---
name: decoder-lodo-findings
description: "Key findings of the LODO decoder benchmark in experiments/decoder_lodo (as of 2026-09-30); results catalog, seed noise, what helped"
metadata:
  node_type: memory
  type: project
  originSessionId: d4dac82a-f3e4-4196-a274-a5c105cb0541
  modified: 2026-10-01T01:26:02.505Z
---

The benchmark lives in `experiments/decoder_lodo`:
- all SuperHugeAAD decoders, run LODO over the 5 datasets;
- scored with HMM p_stay 0.998 on synthetic switches and the official scorer port.

Results go to `results/CATALOG.md` (groups A–H), `summary.xlsx` and `results/research/`. Runner scripts are in `runners/`; tmux session `decoder_lodo`.

Key facts (2026-09-30):
- **Seed noise ≈ 0.018 H.** The best config vlaai_ws + mat + aug + cluster scored 0.726 / 0.709 / 0.708 / 0.710 over 4 seeds, mean ≈ 0.713. Differences under about 0.02 H need 2+ seeds.
- **Real gains** (beyond seed noise):
  - DNN beats WF (0.682);
  - mat prep beats py and fft;
  - full-trial z-score beats running z;
  - VLAAI package config has `use_skip: false`; turning skip on makes it learn;
  - onset-envelope multi-target: H 0.726 / 0.724, SNR1 0.141→0.155, better on all 5 folds.
- **New best (2026-10-01):** CAUSAL running Euclidean alignment on top of onset. Prep `mat_reac` carries the covariance across a subject's trials in recording order. Over 4 seeds: H 0.746 / 0.741 / 0.741 / 0.740 (mean 0.742 ± 0.002), Acc 68.8%, SDT 12.1 s, SNR1 0.166. Covariance augmentation (covmix) adds only +0.002 (paired −0.004..+0.008), so it is not adopted. Plain `mat_rea` (tau 60 s, reset per trial) scored 0.733; tau 30 / 120 s scored ≈0.738. ADT transformer: 0.718.
- **External data downloaded (2026-10-01):** AASD, Fuglsang2020 and Broderick2018 (the user downloaded Broderick manually; it has no audio, only precomputed envelopes, and 128 BioSemi channels), all in `experiments/external_data`.
- **No gain:**
  - HMM tweaks (p_stay, temperature, adaptive emissions, CUSUM; +0.002–0.006);
  - ensembles; Markov switching model (−0.01);
  - windows longer than 5 s; wider band 1–15 Hz; SSMamba;
  - warm-up; balanced sampling; most augmentations.

SparrKULee (2026-10-02): access granted; credentials only in `/root/.netrc_sparrkulee` (never in repo/memory).
- Fetched only `derivatives/preprocessed_eeg` + `preprocessed_stimuli/*envelope.npy` (~19 GB; 64 ch, 64 Hz, BioSemi order).
- The user wants it fully unattended: `decoder_lodo/runners/sparrkulee_chain.sh` (tmux window `spk`, log
  `logs/sparrkulee_chain.log`) downloads -> converts -> queues jobs 110-116 (convert, pretrain s0/s1, S2 xFUGSPK25 s0/s1,
  S1 xFUG_ptSPK s0/s1). Models in `decoder_lodo/models/`.
- Phase 2 `runners/sparrkulee_chain_adt.sh` (window `spk_adt`): ADT transformer jobs 120-127 (baseline xFUG, pretrain, SPK25, ptSPK), 2 seeds.
- Single speaker: env[:,1] = own envelope shifted by half the run (mismatched negative); EEG float16 mmap.

**Why:** the user is iterating on improvements for the ICASSP 2027 EEG-sAAD challenge.

**How to apply:**
- Treat SNR1 as the low-noise metric.
- Give any new idea 2+ seeds.
- The main limit is decoder SNR (≈0.15), not the integrator.

See [[full-corpus-eda]] and [[zscore-allowed]].

Results of 2026-10-02:
- **F1, spectral band equalisation:** 0.684 ❌ (it boosts noisy bands).
- **E2, gamma-envelope channels:** 0.728 ❌.
- **E1, delta/theta phase channels:** 0.692 ❌; seed 0 was OOM-killed.
- **C1 on AASD (real spontaneous switches, about 1 every 11.7 s):** used the best config trained on all 5 datasets and the converter `external_data/AASD/convert_aasd.py`.
  - Audio = mixed_{code-10}; key 179 = left.
  - Window SNR1 ≈ 0.07 (corpus 0.166), acc1 ≈ 53%.
  - HMM 0.998: Acc 51.6%, H 0.63, vs a NULL of 49.7% / 0.61.
  - p_stay 0.9 gives a slightly better H of 0.65.
  - So on AASD the model is barely above chance. Synthetic-switch corpus numbers are optimistic for frequent real switches and new domains.

**Git (2026-10-02):** `experiments/` is a local git repo with only code, config and results JSON (~3 MB). The `.gitignore` excludes cache, folds, logs and raw data. Commit after meaningful code changes.

External data as extra training (LODO, 2 seeds each, 2026-10-02):
- **Best config alone:** 0.743.
- **+AASD:** 0.742 (USTC +0.010).
- **+Broderick:** 0.740 ❌.
- **+Fuglsang:** 0.746, SNR1 0.175 (+4%), DTU +0.009. This is the only positive one, and it is small.
- **+Fuglsang+Broderick:** 0.740; KUL-AV-GC −0.021.
- **AASD real switches:**
  - organisers' WF baseline: Acc 51.0%, H 0.619;
  - best config: 51.6% / 0.63;
  - +AASD fine-tune: 51.8–51.9% / 0.63;
  - null: 49.7% / 0.61.
  - AASD is intrinsically low-SNR.

2026-10-02, later:
- **+Fuglsang over 4 seeds:** 0.749 / 0.743 / 0.748 / 0.743 (mean 0.746 ± 0.003). The paired gain over best is +0.002 to +0.008 on 4/4 pairs; SNR1 0.174. This is the **new best** (USTC +0.017, DTU +0.008).
- **p_stay tuning (`results/research/pstay_dense.csv`):**
  - corpus with sparse synthetic switches: best at 0.998 (0.746);
  - dense synthetic (12 s): best at 0.95 (0.607 vs 0.571 at 0.998);
  - AASD real: 0.648 at 0.9 vs 0.630 at 0.998;
  - minimax-regret choice: 0.99.

**SparrKULee plan (agreed 2026-10-02, waiting for KU Leuven access):**
- Download `derivatives/` (41 GB) and reprocess with the compliant pipeline.
- **S1:** pretrain vlaai_ws on SparrKULee (Pearson loss, envelope + onset), then fine-tune on the 5 datasets + Fuglsang with the contrastive loss.
- **S2:** mix SparrKULee directly into training at about 25% of each batch, using mismatched-segment negatives.
- Then a wider vlaai_ws and a transformer (ADT or contrastive transformer) with pretraining.
- 2 seeds each, compared against the best config (+Fuglsang, H 0.746).
- **+Fuglsang+AASD (2026-10-02):** 0.739 / 0.747 (mean 0.743). Not better than +Fuglsang; USTC 0.764 but KUL2016 −0.014 and DTU −0.012. Keep +Fuglsang only.

2026-10-02, afternoon:
- **Learning curve (corpus only):** 25% 0.719, 50% 0.733, 75% 0.741, 100% 0.742. It saturates, so more same-kind data adds little.
- **+Carta in training:** 0.734 ❌.
- **+Fuglsang+Carta:** 0.743, not better than +Fuglsang (0.746).
- **Carta 2026 real cued switches (~25 s), model trained on 5 datasets:**
  - best config: Acc 56.5 / 55.8%, H 0.616 / 0.610;
  - same evidence with synthetic switches: 61.8 / 60.7%, H 0.673 / 0.665;
  - WF baseline: 54.7%, H 0.592.
- Real switches cost about 0.05 H, consistently on both AASD and Carta.
- Converter: `external_data/Carta2026/convert_carta.py`.

SparrKULee results (2026-10-02 evening; 105 subjects, 793 runs, 190 h converted; pretraining early-stops at ~epoch 17-26):
- **vlaai_ws S1 (pretrain SPK -> fine-tune +FUG):** 0.748 / 0.753 (mean 0.751, SNR1 0.178) vs ref 0.746 (4 seeds). Paired −0.001 / +0.011, so not yet conclusive; USTC +0.026, AVGC −0.010. Candidate new best; needs seeds 2-3.
- **vlaai_ws S2 (SPK 25% of batch):** 0.747 / 0.748 (+0.002), within noise.
- **ADT:** ref +FUG 0.740; S2 0.742; S1 0.745 (paired +0.005 / +0.006). Pretraining helps the transformer, but it stays ≤ vlaai_ws.
- **Gamma pilot:** the challenge data are 128 Hz (AVGC low-passed at ~48 Hz). 35-48 Hz linear decoding scores 50% on DTU/KUL, so this direction is dropped.
- **Final (2026-10-03):**
  - vlaai_ws S1 over 4 seeds: 0.748 / 0.753 / 0.751 / 0.742 (mean 0.749), paired +0.003 with 2/4 positive. Not significant, so the best stays +FUG 0.746.
  - S1 long (52 epochs): 0.746. S1 fugval (early stop on Fuglsang): 0.747, SNR1 0.178, SDT 11.2.
  - HappyQuokka (11M params, added as `happyquokka` in decoders.py): ref 0.739, S1 0.740. Below vlaai_ws.
  - Conclusion: SparrKULee gives ≤ +0.003 H. The bottleneck is not pretraining data.

Glimpse targets (2026-10-03): `build_glimpse.py` (cache/glimpse10; talker >10 dB louder on the raw envelope, causal) adds targets `glim` / `mask`; Fuglsang via `build_glimpse_fug.py` (threshold calibrated to the corpus glimpse fraction 0.42).
- **env+onset+glim+mask:** no FUG 0.745 / 0.749 (ref 0.742), SNR1 0.179 (ref 0.166); +FUG 0.744 / 0.750 (ref 0.746), **SNR1 0.184** (ref 0.174, highest so far). H gain is small and AVGC drops ~0.01. Seeds 2-3 of +FUG queued (jobs 172-173).
- **env+onset+glim only:** 0.737 / 0.741 ❌. The masked part is what helps.
- **glim+mask +FUG over 4 seeds (2026-10-03):** 0.744 / 0.750 / 0.754 / 0.746 (mean 0.749), paired vs ref +0.003 (3/4 positive), SNR1 0.183. Reruns are bit-reproducible.
- **Decision stage for 4 targets (`decision_targets.py`, `results/research/decision_targets.csv`):** LDA target weights (env .18, onset .42, glim .16, mask .22) give +0.002 H (0.7505). Temperature, Student-t, running-z and p_stay give nothing. The decision stage is not the bottleneck; H rises only ~0.5 per unit SNR1. The mask target alone is at chance on AVGC (H 0.546); drop_mask gives AVGC 0.708 vs 0.694.
- Variants queued: soft20 / norm dominance masks, `dom` target (normalised (e1-e2)/(e1+e2)), each also from SparrKULee-pretrained init (jobs 180-199; `--init-from` now copies the first head columns).
- **Dominance target `dom` (2026-10-03, user's idea: g = (e1-e2)/(e1+e2+eps), causal, band-passed; GLIMPSE_THR=norm):** targets env+onset+dom +FUG: **0.756 / 0.753 (mean 0.754, SNR1 0.190, Acc 69.9, SDT 11.3)** vs ref 0.746 / 0.174: the largest gain since causal EA. With SparrKULee-pretrained init: 0.757 (s0). glim+mask(norm)+dom: 0.751 / 0.757; soft/norm glim+mask without dom: 0.749-0.750. Seeds 2-3 queued (jobs 202-213). AVGC stays ~0.69-0.70 (no gain there).
- **CONFIRMED new best (2026-10-03 06:30):** env+onset+dom +FUG over 4 seeds: 0.756 / 0.753 / 0.754 / 0.754 (mean 0.754 ± 0.001), paired vs ref +0.006..+0.011 (4/4), SNR1 0.190, Acc 69.9, SDT 11.1. With SparrKULee-pretrained init, 4 seeds: 0.757 / 0.759 / 0.761 / 0.759 (mean 0.759 ± 0.001, paired vs ref +0.009..+0.016, 4/4), Acc 70.1, SDT 11.0, SNR1 0.191, AVGC 0.706. This is the BEST CONFIG. Command: `GLIMPSE_THR=norm run_experiment.py --decoder vlaai_ws --prep mat_reac --aug-variant cluster --targets env,onset,dom --extra-train Fuglsang2020 [--init-from models/pretrain_spk_vlaai_ws_sN.pt]`.
- Relative-onset targets `orel` / `odiff` (cache `glimpseonorm`, GLIMPSE_THR=onorm) queued as jobs 220-231 (3x2 grid with/without dom).
- **Target-weight re-scoring (2026-10-03, `rescore_weights.py`, `results/research/rescore_weights.csv`):** LDA weights for the evidence (fitted on the other 4 datasets) give +0.0025 H on average over 41 multi-target runs, positive in 36/41 (paired t 7.8). Best config with LDA: 0.7600 / 0.7628 / 0.7609 / 0.7600 = **0.761** (equal weights 0.759); weights ≈ env 0.06, onset 0.59, dom 0.35. Dropping env from the evidence: no gain. Not yet wired into score.py (analysis only).
- **Ablations done:** raw difference targets do not help (ediff 0.747, odiff 0.745; dom+ediff 0.753, +pt 0.754); the normalised ones do (dom +0.008, orel +0.004). dom+orel 0.757 without pretraining, but 0.754 with pretraining (2- or 4-target head init alike), so orel is not added. dom helps mainly as EVIDENCE (dropping it from the evidence: 0.759 -> 0.751), little as a pure auxiliary.
- **LDA evidence weights are now in the pipeline (2026-10-03):** `run_experiment.py --evidence-weights lda` (score.lda_weights / weighted_evidence; default `equal` unchanged and verified to reproduce stored results). BEST CONFIG = vlaai_ws, mat_reac, aug cluster, targets env+onset+dom (GLIMPSE_THR=norm), +Fuglsang, init from SparrKULee pretrain, LDA weights: **H 0.7600 / 0.7628 / 0.7609 / 0.7600 = 0.761**, Acc 70.4, SDT 10.7, SNR1 0.195 (results `..._dom_xFUG_ptSPK_wlda[_sN].json`; folds are symlinks to the equal-weight runs, so re-scoring needs no retraining).
- **Causal temporal whitening per channel (2026-10-03, `causal_tw` in prepare_data.py, preps mat_reac_tw / tw05 / tws):** ❌ dropped at the user's request. Full whitening 0.751 / 0.751 vs 0.754 (DTU −0.010, UHD, USTC down on both seeds; AVGC fold SNR1 up but H inconsistent). Half strength 0.755 (1 seed). The spectral-standardisation variant (whiten + fixed re-colouring) did not equalise the datasets' spectra and was not finished.
- **Follow-ups (2026-10-03 evening):**
  - dom + SparrKULee pretrain early-stopped on Fuglsang, 4 seeds: 0.7584 equal / 0.7600 LDA. No gain over the standard pretrain (0.7588 / 0.7609).
  - **No env target (onset+dom, +FUG, +pt):** 0.7583 / 0.7638 equal (mean 0.761), 0.7599 / 0.7626 LDA (0.761). Same as best+LDA with one target fewer and no need for LDA; AVGC 0.697 (best 0.706), KUL 0.764. 2 seeds only.
  - **Best config on REAL switches (`eval_external.py --targets env,onset,dom --extra-train Fuglsang2020 --init-from ... --evidence-weights lda --tag _best`; dom for AASD/Carta via `external_data/build_glimpse_ext.py`):** the corpus gain barely transfers. AASD p0.998: Acc 52.0 / 51.9, H 0.633 / 0.631 (old best 51.6 / 51.6, 0.628 / 0.632). Carta p0.998: Acc 57.2 / 56.3, H 0.628 / 0.611 (old 56.5 / 55.8, 0.616 / 0.610); at p0.99 no gain (0.637 / 0.626 vs 0.636 / 0.634). Carta window SNR1 unchanged (~0.10 / 0.09). LDA weights add ~0 on real switches.
- **A1 end-to-end sequence training (2026-10-04, `--seq-loss filter|crf`, `DNNDec.fit_sequence`, 40-s sequences with synthetic switches):** no gain. Fine-tuned decoder scored with the standard HMM: filter 0.7573 equal / 0.7599 LDA, crf 0.7587 / 0.7605 (best 0.7580 / 0.7614); no-env + crf 0.7579 / 0.7601 (no-env alone 0.7610). Scoring with the LEARNED unary weights is much worse (0.672 / 0.694): the CE/NLL shrinks the LLR scale ~4x to calibrate the overlapping windows, which slows the filter. Fixing the total scale (`--seq-v-norm`) made it worse still (0.737 equal), the decoder itself degrades. Conclusion: decision stage + objective mismatch, not worth pursuing.
- **A2 direct classifier (2026-10-04, `direct_classifier.py`: antisymmetric g(E,A,B)-g(E,B,A), vlaai_ws backbone on EEG+stimulus channels, synthetic switches applied to the INPUT with the scorer's schedule, evidence flagged pre_switched):**
  - Shortcut controls are clean in every run: EEG zeroed 0.50, EEG shifted by half a trial 0.50, frontal channels zeroed = normal (±0.004); envelope-only model 0.496 window acc, H 0.520. So no envelope or gaze shortcut.
  - Training is UNSTABLE: many folds never leave chance (val acc ~0.51 for 7 epochs, early stop); SparrKULee match-mismatch pretraining seed 0 reached 0.808 val acc (40 epochs, still improving), seed 1 never took off.
  - With the good pretrain (seed 0): H 0.7566, Acc 68.9, **SDT 9.5 s**, 5-s decision acc 0.657 (AVGC 0.675, DTU 0.813, KUL 0.761, UHD 0.745, USTC 0.789), i.e. on par with the best decoder at equal weights (0.7574, SDT 10.7) with faster switch detection but AVGC lower. From scratch 0.628; no-env variants 0.61-0.63 (failed folds).
  - Next if pursued: stabilise training (minimum epochs before early stopping), more pretrain seeds, fuse its logit with the decoder evidence.
- **NEW BEST (2026-10-04): fusion of the best decoder with the direct classifier.** Direct classifier with stabilised training (`--min-epochs`: no LR decay / early stop at the start; SparrKULee match-mismatch pretrain 25-80 epochs, val acc 0.83-0.84, then LODO fine-tune) alone: H 0.7564 / 0.7604 / 0.7583 (seeds 0-2), SDT 8.4-10.0 s, no fold stuck, shortcut controls clean. Fusion `fuse_mix.py` (d = (1-a) z_decoderLDA + a z_classifier, a chosen on the other 4 datasets): **0.7681 / 0.7680 / 0.7710 (mean 0.769)** vs decoder+LDA 0.7600 / 0.7628 / 0.7611; fixed a = 0.3 gives 0.7682 / 0.7693 / 0.7710; Acc ~70.3, SDT ~8.8 s. LDA over all sources does NOT work for this fusion (classifier windows are 5 s, far more autocorrelated; seed 0 old run 0.7593). Seed 3, learned-weight fusion (`fuse_learn.py`), from-scratch v2 and transformer backbones (adt, happyquokka; `--backbone`) are queued (jobs 323-343).
- **Seed ensembles (2026-10-04, `fuse_ensemble.py`, 3 seeds, evidence averaged per step):** decoder 0.7613 -> 0.7626 (+0.001); direct classifier 0.7589 -> 0.7663 (+0.007); fused 70/30 0.7695 -> **0.7725** (a = 0.15: 0.7728), Acc 70.5-70.8, SDT 8.8-9.2 s. Best LODO number so far ≈ 0.773 (one ensemble, no variance estimate). Learned-weight fusion (`fuse_learn.py`) equals the fixed 70/30 mix fold by fold (±0.003) and learns ~25-30 % for the classifier; LDA fusion fails (0.760). Direct classifier from scratch with stabilised training: 0.7545 (seed 0) vs 0.7564 pretrained, so the stabilisation, not the pretraining, was the key.
- **Goal stated by the user (2026-10-04): maximise the competition score (dev phase now, test phase later).** `fuse_general.py` (any decoders + any classifiers, a, p_stay): 4 decoders + 3 classifiers, a 0.3: 0.7733; wider ensembles do not help (10 decoders 0.7732; 5 classifiers incl. weaker ones 0.771). **Re-tuned p_stay for the fused ensemble: optimum ~0.9997 -> H 0.7768** (plateau 0.9995-0.9998; Acc 72.0, SDT 9.7); for one decoder alone 0.998 stays best. Not nested yet, and specific to sparse synthetic switches (a stickier HMM would hurt on dense real switches). Learned-weight fusion over 3 seeds: 0.7688 = fixed 70/30 (0.7695). From-scratch classifier seed 1 still had one stuck fold (0.712), so keep the pretraining. Queued: classifiers with 3-s and 10-s windows (`--window-s`, jobs 335-338), ADT and HappyQuokka classifier backbones (340-343).
- **4 seeds complete (2026-10-04 04:45):** direct classifier (pretrained, stabilised) alone 0.7564 / 0.7604 / 0.7583 / 0.7617 (0.759); fused fixed 70/30 per seed 0.7682 / 0.7693 / 0.7710 / 0.7698 (**0.770**, decoder+LDA alone 0.761; nested a 0.768); LDA fusion 0.761 (useless). Ensemble 4 decoders + 4 classifiers, a 0.3: 0.7731 at p_stay 0.998, **0.7769 at p_stay 0.9997** (Acc 72.0, SDT 10.0). A 4th classifier seed adds nothing over 3.
- 2026-10-04: the user dropped the ADT and HappyQuokka classifier backbones (jobs 340-343 removed before they ran; `--backbone` stays in the code, untested beyond a forward pass).
- **2026-10-04 06:00:** learned-weight fusion over 4 seeds 0.7691 = fixed 70/30 (0.7696) -> use the fixed ratio. Seed correlation of the evidence: decoder 0.94, classifier 0.80 (why decoder ensembles gain only +0.002-0.003). **3-s classifier** (`--window-s 3`): alone 0.7549 / 0.7584 (5-s: 0.7564 / 0.7604), fused with the decoder 0.7694 / 0.7682, but adding it to the 4+4 ensemble gives nothing (0.7731 at p 0.998, 0.7750 at 0.9997 vs 0.7769). ADT / HappyQuokka classifier backbones dropped by the user. Queued: 10-s classifier (337-338), best decoder with saved reconstructions for multi-scale evidence (350-351), ADT decoder with dom+pretrain for decoder diversity (352-353).
- **Unused lookahead budget found (2026-10-04 06:15):** the DNN reconstruction of hop block [a, a+8) uses EEG < a+24 (verified by perturbing the EEG: 0.5-0.75 s future per sample), so a 1-s evidence window ending at t_end needs EEG only up to t_end + 0.5 s. score.py stamps with lookahead 0.75 (labels from t_end - 0.25), i.e. 0.25 s of the 1-s budget is wasted. Scoring with lookahead 0.5 (`score_experiment(exp, "causal", 0.5, "lda")`): best decoder 0.7610 -> **0.7633** on all 4 seeds (+0.0023 each). In the fusion `DEC_OFFSET=3` (fuse_evidence.py; decoder window j with classifier hop j+3): 4+4 ensemble 0.7731 -> 0.7752 (p 0.998), 0.7769 -> **0.7776-0.7778** (p 0.9997, a 0.3-0.5). **Default since 2026-10-04 (commit bf7243c, user: apply from the next experiments on, do not re-score old ones):** config `dnn.evidence_lookahead_s: 0.5` (run_experiment, rescore_weights, decision_targets, eval_external) and `DEC_OFFSET` default 3 in fuse_evidence; the classifier / fusion stamping stays 0.75. Results stored before that use 0.75 (about 0.002 lower H): compare new decoder runs against the 0.5-s reference 0.7624 / 0.7651 / 0.7632 / 0.7623 (seeds 0-3), not against 0.760; the summary column lookahead_s tells which stamping a result used. **End-to-end leakage test PASSED (2026-10-04, `experiments/decoder_lodo/leakage_test.py`, commit 64ea2b3; run with `GLIMPSE_THR=norm ../../.venv/bin/python leakage_test.py`):** inputs replaced by noise from tau on, first changed 128-Hz label measured: decoder with lookahead 0.5, direct classifier (5 s and 10 s) and fusion with DEC_OFFSET=3 all use 0.969 s (limit 1 s; lookahead 0.75 used only 0.719 s); negative controls (stamping 0.25 s too early, DEC_OFFSET=2, offline read-out, second zero-phase filter) are all flagged. So lookahead 0.5 / DEC_OFFSET=3 are rule-compliant; rerun the test after any change of timing, read-out or a new causal step. Caveat: `causal_ea` takes the channel set from `x.std(0) > 0` over the whole trial (= montage metadata, e.g. 1 absent channel in UHD); in submission code pass the channel list explicitly.
- `multiscale_decoder.py` (decoder evidence over the last L s from saved reconstructions, `--save-recon`) written; 1-s evidence reproduces the stored one exactly; per-step SNR on DTU 0.195 (1 s) / 0.329 (3 s) / 0.403 (5 s). 10-s classifier pretrain reaches 0.84-0.85 and 66.4 % 5-s... decision accuracy on AVGC (5-s classifier 62 %).
- **10-s direct classifier and multi-scale decoder evidence (2026-10-04 07:10):** direct classifier with a 10-s window (`--window-s 10`, pretrained `_w10`, exp `direct_clf__mat_reac_w10_ptSPK`), seed 0 alone: **H 0.770** (Acc 71.1, SDT 9.3; 5-s 0.756, 3-s 0.755), better on all 5 datasets, shortcut controls clean. In the ensemble (DEC_OFFSET=3): 4 dec + 4 clf5 + 1 clf10 = 0.7772 (p 0.998) / **0.7795** (p 0.9997, a 0.5) vs 0.7752 / 0.7778 without it; 4 dec + only 1 clf10 = 0.7772 at p 0.998 (one 10-s classifier replaces four 5-s ones; p 0.9997 gives no extra gain there). Multi-scale decoder evidence (3-s / 5-s correlation windows from saved reconstructions, `multiscale_decoder.py`): alone 0.752-0.757, added to the ensemble 0.772-0.777 = **no gain, dropped**. Queued: 10-s seeds 2, 3 (354, 355), 20-s seed 0 (356), fusions `logs/fuse_w10_s01.log`, `logs/fuse_adt.log` (ADT decoder env,onset,dom as extra decoders).
- **Window sweep and ensembles finished (2026-10-04 09:55, queue empty):** 10-s classifier 4 seeds alone 0.7700 / 0.7675 / 0.7681 / 0.7697 (mean 0.769; 5-s mean 0.759); **20-s is worse: 0.7626** (SDT 11.2 s) -> 10 s is the optimum of 3/5/10/20, 20 s dropped (never leak-tested, not needed). Ensembles (DEC_OFFSET=3; p 0.998 / p 0.9997): 4 dec + 4 clf5 = 0.7752 / 0.7778; + 2 clf10 = 0.7774 / **0.7814**; + 4 clf10 = 0.7781 / 0.7805; 4 dec + 4 clf10 only = 0.7791 (a 0.7) / 0.7804 (a 0.5). Plateau at 0.778-0.781: differences of 0.001-0.003 are selection noise (a, p_stay not nested). ADT decoder with env,onset,dom (scored at lookahead 0.5): alone 0.7562 / 0.7521, as extra decoders 0.7736-0.7771 = **no gain, dropped**. Still open: fused setup on real switches (needs fold/all-data models saved), nested a / p_stay.
- **Fused setup on REAL switches (2026-10-04 12:25; `eval_external_clf.py`, `fuse_external.py`, `logs/fuse_external.log`, `results/research/fuse_external.csv`; 2 dec + 2 clf per window, all fitted parameters from the corpus):** H stays 0.62-0.66 (corpus synthetic 0.78). AASD: decoder alone 0.634 (p 0.998) / 0.648 (p 0.95); 5-s classifier alone 0.648 / 0.658; fused with clf5 0.646-0.649 / 0.655-0.657; **10-s classifier is worse on AASD** (alone 0.632 / 0.642, fused 0.637 / 0.648-0.651). Carta: everything 0.636-0.645 with a classifier (decoder alone 0.624), 5 s = 10 s within noise. **p_stay 0.9997 is the worst value on both real sets** (AASD decoder 0.624, Carta decoder 0.597); on AASD H rises monotonically as p_stay falls to 0.95, on Carta flat for 0.95-0.998. Fusion gain on real switches ~+0.01 over the decoder, none over the 5-s classifier alone. So: synthetic-optimal choices (10-s window, p 0.9997) do not transfer; robust choice = 5-s classifiers (or 5 s + 10 s) with p_stay 0.998 or lower. All-5-dataset classifier models saved as `models/all5_direct_clf{_v2,_w10}_s{0,1}.pt`.
- **Spectral-tilt augmentation (aug variant `clustertilt`, p 0.5, exponent +-0.5):** decoder 0.761 / 0.762 vs 0.7624 / 0.7651 without = no gain; 10-s classifier with tilt 0.766 / 0.775 vs 0.770 / 0.768 (mean 0.7705 vs 0.769, inside seed noise); fusion 2 dec + 2 clf10 (`logs/fuse_tilt.log`): reference 0.7783 (p 0.998, a 0.5), tilt both 0.7774, tilt decoder only 0.7767, tilt classifier only 0.7792-0.7811 = **no reliable gain (<= +0.002, 2 seeds); not adopted**. Queue empty at 13:00.
- **Adversarial domain loss (2026-10-04, `--adv dataset|subject`, gradient reversal on time-pooled backbone features, lambda 0.05):** decoder seed 0: dataset-adversarial H 0.760, subject-adversarial 0.761 vs 0.762 without = no gain so far. The adversary sits at chance from the first epoch (dataset 0.25 of 5 classes, subject 0.02 of ~117), so either the features carry little domain information after z-score + causal EA or the probe is too weak; not yet checked with lambda 0. Remaining jobs 371, 373-376 (seed 1, 10-s classifier, fusion `logs/fuse_adv.log`).
- **Adversarial loss, update 15:10:** dataset-adversarial decoder seed 1 = 0.765 (reference 0.7651): no gain on either seed. Probe with lambda 0 (dataset): the adversary climbs slowly from 0.26 to 0.36-0.43 (chance ~0.25), so the pooled features hold only weak dataset information and lambda 0.05 removes it (adversary stays at 0.25) without changing H; subject probe reached 0.04-0.05 vs 0.02 before it was OOM-killed. Two more OOM kills (other users' jobs grew after my jobs had started; the RAM guard only checks at start); 373 requeued.
- **Adversarial loss, decoders final (15:40):** dataset-adversarial 0.7596 / 0.765, subject-adversarial 0.7605 / 0.764, reference 0.7624 / 0.7651 (seeds 0 / 1) = **no gain for the decoder**. 10-s classifier with dataset-adversarial loss still running (374 requeued after a third OOM kill at 15:26, 375); fusion in `logs/fuse_adv.log`.
- **Adversarial loss, final (2026-10-05 00:47 UTC, `logs/fuse_adv.log`):** 10-s classifier with dataset-adversarial loss 0.763 / 0.767 (reference 0.770 / 0.768); fusion 2 dec + 2 clf10 at p 0.998, a 0.5: reference 0.7783, adv-dataset decoder 0.7773, adv-subject decoder 0.7770, adv-dataset classifier 0.7766, both adversarial 0.7749 = **no gain anywhere, dropped** (lambda 0.05 only; pooled mean/std adversary). Queue empty.
- **Work started 2026-10-05 after reading the Terms ([[blind-test-sets]]); user asked for directions 1-4:**
  - (2) Gaze-controlled fold (AVGC = the only training set like both test sets): everything scores 0.70-0.72 there. The best classifier share is SMALLER than on the mean: a 0.15-0.3 gives AVGC 0.714-0.717, a 0.7 gives 0.703, classifier only 0.699, decoder only 0.7065; the 10-s classifier has no advantage on AVGC (0.696 / 0.683 vs 5-s 0.689 / 0.690). So prefer a = 0.3 for the submission and expect a test score nearer 0.70 than 0.78.
  - No-carry baseline (prep `mat_rea`): jobs 381-388 -> `logs/fuse_nocarry.log`, `logs/fuse_nocarry_5_10.log`.
  - (1) EA started from a fixed population prior (prep `mat_reap`, `build_ea_prior.py`, prior = mean covariance of Fuglsang): jobs 389-396 -> `logs/ea_prior_caches.log`, `logs/fuse_prior.log`.
  - (4) HMM bank over p_stay with online model averaging (`hmm_bank.py`, job 384a -> `logs/hmm_bank.log`, synthetic corpus + real AASD/Carta).
  - (3) Within-trial self-adapting linear decoder from pseudo labels (`adapt_within_trial.py`, job 384b -> `logs/adapt_within_trial.log`; includes an oracle-label upper bound).
- **Results of directions 1-4 (2026-10-05 12:00 UTC, queue empty; all with seeds 0,1, lookahead 0.5, DEC_OFFSET 3):**
  - **Cost of the test rule (no EA carry):** single models carry / no carry (`mat_rea`) / population prior (`mat_reap`): decoder 0.7624, 0.7651 / 0.7577, 0.7555 / 0.7534, 0.7573; 10-s classifier 0.7700, 0.7675 / 0.7565, 0.7595 / 0.7674, 0.7654; 5-s classifier 0.7564, 0.7604 / 0.7550, 0.7555 / 0.7533, 0.7567. Fusion 2 dec + 2 clf10 at p 0.998 (a 0.3 / 0.5): carry 0.7768 / 0.7783, no carry 0.7693 / 0.7683, prior 0.7728 / 0.7744; with 2 clf5 added: no carry 0.7709 / 0.7683, prior 0.7710 / 0.7716. So **compliant LODO level = 0.770-0.774** (carry was worth ~0.005-0.010).
  - **(1) EA prior:** helps the 10-s classifier a lot (+0.008, and removes an unstable AVGC fold 0.648 -> 0.699), not the decoder or the 5-s classifier; per fold it helps DTU (+0.007 to +0.012), UHD, KUL, but **USTC drops ~0.010** (prior is BioSemi-only Fuglsang, USTC is NeuroScan like the NJU test set). For a submission build the prior from all training data incl. USTC; the NeuroScan risk is untested.
  - **(4) HMM bank: no gain, dropped.** It behaves like a single p ~0.98-0.99: synthetic 0.764-0.767 vs single 0.998 = 0.7744, 0.9997 = 0.7775; real mean 0.646 vs 0.642 / 0.631. Real-switch H is flat (0.642-0.647) for p 0.95-0.998 and only 0.9997 is clearly worse, so **p_stay 0.998 is the robust choice**.
  - **(3) Within-trial self-adaptation: dropped.** Even with TRUE labels the within-trial linear decoder adds only +0.006 (0.7744 -> 0.7808; its own accuracy after 3 min is 61%); with pseudo labels beta 0.3 = 0.7738 (AVGC +0.008, UHD -0.012), beta 0.6 = 0.756.
  - (2) see above: on the gaze-controlled fold a = 0.3 is best.
- **Queued 2026-10-05 ~12:30 UTC (user: "tune everything"), all on compliant preps, results not yet seen:** 4 dec + 4 clf5 + 2 clf10 without carry (jobs 400-409 -> `logs/fuse_442_mat_reap.log`, `logs/fuse_442_mat_rea.log`); EA hyper-parameter search with the linear decoder as proxy (new `ea_n0` = weight of the EA start value in s, preps `eaI_*` / `eaP_*`, job 399 -> `logs/tune_ea.log`, `results/research/tune_ea.csv`; winners must be confirmed with the DNNs); decision grid a x p_stay (409a -> `logs/tune_decision.log`); one-factor-at-a-time DNN sweep on `mat_reap`, seed 0 only, experiments tagged `_hp-<setting>` (decoder: lr, weight decay, batch, dropout, noise, p_chan, p_cluster = jobs 410-423; 10-s classifier: lr, weight decay, batch, dropout, p_switch = 424-432), table with `python hp_summary.py`. Overrides: `--set key=value`, `--aug-set key=value`. One seed: differences below ~0.004 are noise, re-run promising settings with seed 1.
- **Results 2026-10-05 16:30 UTC:** compliant 4 dec + 4 clf5 + 2 clf10 (p 0.998 a 0.3 / 0.5; p 0.9997 a 0.3 / 0.5): `mat_reap` 0.7725 / 0.7728 / 0.7753 / 0.7738; `mat_rea` 0.7687 / 0.7705 / 0.7721 / 0.7723. More seeds do not beat 2 dec + 2 clf10 (0.7744); the prior stays +0.002-0.004 above identity. Decision grid (a 0.1-0.6 x p 0.99-0.9995, mat_reap 2+2+2): flat plateau 0.770-0.773 for p >= 0.998 and a 0.2-0.6, nothing to gain. **EA search with the linear proxy (`results/research/tune_ea.csv`): a light start value (n0 = 5 s) is best**: prior n0 5 = 0.7083, identity n0 5 = 0.7065, vs mat_reap 0.7051, mat_rea 0.6979, carry 0.7097 (not allowed); it also removes the USTC drop of the prior (0.750 vs 0.736; carry 0.754); tau 120 and shrink 0.3 do not help. DNN confirmation of `eaP_n5_t60` queued (jobs 417a-417h -> `logs/fuse_eaP_n5_t60.log`). DNN sweep so far (decoder, seed 0, base 0.7534): lr 2e-4 +0.0045, everything else (lr 1e-3, weight decay, batch 32) within +-0.001; lr 2e-4 seed 1 and lr 1e-4 queued.
- **User goal stated 2026-10-05:** get back to H >= 0.7814 with a compliant setup (4 dec + 4 clf5 + 2 clf10, p_stay 0.9997, a 0.3, EA prior with n0 = 5 s: jobs 417a-417m -> `logs/fuse_442_eaP_n5_t60.log`; compliant reference for that cell 0.7753 with n0 60, carry 0.7801). User said "tune everything, just do it all": also queued 5-s classifier sweep (433-441) and network-size sweeps with their own SparrKULee pretraining (decoder 442-445, 10-s classifier 446-449; `--set model_blocks=N model_kernels=K`; vlaai_ws shares weights across blocks, so blocks do not change the parameter count). Decoder sweep done so far (seed 0, base 0.7534): lr 2e-4 +0.0045, batch 32 +0.001, weight decay / lr 1e-3 / dropout 0.2 = 0, dropout 0.1 -0.001 (dropout lowers AVGC by 0.003-0.005).
- **Fast EA warm-up + full sweep results (2026-10-06 02:00 UTC):**
  - **`eaP_n5_t60` (population prior weighted as only 5 s of data) is the best compliant prep.** Single models vs n0 = 60: decoder 0.7590 / 0.7663 (was 0.7534 / 0.7573; seed 1 even beats the carry 0.7651), 5-s classifier 0.7561 (0.7533), 10-s classifier 0.7664 (0.7674, but its AVGC fold drops 0.699 -> 0.666 at seed 0). USTC no longer suffers (0.779-0.792 vs 0.757-0.776). **4 dec + 4 clf5 + 2 clf10, p 0.9997, a 0.3: 0.7782** (n0 60: 0.7753, identity: 0.7721, carry: 0.7801); at p 0.998 a 0.5: 0.7755. 2+2+2 gives 0.7761, 2 dec + 2 clf10 0.7755.
  - **Sweep (one seed, `python hp_summary.py`): the only real knob is the DECODER learning rate.** lr 2e-4 +0.0045 (seed 0) and +0.0041 (seed 1) = confirmed; lr 1e-4 +0.0048 (seed 0 only, AVGC -0.005). Everything else is noise or worse: decoder batch/weight decay/dropout/noise/channel-mask/cluster-drop within +-0.003 (dropout lowers AVGC); **10-s classifier: every change is worse** (-0.002 to -0.011, base is optimal); 5-s classifier: all within +-0.003. Queued: 4 decoders with lr 2e-4 on `eaP_n5_t60` + fusion (jobs 450a-450e -> `logs/fuse_lr2.log`), lr 1e-4 seed 1, network-size sweeps (442-449), rest of the 5-s sweep.
- **2026-10-06 08:00 UTC:** decoder lr 2e-4 did NOT transfer to `eaP_n5_t60`: 4 seeds 0.7609 / 0.7632 / 0.7623 / 0.7616, fusion 4+4+2 p 0.9997 a 0.3 = 0.7774 (lr 5e-4: 0.7782) -> keep lr 5e-4; the +0.004 on `mat_reap` was specific to that prep or noise. Rest of the sweep: 5-s classifier all within +-0.003 (batch 128 -0.008); decoder size (3 / 6 blocks, 32 / 128 filters) within +-0.002. **One new signal: 10-s classifier with 3 blocks (`--set model_blocks=3`, own pretrain `pretrain_spk_direct_clf_w10_b3_s0.pt`) 0.7728 vs 0.7674 (+0.0054, AVGC +0.012), one seed**; confirmation seed 1 and a 4+4+2 fusion with the 3-block 10-s classifier queued (jobs 451-454 -> `logs/fuse_b3.log`). Best compliant result stays 0.7782.
- **2026-10-06 11:20 UTC, strongest single model so far:** 10-s classifier with **3 blocks** (`--set model_blocks=3`, pretrain `pretrain_spk_direct_clf_w10_b3_s{0,1}.pt`) on `eaP_n5_t60`, seed 0 ALONE: **H 0.7818** (Acc 71.8, SDT 8.3 s, L 0.860; AVGC 0.710, KUL 0.812) vs 4 blocks 0.7664 / 0.7649. Single-decision accuracy is unchanged (0.712 vs 0.714): the gain is faster switch detection. Controls clean. On `mat_reap` it was +0.0054 (0.7728); 6 blocks -0.002, 128 filters -0.005. ONE seed: confirmation running (jobs 453a seed 1 on eaP_n5_t60, 454a seed 1 on mat_reap, 456 fusion 4 dec + 4 clf5 + 2 clf10-b3 -> `logs/fuse_b3.log`). Not yet tried: 3 blocks (or 2) for the 5-s classifier and the decoder on eaP_n5_t60. Decoder lr 1e-4 seed 1 on mat_reap +0.005 (not pursued: lr 2e-4 did not transfer).
- **NEW BEST COMPLIANT, user's target reached (2026-10-06 13:40 UTC, `logs/fuse_b3.log`):** 4 decoders + 4 five-second classifiers + 2 ten-second classifiers with **3 blocks**, prep `eaP_n5_t60`: p 0.9997 a 0.3 = **0.7813**, a 0.5 = **0.7829**; p 0.998 a 0.3 = 0.7784, a 0.5 = **0.7815** (AVGC 0.713-0.718, SDT 8.0-9.2 s). Previous compliant best 0.7782, old non-compliant record 0.7814. 3-block 10-s classifier alone: 0.7818 / 0.7740 (seeds 0 / 1; 4 blocks 0.7664 / 0.7649), on mat_reap 0.7728 / 0.7755 (4 blocks 0.7674 / 0.7654) = confirmed on 4 comparisons; the gain is faster switch detection (SDT ~8.2 s vs ~10 s), accuracy unchanged. 3 blocks do NOT help the 5-s classifier (0.7564 vs 0.7561). Queued: 2-block 10-s classifier (457), seeds 2, 3 of the 3-block one and fusions without / with clf5 (456a, 462-464 -> `logs/fuse_b3_only.log`, `logs/fuse_b3x4.log`). Component experiments: decoders `vlaai_ws__eaP_n5_t60_aug_cluster_onset_dom_xFUG_ptSPK[_s1.._s3]`, clf5 `direct_clf__eaP_n5_t60_ptSPKv2[_s1.._s3]`, clf10 `direct_clf__eaP_n5_t60_w10b3_ptSPK[_s1]`.
- **NEW BEST COMPLIANT (2026-10-06 14:10 UTC, `logs/fuse_b3_only.log`): 4 decoders + 2 ten-second 3-block classifiers, NO 5-s classifiers, prep `eaP_n5_t60`:** p 0.998: a 0.3 = **0.7882**, 0.5 = 0.7871, 0.7 = 0.7886; p 0.9997: a 0.3 = 0.7882, 0.5 = **0.7904**, 0.7 = 0.7885 (Acc 72.5-73.5, SDT 7.5-9.0 s, AVGC 0.707-0.717, UHD 0.78-0.79). Adding the four 5-s classifiers LOWERS it to 0.7784-0.7829: they dilute the stronger 10-s model. Robust choice now = p 0.998, a 0.3 -> 0.7882. Running: 2-block 10-s classifier, seeds 2, 3 of the 3-block one (`logs/fuse_b3x4.log`), and at the user's request 3-block 5-s classifiers x4 fused with 2 / 4 ten-second ones (jobs 465-469 -> `logs/fuse_c5b3_2.log`, `logs/fuse_c5b3_4.log`; the single 3-block 5-s model showed no gain).
- **NEW BEST COMPLIANT (2026-10-06 15:52 UTC, `logs/fuse_b3x4.log`): 4 decoders + 4 ten-second 3-block classifiers, no 5-s classifiers, prep `eaP_n5_t60`:** p 0.998: a 0.3 = 0.7894, 0.5 = 0.7924, 0.7 = **0.7949**; p 0.9997: a 0.3 = 0.7914, 0.5 = 0.7933, 0.7 = **0.7956** (Acc 72.6-73.8, SDT 6.9-8.7 s, AVGC 0.716-0.726). With the four 4-block 5-s classifiers added: 0.7815-0.7860 (worse). 3-block 10-s classifier alone, 4 seeds: 0.7818 / 0.7740 / 0.7937 / 0.7703 (mean 0.780, large seed spread); 2 blocks 0.7750 (one seed), so 3 blocks is the optimum. Best share shifted to the classifier (a 0.5-0.7) on the mean, AVGC flat across a. Components: decoders `vlaai_ws__eaP_n5_t60_aug_cluster_onset_dom_xFUG_ptSPK[_s1.._s3]`, classifiers `direct_clf__eaP_n5_t60_w10b3_ptSPK[_s1.._s3]`. Still running at the user's request: 3-block 5-s classifiers x4 fused with 2 / 4 ten-second ones (465-469). the method is documented in experiments/decoder_lodo/README_BEST.md (written 2026-10-06, replaces BEST_COMPLIANT_METHOD.md).
- **Number of decoders hardly matters now (2026-10-06 16:05 UTC):** 2 decoders + 4 ten-second 3-block classifiers: seeds 0,1 = 0.7900 / 0.7927 / 0.7941 (p 0.998, a 0.3 / 0.5 / 0.7), 0.7956 at p 0.9997 a 0.7; seeds 2,3 = 0.7874 / 0.7919 / 0.7942, 0.7963 at p 0.9997 a 0.7 -> same as 4 decoders (0.7894 / 0.7924 / 0.7949, 0.7956). **The 4 classifiers ALONE (a = 1.0) reach 0.7932 (p 0.998) / 0.7914 (p 0.9997), AVGC 0.727**: the decoders now add only ~0.001-0.002. Simplest strong submission candidate = 4 ten-second 3-block classifiers (+ optionally 2 decoders), p 0.998.
- **Real switches, first numbers for the new best components (2026-10-06 17:00 UTC, prep `eaP_n5_t60`, models trained on all 5 + Fuglsang):** 3-block 10-s classifier seed 0 alone: AASD H 0.628 (p 0.998) / 0.640 (p 0.95), Carta 0.632 / 0.644 = the SAME as the old 4-block 10-s classifier (0.632 / 0.634); decoders seeds 0, 1 (LDA): AASD 0.633-0.636, Carta 0.620-0.626 (old 0.634 / 0.624). **The +0.015 of the 3-block classifier on synthetic switches does not show on real switches so far.** Fusion of 4 + 4 pending (`logs/fuse_external_best.log`). Jobs occasionally die with rc 143 (SIGTERM from outside, cause unknown, 4 times on 2026-10-06): just requeue.
- **Best compliant method on REAL switches (2026-10-07 00:36 UTC, `logs/fuse_external_best.log`; 4 decoders + 4 ten-second 3-block classifiers, prep `eaP_n5_t60`, trained on all 5 + Fuglsang):** AASD H 0.632-0.653, Carta 0.599-0.648 over the whole grid = **no better than the older setups** (decoder + 5-s classifier, carry: AASD 0.646 / Carta 0.639 at p 0.998). At p 0.998: AASD decoder only 0.635, a 0.3 = 0.639, 0.7 = 0.637, classifier only 0.633; Carta 0.624 / 0.638 / 0.643 / 0.642. On AASD the decoder-heavy mix and a LOW p_stay are best (0.9997 worst everywhere); on Carta the classifier-heavy mix. **So the synthetic-LODO gain 0.78 -> 0.795 does not transfer to real switches at all**; a = 0.3-0.5 with p 0.99-0.998 is the robust compromise. User's two requested fusions with 3-block 5-s classifiers: + 2 clf10-b3 = 0.7769-0.7813, + 4 clf10-b3 = 0.7806-0.7857 (both below the version without 5-s classifiers, 0.7949). Queue empty.
- **Queued 2026-10-07 01:20 UTC (user asked for more augmentation of the classifier):** 10-s 3-block classifier on `eaP_n5_t60`, seeds 0 and 1 (bases 0.7818 / 0.7740), experiments `direct_clf__eaP_n5_t60_w10b3_ptSPK[_s1]_aug-<name>`: noise02, pclus05, covmix (existing knobs) and new ones in `direct_classifier.windows` (`--aug-set clf_lag_jitter=2`, `clf_p_feat_drop=0.2`, `clf_p_switch2=0.5` with `--p-switch 0.5`), jobs 470-481. Seed spread of this model is ~0.02, so only a gain on BOTH seeds counts; a winner should then be checked on real switches (eval_external_clf.py).
- **Classifier augmentation results (2026-10-07 07:30 UTC; 10-s 3-block on eaP_n5_t60, seeds 0 / 1, base 0.7818 / 0.7740):** sw2 (p-switch 0.5, half with a second switch) +0.007 / +0.002 = only one positive on both seeds; covmix +0.001 / +0.001; noise 0.2 0 / -0.002; featdrop -0.006 / +0.011; p_cluster 0.5 -0.011 / 0; **EEG-stimulus lag jitter +-2 samples HURTS badly -0.034 / -0.029** (timing precision matters). Queued: sw2 seeds 2, 3 + LODO fusion (482-484 -> `logs/fuse_sw2.log`) and sw2 on real switches x4 + fusion with the 4 decoders (4850-4859 -> `logs/fuse_external_sw2.log`; compare with `logs/fuse_external_best.log`).
- **Two-switch augmentation `sw2` confirmed on LODO, neutral on real switches (2026-10-07 14:12 UTC):** 10-s 3-block classifier with `--p-switch 0.5 --aug-set clf_p_switch2=0.5`, 4 seeds alone: 0.7887 / 0.7761 / 0.7944 / 0.7785 (base 0.7818 / 0.7740 / 0.7937 / 0.7703: +0.007 / +0.002 / +0.001 / +0.008, 4/4 positive, mean 0.7844 vs 0.7800). **Fusion 4 dec + 4 clf-sw2 (`logs/fuse_sw2.log`) = NEW BEST COMPLIANT LODO: p 0.998 a 0.3 / 0.5 / 0.7 / 1.0 = 0.7929 / 0.7981 / 0.7966 / 0.7951; p 0.9997 = 0.7914 / 0.7993 / 0.8003 / 0.7954** (was 0.7894 / 0.7924 / 0.7949 / 0.7932 and 0.7956; AVGC 0.718-0.738). On REAL switches (`logs/fuse_external_sw2.log`) no change: AASD p 0.998 a 0 / 0.3 / 0.7 / 1.0 = 0.635 / 0.639 / 0.638 / 0.632 (without sw2: 0.635 / 0.639 / 0.637 / 0.633), Carta a 0.3 = 0.639 (0.638). Experiments `direct_clf__eaP_n5_t60_w10b3_ptSPK[_sN]_aug-sw2`. Running: sw2 + featdrop (`swfd`, jobs 4860-4881 -> `logs/fuse_swfd.log`, `logs/fuse_external_swfd.log`).
- **sw2 + feature-drop (`swfd`) = no better than sw2 alone (2026-10-07 17:21 UTC):** classifier alone 0.7893 / 0.7845 / 0.7899 / 0.7800 (mean 0.7859 vs sw2 0.7844, plain 0.7800; seed 2 below plain). LODO fusion 4 dec + 4 clf (`logs/fuse_swfd.log`): p 0.998 a 0.3 / 0.5 / 0.7 / 1.0 = 0.7925 / 0.7961 / 0.7932 / 0.7949, p 0.9997 = 0.7880 / 0.7941 / 0.7960 / 0.7923 -> below sw2 (0.7929 / 0.7981 / 0.7966 / 0.7951; 0.8003), AVGC lower (0.697-0.723 vs 0.718-0.738). Real switches (`logs/fuse_external_swfd.log`, p 0.998, a 0.3 / 0.5): AASD 0.640 / 0.639, Carta 0.642 / 0.642 = within +-0.003 of sw2 and of no augmentation. **Decision: keep sw2 only; best compliant stays 4 dec + 4 clf10-b3-sw2 (0.7981 at p 0.998, a 0.5).** Queue empty 2026-10-08 01:30 UTC.
- **2026-10-28 pointer:** real-switch experiments A1-A4, 4-s evidence lag, HMM cross-fit, schedule-noise (best fusion 0.7975 +- 0.0015 over 5 schedules) and the eeg-thao stem_ms96 audit are in [[real-switch-findings]]; priorities in [[design-priorities]].
