": {
  "workspace": "/root/disk_3tb/eeg-thao",
  "dataset_repo": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset",
  "python": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python",
  "gpu": "1x Tesla V100-SXM2-32GB; take flock on /root/disk_3tb/eeg-thao/multi_agent_opt/gpu.lock before GPU jobs; box shared, 24 GB RAM (docs), no swap",
  "organiser_cnd_raw": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/data/cnd/{AVGCAAD2024,DTU2018,KUL2016,KUL_AAD_UltraHD2021,USTC2026}",
  "five_dataset_keys": {
   "AVGC": "KUL-AV-GC 13 subj/99 trials/~600 s",
   "DTU": "DTU 18/1080/~50 s",
   "KUL": "KUL2016 16/320/~124 s",
   "UHD": "KUL-UHD 30/120/~391 s",
   "USTC": "USTC 18/360/~120 s (only NeuroScan train set)"
  },
  "external_root": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/external_data",
  "AASD": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/external_data/AASD/cache/{eaP_n5_t60,mat,mat_reac,lap_eaP,lp7_eaP,bipolar_eaP,env,index.csv}  (900 npz)",
  "Carta2026": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/external_data/Carta2026/cache/ (480 trials)",
  "Fuglsang2020": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/external_data/Fuglsang2020",
  "SparrKULee": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/external_data/SparrKULee",
  "Broderick2018": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/external_data/Broderick2018",
  "foundation_models": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/external_data/foundation_models",
  "fold_evidence_workspace": "/root/disk_3tb/eeg-thao/folds/<exp>/<ds>.pkl",
  "results_workspace": "/root/disk_3tb/eeg-thao/results/*.json, /root/disk_3tb/eeg-thao/results/research/*",
  "fold_evidence_sibling": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/results/research/*evidence*.pkl",
  "ledgers_workspace": "/root/disk_3tb/eeg-thao/multi_agent_opt/{ALREADY_TRIED.md,CLAIMS.md,CURRENT_BEST.md,STATUS_GAPS.md,results_*.md}",
  "ledgers_sibling": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/{CONTEXT_1_STATE.md,CONTEXT_2_HISTORY.md,CONTEXT_3_RULES.md,CONTEXT_4_OPERATIONS.md,REAL_SWITCH_STRATEGY.md,EXPERIMENTS_LOG.md(to 2026-10-15),CONTEXT_MEMORY_real-switch-findings.md}",
  "best_documented_recipe": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/README_BEST.md",
  "challenge_rules": "/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/CONTEXT_3_RULES.md; /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/docs_challenge/*.txt",
  "models": "/root/disk_3tb/eeg-thao/models/ (all5.pt, held-*.pt per method); /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/decoder_lodo/models/",
  "agent_transcripts": "/root/.cursor/projects/root-disk-3tb-eeg-thao/agent-transcripts (2026-10-24 v4 run d58e57c5-b97f-4f13-8baa-3eee5c1b79c5)"
 },
 "gaps": [
  {
   "id": "GAP-01",
   "severity": "critical",
   "text": "No antisymmetry/column-swap/position-invariance test has ever been run on the workspace stem_ms96 family (non-odd net: y=(att==0), zero-init bias, zero-padded start, abs-pos). All LODO tip numbers 0.87-0.884 are unverified; sibling audit of the same architecture saw 0.863->0.827 symmetrised and real-switch collapse (Carta 0.55). Real-switch results of the stem tip (AASD 0.622/51.2, Carta 0.553/49.4) confirm it does not transfer."
  },
  {
   "id": "GAP-02",
   "severity": "critical",
   "text": "Synthetic LODO does not predict real switches: all points lie on one frontier set by HMM p_stay; gains 0.78->0.80 synthetic did not move AASD/Carta (H ~0.63-0.65). Report both always; accept only if both rise (REAL_SWITCH_STRATEGY sec 6/9)."
  },
  {
   "id": "GAP-03",
   "severity": "high",
   "text": "AASD frame accuracy 52-54 % (acc1 0.50-0.54) is the cap; Acc>60 on AASD unreachable by any decision-layer/logit post-processing tried (>150 variants). Needs a better EEG evidence view, not another readout."
  },
  {
   "id": "GAP-04",
   "severity": "high",
   "text": "Only 2 real-switch sets, seed noise +-0.005..0.014 (2 seeds), schedule noise +-0.0015 (ensemble) / +-0.005..0.01 per dataset. No subject-level CIs, no automated acceptance report; differences < 0.005 are not evidence."
  },
  {
   "id": "GAP-05",
   "severity": "high",
   "text": "Classifier-family hyperparameters (a, p_stay, blocks, aug, EA weight) chosen on LODO itself (README_BEST sec 11), not nested; stem family selection is nested but on the suspect evidence."
  },
  {
   "id": "GAP-06",
   "severity": "high",
   "text": "No submission pipeline: final model on all 5 datasets + inference script emitting 128 Hz labels per trial not written (README_BEST sec 11); submissions/ is empty. Test phase 2026-11-09..12-08, only 2 submissions, last counts."
  },
  {
   "id": "GAP-07",
   "severity": "medium",
   "text": "leakage_test.py for 3-block eaP_n5_t60 classifier only covered by leakage_b1 (seed 0, B1 rule); not run for any stem/second-stage net, BMMS/recipe/antisym transformers (audit_transformer_leakage.py is a surrogate only)."
  },
  {
   "id": "GAP-08",
   "severity": "medium",
   "text": "CLAIMS board lists dual-stream-corr-acc as 'claimed' (2026-10-30) but no script/result exists; treat as not run (EEG<->env0/env1 causal correlation features -> DirectNet head)."
  },
  {
   "id": "GAP-09",
   "severity": "medium",
   "text": "Duplicate code trees (workspace vs decoder_lodo vs eeg-thao_unzipped) and two ledgers; workspace ledgers stop 2026-10-30 (+ Nov results in results_preprocessing.md, causal_eeg_nested_lodo), sibling ledger has Nov 3-6 BMMS/recipe/antisym results undocumented in ALREADY_TRIED.md."
  },
  {
   "id": "GAP-10",
   "severity": "medium",
   "text": "Population EA prior from Fuglsang (BioSemi) only; USTC is the only NeuroScan train set and NJU-SUSTech test is NeuroScan; prior from all train data not computed."
  },
  {
   "id": "GAP-11",
   "severity": "low",
   "text": "Not yet tried or only screened (candidates, none claimed): exactly-antisymmetric second stage over the 8 raw logits+HMM with no pos/bias; per-trial causal adaptation of the classifier head inside one trial; larger classifier seed bank beyond 5; multi-view EEG evidence that is not frozen-FM (see ARC-14); ensembles over eaP + lap/lp7 trained from scratch (only 1 seed each banked); cross-corpus pretraining on Broderick under SPK recipe with antisymmetry."
  }
 ],
 "counts": {
  "forbidden_entries": 79
 }
}