# Already tried — do not repeat

Inventory date: 2026-10-26. Sources searched:

- `/root/disk_3tb/eeg-thao` (code, `results/`, `experiments/`, `models/`)
- `/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset` (`EXPERIMENTS_LOG.md` updated 2026-10-15, 408 stored results; `experiments/decoder_lodo/README_BEST.md`; `experiments/decoder_lodo/results/research/fuse_external.csv`; `experiments/external_data/{AASD,Carta2026}`)
- Unzipped snapshot `.../eeg-thao_unzipped/eeg-thao/` (v2 external CSV and `--mode external`, which the live workspace script no longer has)
- Agent transcripts under `/root/.cursor/projects/root-disk-3tb-eeg-thao/agent-transcripts` (the 2026-10-24 run is `d58e57c5-b97f-4f13-8baa-3eee5c1b79c5`; it produced v4)
- Sibling dirs under `/root/disk_3tb`. `eeg_challenge/` is a different challenge (ds0055xx). No extra sAAD scores there.

**DO NOT REPEAT** means: do not rerun the same method, block, loss, augmentation, fusion, or hyperparameter grid, including a rename. A method is free only if it is absent from this file and from `EXPERIMENTS_LOG.md` sections 3–5.

H is a fraction (0.85 = 85). Acc is percent. Train numbers are leave-one-dataset-out means of the five corpus sets unless a dataset is named. External numbers are real switches on AASD and Carta2026.

---

## A. Logit transformers (the runs agents will be tempted to redo)

All of these read the four direct-classifier seed logits only. They never see raw EEG. Seed `20270101`. Official scorer via `score.score_dataset`.

### A1. v1 causal encoder — DO NOT REPEAT

Code: `/root/disk_3tb/eeg-thao/ensemble_clf_transformer.py`
Checkpoints: `models/logit_tf/`
Results: `results/research/ensemble_clf_transformer.csv` (+ `_inner.csv`)
Command: `python ensemble_clf_transformer.py` (800 steps, batch 512, d_model 32, 4 heads, GELU pre-norm, absolute positions, causal mask).

| Model | What | Mean H | Mean Acc | External |
|---|---|---:|---:|---|
| `tf_raw_w16` | raw 4 logits, 4 s window, 1 layer | 0.8027 | 74.07 | not scored |
| `tf_raw_w48` | raw 4 logits, 12 s, 1 layer | 0.8108 | 76.74 | not scored |
| `tf_hp_w32` | causal 120 s high-pass of logits, 8 s, 1 layer | 0.8173 | 76.85 | not scored |
| `tf_hp_w48` | same features, 12 s, 2 layers. acc74 pick on every fold | **0.8208** | **78.09** | not scored |

acc74 and acc73 selected the same `tf_hp_w48` point (p_stay 0.996, scale 0.9) on every fold. Per fold H/Acc: AVGC 0.738/69.62, DTU 0.871/85.79, KUL 0.836/78.19, UHD 0.793/76.12, USTC 0.866/80.70.

### A2. v2 recipe — DO NOT REPEAT every named model

Code: `/root/disk_3tb/eeg-thao/ensemble_clf_transformer_v2.py` (workspace copy has **no** `--mode`)
External driver: `.../eeg-thao_unzipped/eeg-thao/external_train_v2.py` via `--mode external`
Checkpoints: `models/logit_tf_v2/` and unzipped `models/logit_tf_v2_external/`
Results: `results/research/ensemble_clf_transformer_v2.csv`, `_inner.csv`, and unzipped `ensemble_clf_transformer_v2_external.csv`
Log: `experiments/transformer_v2.log` (this run used `STEPS=1000`, not 800)
Train command: `STEPS=1000 BATCH=512 python ensemble_clf_transformer_v2.py`

Shared recipe, all four models: every causal position supervised, dataset-balanced batches, cosine LR, AdamW, EMA 0.995, input noise, seed dropout, time masking, post-switch loss weights, zero-init head, linear skip initialised to the mean of the four raw logits. HMM grid p_stay in {0.994, 0.996, 0.997, 0.998}, emission scale in {0.8, 0.9, 1.0}, chosen by acc74 on inner folds only.

| Model | Block / features | Mean H | Mean Acc | AASD H / Acc (preset 0.994, 0.8) | Carta H / Acc |
|---|---|---:|---:|---|---|
| `recipe_hp48` | hp120 features, d=32, 2 layers, window 48, no conv stem, lr 1e-3 | 0.8323 | 77.75 | 0.6229 / 51.13 | 0.6164 / 56.78 |
| `ms_w96` | 24-d multi-scale (raw, hp60, hp120, z60, 1 s slope, 2 s mean), d=64, window 96, lr 8e-4 | 0.8480 | 80.98 | 0.6128 / 50.98 | 0.5894 / 52.46 |
| `ms_w192` | same features, 48 s window, d=48 | 0.8304 | 82.45 | 0.6125 / 50.78 | 0.5768 / 51.77 |
| **`stem_ms96`** | depthwise causal conv (kernel 9, zero-init) + 4 dilated residual convs (dilation 1,2,4,8, kernel 5) in front of the `ms_w96` transformer | **0.8677** | **82.42** | 0.6158 / 51.54 | 0.5695 / 51.03 |
| `ens_all` | equal average of the four evidence streams | 0.8564 | 82.93 | 0.6183 / 51.19 | 0.5967 / 53.29 |

`stem_ms96` per fold is in `CURRENT_BEST.md`. acc74 mixture mean is H 0.8562 / Acc 82.75. **DO NOT REPEAT** `stem_ms96`, the dilated stem, the 24-d multi-scale feature stack, windows 48/96/192, or averaging the four recipes.

### A3. Image-split retrain of `stem_ms96` — DO NOT REPEAT

Code: `stem_ms96_image_split.py`
Checkpoints: `models/stem_ms96_image_split/stem_ms96__held-*.pt`
Results: `results/research/stem_ms96_image_split.csv`, `_manifest.csv`, `stem_ms96_image_split_rescore_old_hmm.csv`
Log: `experiments/stem_ms96_image_split.log`
Command: `STEPS=1000 python stem_ms96_image_split.py`

Same network, loss, seed, 1000 steps, and features. Only the split changes: one validation subject from each of the other four datasets (`no_decoder_splits.csv`), no refit. Mean H **0.8627**, Acc **82.50** for both `image_train_emis` and `image_all4_emis`. Old vs new HMM rescore does not change the conclusion. Fuglsang and SparrKULee were **not** in this split (no 4-seed logit streams). AASD and Carta were not scored here.

### A4. v3 gated local attention — DO NOT REPEAT

Code: `ensemble_clf_transformer_v3.py`
Checkpoints: `models/logit_tf_v3/gate_local__screen-*.pt`, `gate_robust__screen-AVGC.pt` (robust was only finished for AVGC)
Results: `results/research/transformer_v3_screen.csv`, `transformer_v3_screen_avgc.csv`
Logs: `experiments/transformer_v3_screen.log`, `transformer_v3_screen_avgc.log`
Command that actually finished: `MODE=screen STEPS=1000 python ensemble_clf_transformer_v3.py`
`MODE=nested` was **not** run (no `ensemble_clf_transformer_v3.csv`). The screen already scored all five held-out sets, so rerunning nested on the same blocks is a repeat.

| Model | What | Mean H | Mean Acc |
|---|---|---:|---:|
| `gate_local` | 6 groups × 4 seeds, per-seed projection, causal depthwise conv, same-time attention mix across seeds, zero-init dilated stem, **local causal attention with relative bias span 6.25 s**, 2 layers, d=64, stem_ms96 loss | 0.8631 | 81.97 |
| `gate_robust` | same block plus label smoothing, cross-window mixup, per-channel gain, mild group-DRO, CORAL weight 0.005 across train datasets, temporal-smoothness penalty | AVGC only: H 0.7963, Acc 74.08 (worse than `gate_local` AVGC 0.8015 / 74.51) | |
| `ens_gate_local_stem` | average of `gate_local` and `stem_ms96` evidence | 0.8676 | 82.75 |

Per-fold `ens_gate_local_stem` H/Acc: AVGC 0.8023/75.10, DTU 0.8971/90.20, KUL 0.8781/82.93, UHD 0.8414/79.36, USTC 0.9193/86.18. Mean H does not beat `stem_ms96`. No external score. **DO NOT REPEAT** seed-gating, local relative attention, CORAL-on-logits, mixup, group-DRO, or this ensemble.

### A5. v4 local attention and 30 s high-pass — DO NOT REPEAT

Code: `ensemble_clf_transformer_v4.py`
Checkpoints: `models/logit_tf_v4/rel_ms.pt`, `rel_hp.pt` (`linear_*` have no weights)
Results: `results/research/ensemble_clf_transformer_v4_external.csv`, `ensemble_clf_transformer_v4_hp_lodo.csv`
Logs: `experiments/transformer_v4_external.log`, `transformer_v4_hp_lodo.log`
Commands: `MODE=external STEPS=800 python ensemble_clf_transformer_v4.py` and `MODE=lodo_hp python ensemble_clf_transformer_v4.py`

| Model | What | Train mean H / Acc | AASD H / Acc | Carta H / Acc |
|---|---|---|---|---|
| `linear_mean` | mean of the four raw logits, no new net | 0.7963 / 72.95 | 0.6417 / 52.91 | 0.6454 / 59.28 |
| **`linear_hp30`** | causal one-pole high-pass, tau **30 s**, then the mean. External winner. Train loser | 0.7899 / 69.21 | **0.6422 / 53.12** | **0.6520 / 60.54** |
| `rel_ms` | local causal attention, relative bias 6.25 s, dilated stem, stochastic depth 0.1, d=48, window 96, dropout 0.2. Skip = mean of four logits | not on the 5-fold table | 0.6214 / 51.25 | 0.6065 / 54.48 |
| `rel_hp` | same block on 30 s high-passed logits, anchor penalty 0.25 toward that mean, dropout 0.25, drop-path 0.15 | not on the 5-fold table | 0.6334 / 52.12 | 0.6229 / 57.04 |

External rows above are the pre-specified HMM (p_stay 0.994, scale 0.8). The extra grid point p_stay 0.998 / scale 1.0 is already in the CSV and is not better on AASD accuracy (Acc stays 50.9–52.4). **DO NOT REPEAT** tau-30 high-pass, stochastic depth, or another relative-bias local-attention block on these logits.

---

## B. Decision layer on the same four logits — DO NOT REPEAT

These do not train a transformer. They reweight or filter the four seed logits and retune the HMM. All causal. Held-out labels are not used to pick the fold's hyperparameters.

| Script | What | Mean H | Mean Acc | Result file |
|---|---|---:|---:|---|
| `ensemble_clf.py` | equal / inv-variance / single-seed average of 4 classifiers | equal **0.7942** / 73.05 | | `ensemble_clf.csv` |
| `ensemble_clf_meta.py` | LDA, inv-std, median, logistic regression stack | LDA 0.7947 / 73.05; logreg 0.767 / 73.25 | | `ensemble_clf_meta.csv` |
| `ensemble_clf_nested.py` | families `lda4` and `dyn` (logits + 1 s slope + 1 s mean); transforms `id`, `hp60`, `z60`, `sm1`; p_stay and emission scale grid | nested 0.8065 / 72.64; oracle (illegal as a selector) 0.8134 / 72.63; baseline lda4 at 0.998 is 0.7965 / 73.01 | | `ensemble_clf_nested.csv`, `_inner.csv`, `_pareto.csv` |
| `ensemble_clf_hpfrac.py` | fractional high-pass beta in {0, 0.25, 0.5, 1.0} at tau 60/120/240 s, plus slope kicks {0, 0.05, 0.1, 0.2} on the identity | joint 0.8068 / 73.03; maxH 0.8089 / 72.71. Winner variants are `hp120b100` and `hp240b100`, kick 0 | | `ensemble_clf_hpfrac.csv` |

Per-fold nested `lda4/hp60` H/Acc: AVGC 0.744/68.17, DTU 0.823/72.98, KUL 0.838/75.86, UHD 0.801/73.55, USTC 0.826/72.62. **DO NOT REPEAT** any of these transforms, kicks, or stacks. Full 60 s high-pass was already known to push H above 0.8 while dropping Acc below 73.

---

## C. Front-end, preprocessing, loss, augmentation (through 2026-10-15)

Exhaustive row-level log: `/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/EXPERIMENTS_LOG.md` (sections 4 and 5, 408 results). Best legal stack write-up: `experiments/decoder_lodo/README_BEST.md`. Code for that stack lives in **both** `/root/disk_3tb/eeg-thao/` and `experiments/decoder_lodo/` (`decoders.py`, `direct_classifier.py`, `prepare_data.py`, `build_ea_prior.py`, `build_glimpse.py`, `fuse_general.py`, `fuse_external.py`, `score.py`, `config.yaml`).

Headline path (mean LODO H): linear baseline 0.682 → `vlaai_ws` 0.715 → cluster aug + onset 0.723 → causal Euclidean alignment → Fuglsang 0.746 → `dom` target 0.754 → SparrKULee pretrain 0.759 → LDA evidence weights 0.761 → direct classifier fusion → drop the 5 s classifier, 10 s × 3 blocks × 4 seeds → `sw2` aug **0.798** (p_stay 0.998, classifier weight 0.5) or 0.800 at p_stay 0.9997. On real switches that stack stays about **H 0.64** and Acc 52–54 (AASD) / 57–60 (Carta). Later front-end changes moved synthetic-switch H and did not move real-switch H.

### C1. Architectures — DO NOT REPEAT

| Method | Where | Measured H | Flag |
|---|---|---:|---|
| Organiser linear Wiener filter `wf__mat` | baselines | 0.682 | baseline only |
| `vlaai_ws` shared-weight conv, **3 blocks**, 64 filters, window 10 s | `decoders.py`, `direct_classifier.py` | decoder ~0.76; classifier seeds 0.776–0.794; fusion **0.798** | this is the current front-end. Do not retune block count |
| `vlaai_ws` 2 / 4 / 6 blocks, 32 or 128 filters | same | 2 blocks 0.775; 4 blocks 0.766; 6 blocks worse than 4 by ~0.002; 32 and 128 filters −0.005 to −0.006 | DO NOT REPEAT |
| `deformer`, `deformer_ws`, `lsm_cnn`, `ssmamba`, `simple_cnn` | decoder benchmark | 0.702–0.710 | DO NOT REPEAT |
| CCA, FBCCA, per-subject WF, Riemann WF, forward WF | decoder benchmark | 0.628–0.683 | DO NOT REPEAT |
| Original VLAAI (skip off, skip on, 10-epoch warm-up) | decoder benchmark | 0.505–0.712 | DO NOT REPEAT |
| ADT transformer decoder, with and without SparrKULee / onset+dom | `decoders.py` | 0.739–0.756, below `vlaai_ws` | DO NOT REPEAT as a reconstruction decoder |
| HappyQuokka (~11M) | `decoders.py` | 0.734–0.745 | DO NOT REPEAT |
| FBCSP and Riemann left/right from EEG | spatial decode | 49–53% held-out | DO NOT REPEAT |
| Transformer **body of the direct classifier** | `tf_directnet_lodo.py`, `decoders.TFDirect` | **measured 2026-10-29**: swapΔ=0.0255; AVGC chance-probe mean_loss **0.6958** ≥0.685 @50 steps → abandoned. No LODO/external | DO NOT REPEAT. Was the last untried C1 front-end; now closed. `multi_agent_opt/results_tf_directnet.md` |

### C2. Preprocessing — DO NOT REPEAT

Prep key of the legal stack: **`eaP_n5_t60`** (`prepare_data.py`, `build_ea_prior.py`, `config.yaml`).

Order: organiser `pre_*` files (Butterworth order-2, 1–9 Hz, filtfilt, bad-channel repair, 32 Hz) → 58-channel union `b64ustc58` with **zeros for missing channels** → CAR on present channels → drop 1 s at each end → one full-trial z-score on EEG and on envelopes → causal Euclidean alignment restarted every trial, population covariance from Fuglsang, initial weight 5 s, forgetting 60 s, 10% shrink to identity.

| Variant | H vs the reference of that day | Flag |
|---|---|---|
| `mat` organiser files | 0.715 vs self-built causal `py` 0.700 | keep `mat` |
| Running z-score `*_rz` | 0.694 | DO NOT REPEAT |
| Full-trial FFT, FFT 1–15 Hz | 0.700–0.710, and FFT is illegal at test | DO NOT REPEAT |
| 3-band spectral EQ `beq` | 0.673–0.694 | DO NOT REPEAT |
| Gamma envelope channels 30–40 Hz, and 35–48 Hz | 0.727–0.729 or chance on DTU/KUL | DO NOT REPEAT |
| Delta/theta phase channels | 0.692 | DO NOT REPEAT |
| Causal time whitening AR(6) `tw`, half-strength `tw05`, spectral shape `tws` | 0.751–0.755 vs 0.754 | DO NOT REPEAT |
| EA forget 30 s or 120 s, shrink 0.3 | no gain (`tune_ea.py`, 16 configs) | DO NOT REPEAT |
| EA carried across trials of one subject `mat_reac` | +0.007 to +0.010 | **illegal at test**. DO NOT REPEAT |
| EA from identity, 60 s | 0.772 | worse than `eaP_n5_t60` |
| Population init, weight 60 s `mat_reap` | small, hurts USTC | DO NOT REPEAT |

Speech features already in the stack (`build_glimpse.py`): `env`, causal `onset` (positive first difference), causal `dom` (relative envelope dominance, 25 ms smooth, 1–9 Hz). **DO NOT REPEAT** adding `glim`/`mask`, soft mask, `orel`, `odiff`, `ediff`, or delta/theta bands. Dropping `env` and keeping onset+dom tied the 3-target decoder; it does not beat it. InfoNCE as the reconstruction loss scored 0.703.

### C3. Losses — DO NOT REPEAT

| Loss | Where | Result |
|---|---|---|
| Contrastive Pearson (−corr attended + corr unattended), mean over env/onset/dom | decoder `fit` | in the 0.76 decoder |
| Binary cross-entropy on the antisymmetric logit `g(EEG,A,B) − g(EEG,B,A)` | `direct_classifier.py` | in the 0.78–0.79 classifier |
| End-to-end CRF / sequence loss, 40 s with fake switches | decoder training | 0.757–0.762, no gain |
| Dataset-adversarial and subject-adversarial | classifier HP search | −0.004 or inside seed noise |
| Label smoothing, mixup, group-DRO, CORAL, TV smoothness | v3 `gate_robust` only | worse on AVGC. DO NOT REPEAT on the logit net |
| Post-switch reweighting inside the logit BCE | v2 recipe | already inside `stem_ms96` |

### C4. Augmentations — DO NOT REPEAT

| Augmentation | Result |
|---|---|
| Cluster channel drop (3–8 neighbouring electrodes, 30%), per-channel drop 0.2, gain ±25%, Gaussian 0.1, zero-fill start of window 10% | kept for the decoder (`aug_variants.cluster`) |
| `sw2`: 50% of windows swap the two speakers once; half of those swap back after ≥2 s (`--p-switch 0.5 --aug-set clf_p_switch2=0.5`) | **kept**. +0.001 to +0.008 on all 4 classifier seeds. Fusion 0.798 |
| `featdrop` (drop one speech feature) | inconsistent across seeds; fusion worse than `sw2` alone |
| `sw2`+`featdrop` | mean +0.006 alone, worse once fused |
| `noise02`, `pclus05` (cluster drop 50%) | no gain or −0.011 |
| `covmix` covariance mix | +0.001 to +0.002, not used |
| Spectral tilt | inside seed noise |
| EEG–speech lag jitter ±60 ms `lag2` | **−0.03**. DO NOT REPEAT |
| v2 input noise, time mask, seed dropout | already inside the logit transformer recipe |

### C5. Data added at train time — DO NOT REPEAT

| Data | Result | Flag |
|---|---|---|
| Fuglsang 2020 as extra train and as the EA population prior | +0.004, in the stack | keep, do not retune the mix |
| SparrKULee pretrain (match/mismatch), then fine-tune. Classifier: 25 epoch min, 80 max. Decoder Pearson on env+onset | +0.005 | keep the existing checkpoints `models/pretrain_spk_*`. Do not run another pretrain schedule |
| SparrKULee mixed at 25% of each batch | no gain | DO NOT REPEAT |
| Broderick 2018 | 0.740, hurts KUL-AV-GC | DO NOT REPEAT. Also already concatenated into the v2 external train pool |
| AASD added to decoder training | LODO 0.742, hurts KUL and DTU | **DO NOT REPEAT. AASD is an evaluation set for this effort** |
| Carta added to decoder training | LODO 0.734 | **DO NOT REPEAT. Carta is an evaluation set** |
| Learning-curve subsamples 25/50/75/100% of the corpus | 0.719 / 0.733 / 0.741 / 0.742 | corpus is saturated |

### C6. Fusion and HMM — DO NOT REPEAT

| Method | Result |
|---|---|
| 4 decoders + 4 classifiers, weight 0.5, p_stay 0.998, `DEC_OFFSET=3`, LDA target weights, emissions from train only | **H 0.798** mean. Per dataset at that point (README_BEST §9.1): AVGC 0.726, DTU 0.842, KUL 0.809, UHD 0.791, USTC 0.823 |
| Same with p_stay 0.9997 | H 0.800 on synthetic switches, **worst** on real switches. DO NOT USE 0.9997 |
| Classifier weights 0.3 / 0.5 / 0.7 / 1.0 and p_stay in {0.95, 0.99, 0.998, 0.9997} on AASD and Carta | already in `fuse_external.csv`. Best swept AASD is H 0.6525 Acc 54.03. Acc never reaches 60 |
| Adding the 5 s classifier, a 3 s classifier, a 20 s classifier, a second decoder family (ADT), or multi-scale decoder evidence (3 s, 5 s) | 0 to −0.01 | DO NOT REPEAT |
| Learned fusion weights (sequence loss) vs fixed 0.5 | tied | DO NOT REPEAT |
| Reliability weights, temperature, Student-t emissions, running z-score of evidence, CUSUM, multi-rate HMM, mode-switching Markov model | +0.003 or worse (`integrators.py`, `hmm_bank.py`, `msm.py`, `reliability.py`) | DO NOT REPEAT |
| Within-trial self-training on pseudo-labels | −0.018 to −0.001 (`adapt_within_trial.py`) | DO NOT REPEAT. Also illegal if it uses other trials |
| Hyperparameter search of lr, weight decay, batch, dropout, noise, switch rate on the 10 s classifier | −0.002 to −0.011 (`hp_summary.py`) | the published point (AdamW 5e-4, wd 5e-4, batch 64, 250 steps/epoch, max 40, patience 6, min 10 stable epochs) is already the best |

Real-switch table for the legal stack is README_BEST §9.5 and `experiments/decoder_lodo/results/research/fuse_external.csv`. Names `RB`, `RS`, `RSF`. At the safe p_stay 0.998 the scores sit near H 0.63–0.65 with Acc in the low-to-mid 50s (AASD) and high 50s (Carta). `stem_external.csv` `mean4` is H 0.633 / Acc 52.11 on AASD and H 0.643 / Acc 58.73 on Carta (p_stay 0.998).

### C7. Splits and checks already done

- Nested inner-dataset selection (acc74 at 74, acc73 at 73, joint at inner Acc 73, maxH). DO NOT REPEAT the threshold search.
- Subject-level image split of `stem_ms96` (section A3).
- Shortcut checks: EEG=0, EEG rolled by half a trial, drop frontal channels, envelope-only. Classifier is not using a shortcut.
- Lookahead audit `leakage_test.py`: 0.969 s. Do not add a non-causal stage.
- `mat_reac` state carried across trials. Illegal. DO NOT REPEAT.

---

## D. Explicit non-results (so they are not "new")

- No system on disk has mean train H >= 0.85 **and** Acc >= 78 **and** AASD Acc > 60. Train mean is solved by `stem_ms96`. AASD Acc is not.
- `acc1` on AASD for every logit model is 0.50–0.54. A new head, HMM, high-pass, or ensemble on those logits cannot create Acc 60.
- Training on AASD or Carta to chase the external number leaks the test this ledger is scoring. It was also already tried for LODO and did not help.
- v3 `MODE=nested` did not finish, but `MODE=screen` already measured `gate_local` on all five datasets. Do not spend a GPU on the same block.

---

## E. Agent addenda (2026-10-27) — DO NOT REPEAT

- **channel-mask + XY FiLM** on vlaai_ws 10 s (`eaP_n5_t60`, sw2): LODO mean H 0.7669 Acc 70.35; AASD 0.6371/52.57; Carta 0.6420/59.31. Below stem_ms96 and linear_hp30. `multi_agent_opt/results_channel_mask.md`.
- **SWA last-200** fine-tune of `stem_ms96` from published checkpoints (equal avg of last 200 of 1000 steps; evaluate SWA weights): LODO mean H 0.8700 Acc 82.61 (AVGC 0.8015/74.91, UHD 0.8553/80.09); AASD 0.6202/51.08; Carta 0.5507/48.68. UHD H cleared 0.85; external Acc not improved. Do not rerun SWA/Polyak averaging on this stem. `multi_agent_opt/results_swa.md`.
- **MixStyle** on vlaai_ws direct classifier (shared perm/lambda across antisymmetric pair; eaP_n5_t60, 10 s, sw2): LODO mean H 0.7815 Acc 72.30; AASD p0.998 H 0.6307 Acc 51.71; Carta p0.998 H 0.6300 Acc 58.18. Below stem_ms96 and linear_hp30. Do not rerun MixStyle on this front-end. `results/direct_clf__eaP_n5_t60_w10b3_ptSPK_mixstyle.json`, `results/research/technique_mixstyle_external.csv`.
- **label-free causal TTA** on frozen stem_ms96 / SWA all5 (scale-only online entropy-min on the logit; mild feature-stat and aggressive feat+affine ablated): LODO mean H 0.8684 Acc 82.44 (AVGC 0.7955/74.08); AASD 0.6199/51.08; Carta 0.5498/48.82. Negligible vs stem/SWA; AASD Acc unchanged vs SWA and below linear_hp30. Do not rerun entropy-min / feature-stat TTA on this stem. `multi_agent_opt/results_tta.md`.
- **RSC** (Representation Self-Challenging) fine-tune of stem_ms96 (mute top-33% gradient feature channels w/ p=0.5; init published held-* ckpts; EMA eval): LODO mean H 0.8702 Acc 82.51 (AVGC 0.8042/75.12, UHD 0.8500/79.79); AASD 0.6196/51.13; Carta 0.5481/48.48. Negligible vs SWA; AASD Acc unchanged vs SWA, below linear_hp30. Do not rerun RSC on this stem. `multi_agent_opt/results_rsc.md`.
- **IRM** across 5 train datasets as environments (fine-tune stem_ms96 from SWA; IRMv1 dummy-scale penalty; smoke λ∈{0.1,1,10}→λ=0.1; restore causal mask corrupted by SWA buffer avg; EMA eval): LODO mean H 0.8695 Acc 82.35 (AVGC 0.8095/75.12 best AVGC H so far, UHD 0.8530); AASD 0.6207/51.01; Carta 0.5426/48.04. AASD Acc flat vs SWA/RSC; no new target met. Do not rerun IRM on this stem. `multi_agent_opt/results_irm.md`.
- **AVGC loss upweight 2.5×** fine-tune of stem_ms96 from IRM (smoke 2.5 vs 3.0→2.5; other datasets weight 1; 800 steps; EMA eval): LODO mean H 0.8632 Acc 81.92 (AVGC 0.8025/75.09 worse than IRM; UHD 0.8434 lost the 0.85 bar); AASD 0.6213/51.21; Carta 0.5404/47.88. No new target met; AASD Acc flat. Do not rerun AVGC sample-loss upweight on this stem. `multi_agent_opt/results_avgc_upweight.md`.

- **KD ens_all → vlaai_ws 10 s** (α=0.6, T=2.0, eaP_n5_t60, sw2, SPK init; teachers stem_ms96+ms_w96+ms_w192+recipe_hp48 held-*): LODO mean H 0.7595 Acc 68.97 (AVGC 0.7131/66.32); AASD 0.6408/52.71; Carta 0.6343/58.66. Do not rerun this KD recipe. `/root/disk_3tb/eeg-thao/multi_agent_opt/results_kd.md`.
- **soft-H aux** fine-tune of stem_ms96 from IRM (BCE + λ·(1−soft harmonic mean of sens/spec); smoke λ∈{0.1,0.3,1.0}→λ=0.1; 800 steps; EMA eval): LODO mean H 0.8671 Acc 82.15 (AVGC 0.8088/75.29 ≈ IRM; UHD 0.8475 lost the 0.85 bar); AASD 0.6226/50.94; Carta 0.5395/47.82. No new target met; AASD Acc flat. Do not rerun soft-H / soft Dice-style H surrogate on this stem. `multi_agent_opt/results_softH.md`.
- **logit temperature / Platt calibration** on frozen stem_ms96 (+ SWA), fit on LODO non-held-out folds only, fixed HMM 0.994/0.8 + hp30: temperature ≈ invariant under Gaussian emissions; Platt stem LODO mean 0.8619/83.37; best AASD Acc 51.38. Does not move AASD Acc toward 60. Distinct from prior HMM-emission “temperature” in §C6. Do not rerun logit T/Platt on this stem. `multi_agent_opt/results_calib.md`.
- **vlaai_ws 15 s alone** (eaP_n5_t60, 3 blocks, sw2, partial SPK w10_b3 init; 4×250 steps): LODO mean H 0.7693 Acc 71.82; AASD 0.6290/51.48 (hp30 0.6310/52.02); Carta 0.6245/58.13 (hp30 0.6526/60.43). Below stem_ms96 and linear_hp30 on AASD Acc. Do not rerun 15 s vlaai_ws alone. `multi_agent_opt/results_phase2.md`.
- **ATCNet** 10 s direct-classifier backbone (Altaheri stack adapted: Conv1d front + MHSA + TCN residuals → dense (B,W,d); MI EEGNet electrode-collapse front stayed at chance in smoke). Prep eaP_n5_t60, sw2, random init, max 40 / min 10 ×250 (gru_ffn schedule). LODO mean H 0.7425 Acc 68.06; AASD 0.6290/52.05 (hp30 0.6346/52.30); Carta 0.6310/58.80 (hp30 0.6377/59.75). Below stem_ms96 / SWA / linear_hp30. Do not rerun ATCNet on this front-end. `multi_agent_opt/results_atcnet.md`.
- **EEG Conformer** 10 s direct-classifier backbone (Song et al. dense Macaron: Conv1d patch embed over EEG+speech → ½FFN→MHSA→depthwise-conv→½FFN ×4; no electrode-collapse; swap-test Δ=0.0144). Prep eaP_n5_t60, sw2, random init, max 40 / min 10 ×250, batch 64, d=64. LODO mean H 0.7471 Acc 67.93; AASD 0.6398/52.65 (hp30 0.6401/52.81); Carta 0.6255/58.66 (hp30 0.6358/59.27). Below stem_ms96 / SWA / linear_hp30. Do not rerun EEG Conformer on this front-end. `multi_agent_opt/results_conformer.md`.
- **EEG-TCNet** 10 s direct-classifier backbone (Ingolfsson: Conv1d front + TCN residuals, no MHSA; eaP_n5_t60, sw2). SwapΔ=0.0077; abandoned: AVGC fold train loss 0.6907 after 50 steps (~chance). No LODO/external scores. Do not rerun EEG-TCNet on this front-end. `multi_agent_opt/results_tcnet.md`.
- **SWA-2000 / last-400** fine-tune of stem_ms96 from SWA-1000 ckpts (equal avg; causal mask restored; distinct from SWA last-200 of 1000): LODO mean H 0.8596 Acc 81.81 (AVGC 0.8015/75.12, UHD 0.8388 lost 0.85); AASD 0.6199/50.98; Carta 0.5432/48.14. Regressed vs SWA-1k 0.8700/82.61; AASD Acc flat. Do not rerun longer SWA/Polyak on this stem. `multi_agent_opt/results_swa2k.md`.
- **learned channel-adjacency mixer** on vlaai_ws 10 s (residual A=I+sym(Δ), Δ zero-init; eaP_n5_t60, sw2, SPK w10_b3 init; NOT channel-mask+XY FiLM; swapΔ=0.0073): LODO mean H 0.7707 Acc 70.55; AASD 0.6381/52.71 (hp30 52.81); Carta 0.6421/59.33 (hp30 0.6513/60.46). Below stem_ms96/SWA/IRM; AASD Acc below linear_hp30. Do not rerun graph-adj / electrode adjacency mix on this front-end. `multi_agent_opt/results_graph.md`.
- **IRM↔SWA logit blend** (frozen held-*/all5; w∈{0,0.25,0.5,0.75,1}; primary w=0.5; HMM 0.994/0.8; also blend+hp30): LODO mean H 0.8725 Acc 82.58 (AVGC 0.8100/75.14 tiny vs IRM; UHD 0.8580 kept ≥0.85); AASD 0.6179/50.94 (hp30 ~51); Carta 0.5461/48.38. AASD Acc flat below linear_hp30 53.12. Do not rerun IRM/SWA weight blends or hp30-on-blend of these two. `multi_agent_opt/results_irm_swa_fuse.md`.
- **ShallowConvNet** 10 s direct-classifier backbone (Schirrmeister dense Conv1d+square+log; eaP_n5_t60, sw2, random init; swapΔ=0.1068; AVGC chance-probe 0.6835 then full LODO). LODO mean H 0.7537 Acc 68.64 (AVGC 0.6764/64.01); AASD 0.6365/52.39 (hp30 52.76); Carta 0.6325/58.23 (hp30 59.50). Below stem_ms96/SWA/IRM↔SWA fuse and linear_hp30. Do not rerun ShallowConvNet on this front-end. `multi_agent_opt/results_shallow.md`.
- **DeepConvNet** 10 s direct-classifier backbone (Schirrmeister stacked Conv1d+ELU; eaP_n5_t60, sw2, random init; swapΔ=0.0120). Abandoned: AVGC fold mean loss 0.6945 after 50 steps (~chance). No LODO/external scores. Do not rerun DeepConvNet on this front-end. `multi_agent_opt/results_deep.md`.
- **multi-rate envelope banks** on vlaai_ws 10 s (causal EWMA τ∈{0,50,200,800} ms + onset, F=5; eaP_n5_t60, sw2, random init; distinct from delta/theta spectral bands and from onset+dom noenv decoder ablation): LODO mean H 0.7550 Acc 69.01 (AVGC 0.6925/65.16); AASD 0.6405/52.97 (hp30 52.88); Carta 0.6373/58.08 (hp30 58.99). Below fuse/stem and linear_hp30. Do not rerun this temporal multi-rate speech bank on this front-end. `multi_agent_opt/results_phase2_gap.md`.
- **ALiBi** full causal attention bias on stem_ms96 (replace absolute `pos`; fixed slopes; dilated stem + ms/w96 kept; SWA init; EMA eval; NOT v3/v4 local rel-bias 6.25s, NOT RoPE): LODO mean H 0.8320 Acc 75.04 (AVGC 0.7654/69.15, UHD 0.8117 lost 0.85); AASD 0.6308/51.53 (hp30 51.34); Carta 0.6384/58.11 (hp30 59.10). Below SWA/IRM↔SWA fuse and linear_hp30; no open target moved. Do not rerun ALiBi / abs-pos→ALiBi swap on this stem. `multi_agent_opt/results_stem_alibi.md`.
- **frozen-GRU** residual on SWA stem (freeze encoder; softplus gate+GRU+head): LODO mean H 0.8702 Acc 82.62; AVGC 0.8036/75.12; AASD 0.6181/51.03; Carta 0.5457/48.31. No new target. Do not rerun frozen residual GRU on this stem. `results_stem_next.md`.
- **causal ms-feature EWMA AdaLN/FiLM** on stem_ms96 (cond = EWMA μ/σ of 24-d ms inputs at τ∈{1s,4s}@4Hz → zero-init γ/β after DilatedResidual LNs + post-enc FiLM; SWA init; EMA; NO subject/dataset/trial ID; distinct from MixStyle and channel-XY/graph-adj): LODO mean H 0.8690 Acc 82.17 (AVGC 0.8087/75.03, UHD 0.8566); AASD 0.6237/51.09 (hp30 51.02); Carta 0.5391/47.81. ≈SWA; no open target moved. Do not rerun ms-feat-stat AdaLN/FiLM on this stem. `multi_agent_opt/results_speech_adaln.md`.
- **causal MoE gate fuse↔linear_hp30** (IRM↔SWA w=0.5 vs 4-seed hp30; g from |fuse|/EWMA/disagree; selected by mean inner Acc/H; logistic fit lost): selector → **const_fuse** ≡ fuse LODO mean 0.8725/82.58 (AVGC 0.8100/75.14); AASD 0.6155/50.77; Carta 0.5349/47.29. Mixing runners (ewma_abs_fuse) tiny LODO lift, AASD Acc still ~51. Distinct from fixed-w blend+hp30. Do not rerun this fuse↔hp30 causal gate + inner Acc/H selector. `multi_agent_opt/results_moe_fuse.md`.
- **MoE Acc@lower-p_stay proxy** (`moe-pstay-proxy`; same fuse↔hp30 gate grid; select by mean inner Acc then H at p_stay∈{0.95,0.99,0.994} scale 0.8; final score @0.994/0.8; never AASD/Carta labels): Acc@0.95 / Acc@0.99 / Acc@0.994 **all** → **const_fuse**. LODO 0.8725/82.58; AASD 0.6155/50.77 vs linear_hp30 53.12; Carta 0.5349/47.29. Lower-p_stay proxy does not promote hp30-favoring gates. Do not rerun Acc@{0.95,0.99,0.994} reselect on this grid. `multi_agent_opt/results_moe_pstay_proxy.md`.
- **SAM** (Sharpness-Aware Minimization) fine-tune of stem_ms96 from SWA (ascent ρ·g/‖g‖ then AdamW descent; smoke ρ∈{0.02,0.05}→ρ=0.02; abs pos + dilated stem unchanged; EMA eval; distinct from SWA/Polyak): LODO mean H 0.8721 Acc 82.64 (AVGC 0.8088/75.16, UHD 0.8538); AASD 0.6197/51.03; Carta 0.5449/48.40. ≈IRM↔SWA fuse / SWA; AASD Acc flat below linear_hp30; no open target met. Do not rerun SAM / sharpness-aware fine-tune on this stem. `multi_agent_opt/results_sam.md`.
- **PCGrad** multi-dataset gradient surgery on stem_ms96 from SWA (per-env BCE grads; project when `g_i·g_j<0`; mean projected step; conflict_frac≈0.45–0.51; abs pos + dilated stem unchanged; EMA; **not** IRMv1 λ-penalty): LODO mean H 0.8684 Acc 82.38 (AVGC 0.8015/75.05 worse than IRM 0.8095 / fuse 0.8100; UHD 0.8531); AASD 0.6204/51.05; Carta 0.5425/48.00. AASD Acc flat below linear_hp30; no open target met. Do not rerun PCGrad / multi-env gradient projection on this stem. `multi_agent_opt/results_pcgrad.md`.
- **EEGNet DWS 10 s** (eegnet-seed Path A): Lawhern temporal→depthwise spatial→separable, dense (B,W,d); swapΔ≈0.014; chance-probe loss~0.693/50. Do not rerun EEGNet / MI electrode-collapse fronts.
- **Filterbank-only C=232 → vlaai_ws** (eegnet-seed): bands 1–3/3–5/5–7/7–9 Hz stacked as channels; chance (BCE≈0.693). Distinct from residual FIR below. Do not rerun C=232-only stack.
- **In-net FIR filterbank residual → vlaai_ws + SPK** (eegnet-seed Path B `fb_vlaai`): frozen causal FIR banks + zero-init mix on 58-ch, SPK→`bb.bb`; eaP_n5_t60 sw2; swapΔ=0.0111. LODO mean H 0.7639 Acc 70.05. AASD 0.6372/52.84 (hp30 52.98) vs linear_hp30 53.12. Carta hp30 Acc 60.52. No AASD Acc>60. Do not rerun this FIR residual. `multi_agent_opt/results_eegnet.md`.
- **causal WaveNet** 10 s direct-classifier backbone (gated dilated residual stacks, dilations 1..64, tanh⊗σ + residual/skip, dense (B,W,d); eaP_n5_t60, sw2, Fuglsang; NOT EEG-TCNet ELU-TCN / ATCNet / Conformer). SwapΔ=0.0033 (≈zero-EEG); abandoned: AVGC fold mean loss 0.6929 after 50 steps (~chance). No LODO/external scores. Do not rerun this WaveNet gated stack on this front-end. `multi_agent_opt/results_wavenet.md`.
- **Squeeze-and-Excitation residual** on vlaai_ws 10 s (GAP_t → FC→ReLU→FC zero-init; h'=h·(1+gate) after blocks; eaP_n5_t60, sw2, SPK w10_b3 init; NOT channel-mask FiLM / graph-adj; swapΔ=0.0073; early loss@50=0.6714): LODO mean H 0.7693 Acc 70.37; AASD 0.6363/52.59 (hp30 52.89); Carta 0.6402/59.38 (hp30 60.67). ≈chanmask/graph; below published vlaai 0.7846/72.14 and linear_hp30 AASD Acc. Do not rerun SE / channel-gate residual on this front-end. `multi_agent_opt/results_se_vlaai.md`.
- **Mamba DirectNet** 10 s classifier body (`mamba_ssm` pre-norm Mamba+FFN ×4 → dense (B,W,d); eaP_n5_t60, sw2, Fuglsang; NOT decoder-only SSMamba / WaveNet / ATCNet / Conformer). SwapΔ=0.0149; abandoned: AVGC fold mean loss 0.6926 after 50 steps (~chance). No LODO/external. Do not rerun Mamba/S4 as this DirectNet front-end. `multi_agent_opt/results_s4.md`.
- **label-free CORAL/affine on frozen 4-seed logits** (`ext-coral-logits`; unlabeled AASD/Carta moments; CPU; NOT train-time gate_robust CORAL, NOT entropy TTA, NOT Platt): primary `coral4_hp30_AASD` AASD 0.6414/53.11 (Δ−0.01 vs linear_hp30 53.12); train-only whiten 53.16; Carta ~60.5 flat. LODO mean ~0.78/68 at fixed 0.994/0.8. Do not rerun global CORAL/mean-std affine of 4-seed raw/hp30 onto external unlabeled moments. `multi_agent_opt/results_ext_coral.md`.
- **causal-online EWMA moment match on linear_hp30** (`online-moment-hp30`; per-subject/per-trial past-only μ/σ affine-to-train or whiten; τ∈{5,15,30,60}s + scale-only; NO labels/entropy-min; CPU 4-seed; NOT global CORAL, NOT TTA): full Acc/H grid → select **unadapted linear_hp30**. AASD **0.6422/53.12** (Δ0.00). All online variants regress LODO (best adapted `affine_tau30_scale` 0.7678/67.03; external illegal best +0.02 Acc). Prior narrower pick `trial_scale_t60` was 53.04/LODO 0.7809 — still flat. Do not rerun EWMA moment match/whiten on 1-D hp30 at this HMM. `multi_agent_opt/results_online_moment.md`.
- **causal-online EWMA moment match on IRM↔SWA fuse w=0.5** (`online-moment-fuse`; per-trial/per-subject past-only μ/σ affine; train-calibrated or whiten; τ=60 s; NO labels/entropy-min; CPU on MoE fuse cache; NOT hp30 stream): selected `trial_scale_t60` AASD 0.6183/50.97 (Δ+0.03 vs fuse 50.94; Δ−2.15 vs linear_hp30 53.12); LODO mean 0.8769/83.01 (AVGC 0.8168/75.70 tiny lift). Match/whiten hurt LODO. Do not rerun EWMA moment match/whiten on 1-D fuse w=0.5 at this HMM. `multi_agent_opt/results_online_moment_fuse.md`.
- **domain emission-LL gate fuse↔linear_hp30** (`domain-emis-ll`; train-fit Gaussian emissions per stream; causal frame LL under 2-state ±μ or IID; soft EWMA/raw sigmoid or hard cum/EWMA of LL_fuse−LL_hp30; select on LODO inner Acc then H; CPU on MoE fuse caches; **not** stream-stat MoE `|fuse|`/disagree): selector → **const_fuse** ≡ fuse LODO mean 0.8725/82.58; AASD 0.6155/50.77; Carta 0.5349/47.29. Soft LL mix ≤88.21 inner Acc; hard LL → hp30 (~68). Do not rerun this framewise emission-LL fuse↔hp30 gate family. `multi_agent_opt/results_domain_emis.md`.
- **whole-trial energy/MMD/Mahalanobis stream assignment** (`trial-mmd-stream`; per-trial mom4/mom_q summaries → train-only mean/cov ball per stream; mahal/euclid/energy/RBF-MMD; hard or soft gate; LODO Acc/H select; CPU on MoE fuse cache + hp30; **not** framewise domain-emis-LL, **not** MoE `|fuse|`): selector → **const_fuse** ≡ fuse LODO mean 0.8725/82.58; AASD 0.6155/50.77 vs linear_hp30 53.12; Carta 0.5349/47.29; frac→hp30=0. Best non-const soft|mom4|mmd|T1 inner 88.23. Do not rerun this whole-trial summary→ball fuse↔hp30 family. `multi_agent_opt/results_trial_mmd.md`.
- **mixing-only MoE Acc@0.95** (`moe-mix-forced`; same fuse↔hp30 gate grid **minus const_***; select Acc then H @0.95; final @0.994/0.8): Acc@0.95/0.99/0.994 → **`ewma_abs_fuse|t0.5|s4`**. LODO mean **0.8775 / 83.01** (AVGC 0.8170/75.49); AASD **0.6204 / 50.86** (Δ+0.09 vs fuse/`const_fuse` 50.77; Δ−2.26 vs linear_hp30 53.12); Carta 0.5461/48.60; mean g≈0.95/0.91. Tiny LODO lift, flat external Acc. Do not rerun mixing-only Acc@{0.95,0.99,0.994} reselect on this fuse↔hp30 grid. `multi_agent_opt/results_moe_mix_forced.md`.
- **cross-stream causal scale fuse→train-hp30 σ** (`cross-stream-scale`; past-only EWMA τ=60 fuse scale/affine toward unlabeled train **hp30** μ/σ; also 2-D align both streams to hp30 moments then soft mix g∈{0.5,0.75}; LODO Acc/H select; **not** same-stream online-moment-fuse, **not** MoE/emis-LL): selected `fuse_xσ_t60` AASD 0.6147/50.29 (Δ−0.65 vs fuse 50.94; Δ−2.83 vs linear_hp30 53.12); LODO mean 0.8139/82.13 (AVGC 0.7703/73.40 regress vs fuse/omf). Soft-mix LODO Acc≤77.79 (illegal external Acc peek 52.16). Do not rerun fuse→hp30σ affine or fixed-g 2-D align soft mix at this HMM. `multi_agent_opt/results_cross_stream_scale.md`.
- **confidence-gated fuse→hp30σ residual** (`conf-gate-xstream`; past-only EWMA `|fuse|` τ=4 s gate; apply fuse_xσ / soft residual toward fuse_xσ only when |fuse| low; thr∈{0.5,1,1.5} α∈{0.5,1}; LODO Acc/H select; **not** always-on cross-stream, **not** MoE raw fuse↔hp30 mix): selected `soft_resid_a0.5_t0.5_s4` AASD 0.6146/50.72 (Δ−0.22 vs fuse 50.94; Δ−2.40 vs linear_hp30 53.12; Δ+0.43 vs always-on xss); LODO mean 0.8713/82.56 (AVGC 0.8102/74.95 below omf 0.8168); adapted frac ≈7% ext. Recovers H vs always-on; Acc flat. Do not rerun EWMA-|fuse|-gated fuse→hp30σ/resid at this HMM. `multi_agent_opt/results_conf_gate_xstream.md`.
- **seed-disagreement soft/hard gate on 4-seed evidence** (`seed-disagree-gate`; pairwise `mean|s_i−s_j|` ±EWMA τ∈{0,1,4,8}s; soft sigmoid shrink or hard hold/zero; thr=train p50/p75; apply to raw mean / linear_hp30 / median; LODO Acc/H select; **not** fuse↔hp30): selected **hard_hold_raw_mean_t8_p75** LODO 0.7999/75.02; AASD **0.6198/51.12** (Δ−2.00 vs linear_hp30 53.12); Carta 0.5896/54.63; soft hp30≈identity. Do not rerun this pairwise-seed gate grid. `multi_agent_opt/results_seed_disagree.md`.
- **label-free CORAL/affine on 24-d stem ms features** (`ms-feat-coral`; train→unlabeled AASD moments; CORAL s2t/t2s, train-whiten, diag affine; frozen SWA + fuse identity; LODO Acc/H select; **not** 4-seed logit CORAL, **not** 1-D online-moment): select→**identity** SWA LODO 0.8700/82.61 (AVGC 0.8015/74.91); AASD 0.6202/51.08 (Δ−2.04 vs linear_hp30 53.12); Carta 0.5507/48.68; fuse id AASD 50.94. Adapted maps tank LODO (coral_AASD 0.49/72; affine 0.28/51; best adapted t2s 0.86/81). Do not rerun global 24-d ms CORAL/diag-affine/whiten through frozen stem_ms96 SWA. `multi_agent_opt/results_ms_feat_coral.md`.
- **multi-τ causal one-pole HP on 4-seed logits** (`multi-tau-hp`; τ∈{10,15,20,25,35,40,45,50} mean/`med_hp`/2–4 `bank_mean`/soft 0.5·raw+0.5·hp; LODO Acc/H select @0.994/0.8; CPU; **not** τ=30 alone, **not** hpfrac 60/120/240, **not** fuse↔hp30): selected **`soft_hp50`** LODO mean **0.8051/72.04**; AASD **0.6407/52.94** (Δ−0.18 vs linear_hp30 53.12); Carta **0.6537/60.04**. Illegal best AASD Acc `hp10` 53.38. No open target. Do not rerun this multi-τ HP grid on 4-seed logits. `multi_agent_opt/results_multi_tau_hp.md`.
- **causal seed-reliability weighted mean** (`seed-reliability-weights`; past-only inv-var EWMA τ∈{4,8,15,30} / EWMA-agreement to median-others β∈{0.5,1,2,4} / softmax(−EWMA|s_i−mean|/T); raw and/or per-seed hp30 then mean; **always emit**; LODO Acc/H select @0.994/0.8; CPU; **not** seed-disagree hard-hold, **not** multi-τ bank, **not** fuse↔hp30): selected **`invvar_raw_t4`** LODO mean **0.7914/72.92** (AVGC 0.7127/68.90); AASD **0.6412/52.82** (Δ−0.30 vs linear_hp30 53.12); Carta **0.6451/59.11**. Illegal best AASD Acc `softneg_hp30_t4_T0p25` 53.14. No open target. Do not rerun this invvar/agree/softneg 4-seed weight grid. `multi_agent_opt/results_seed_reliability.md`.
- **Transformer DirectNet** (`tf-directnet`; `decoders.TFDirect` causal patch+MHSA+FFN; eaP_n5_t60 sw2 10s; random init; swapΔ=0.0255): abandoned — AVGC chance-probe mean_loss **0.6958** ≥0.685 @50 steps. No LODO/external. Do not rerun this TFDirect stack. `multi_agent_opt/results_tf_directnet.md`.
- **partial CORAL on raw-4 inside 24-d ms** (`ms-partial-coral`; CORAL/t2s/whiten/diag-affine/scale on dims **0:4 only**; dims 4:24 identity; frozen SWA; LODO Acc/H @0.994/0.8; **not** global 24-d ms-feat-coral): select→**coral_t2s_AASD** LODO 0.8688/82.85 (AVGC 0.7935/75.34); AASD **0.6218/51.05** (Δ−2.07 vs linear_hp30 53.12); Carta 0.5494/48.74; identity AASD 51.08. Safer than global CORAL on LODO; flat external Acc. Do not rerun raw-4-only CORAL/affine through frozen stem SWA. `multi_agent_opt/results_ms_partial_coral.md`.
- **causal past-only rank→Gaussianize + hp30** (`rank-gauss-hp30`; per-seed empirical CDF→N(0,1) full past or W∈{15,30,60,120}s; then mean / RG→hp30 / mean→hp30 / hp30→RG; LODO Acc/H @0.994/0.8; CPU; **not** multi-τ / seed-reliability / seed-disagree / online-moment / CORAL / fuse↔hp30): select→**equal_mean** LODO **0.7932/72.68** (AVGC 0.7173/68.51); AASD **0.6417/52.91** (Δ−0.21 vs linear_hp30 53.12); Carta 0.6454/59.28. All RG variants lose LODO Acc (best RG `rg_w120_mean` 68.57). Illegal AASD Acc peek +0.02. Do not rerun this rank-Gauss×hp30 order/window grid. `multi_agent_opt/results_rank_gauss.md`.
- **causal seed power-mean / trim / softabs-|logit|** (`seed-power-trim`; sign-aware |x|^p mean p∈{0.5,1,1.5,2,3}; drop farthest-from-median or lowest EWMA(|logit|) τ∈{4,8}; softmax(|s|) static or EWMA τ×T∈{0.25,0.5,1,2}; ± post 1-D hp30; LODO Acc/H @0.994/0.8; CPU; **not** rank-gauss / seed-reliability / multi-τ / CORAL / fuse↔hp30): select→**`pmean_p0p5`** LODO **0.7932/72.86** (AVGC 0.7226/68.77); AASD **0.6383/52.69** (Δ−0.43 vs linear_hp30 53.12); Carta **0.6441/59.17**. Illegal best AASD Acc `pmean_p2p0_hp30`/`pmean_p3p0_hp30` 53.23 (+0.11). No open target. Do not rerun this power/trim/softabs grid. `multi_agent_opt/results_seed_power_trim.md`.

- **RoPE** on stem_ms96 attention Q/K (replace absolute `pos`; custom pre-norm encoder; dilated stem + ms/w96 kept; SWA init; EMA; NOT ALiBi, NOT v3/v4 local rel-bias 6.25s): LODO mean H 0.8711 Acc 82.68 (AVGC 0.8124/75.66, UHD 0.8541); AASD 0.6216/51.04 (hp30 51.04); Carta 0.5392/48.06 (hp30 49.66). ≈SWA/fuse; AASD Acc flat below linear_hp30 53.12; no open target moved. Do not rerun RoPE / abs-pos→RoPE swap on this stem. `multi_agent_opt/results_stem_rope.md`.
- **stem denser synth-switch** fine-tune from SWA (`stem-synth-switch`; P_SWITCH=0.7 P_SWITCH2=0.7 MIN_GAP=8; SW_GAIN=5.0 post-switch BCE; abs pos + ms/w96 kept; EMA; **not** classifier technique-synth-switch / mix820): LODO mean H 0.8772 Acc 82.78 (AVGC 0.8118/74.91, UHD 0.8468 lost 0.85); AASD 0.6174/50.98 (Δ−2.14 vs linear_hp30 53.12); Carta 0.5516/49.50. Tiny LODO lift vs SWA; AASD Acc flat. Do not rerun denser in-window synth-switch + stronger post-switch gain on this stem. `multi_agent_opt/results_stem_synth_switch.md`.
- **IRM↔RoPE logit blend** (`irm-rope-fuse`; frozen held-*/all5; w∈{0,0.25,0.5,0.75,1}; primary w=0.5; HMM 0.994/0.8; also blend+hp30; never AASD/Carta for w): LODO mean H **0.8724** Acc **82.62** (AVGC 0.8139/75.38 tiny vs RoPE; UHD 0.8566 kept ≥0.85); AASD 0.6222/51.10 (hp30 ~51.0) vs linear_hp30 53.12 (Δ−2.02); Carta 0.5422/48.09. ≈IRM↔SWA fuse; AASD Acc flat. Do not rerun IRM/RoPE weight blends or hp30-on-blend of these two. `multi_agent_opt/results_irm_rope_fuse.md`.
- **SW2 fuse ↔ linear_hp30 soft/hard blend** (`sw2-hp30-blend`; published 4 dec+4 clf10 sw2 via `fuse_general` a=0.5 ↔ 4-seed τ=30 HP; DEC_OFFSET align; w∈{0,0.25,0.5,0.75,1} + |sw2|/disagree gates; LODO Acc then H @0.994/0.8; **not** IRM↔SWA↔hp30): select→**`soft_w=1`≡pure SW2**. LODO mean **0.7914/72.09**. AASD **0.6439/53.46** (Δ**+0.34** vs linear_hp30 53.12; H>0.63) but Acc≪60. Carta **0.6470/59.60**. Mixing toward hp30 loses LODO Acc. Do not re-grid SW2↔hp30 soft/hard under this selector. `multi_agent_opt/results_sw2_hp30_blend.md`, `sw2_hp30_blend.py`.
- **IRM↔stem-synth-switch logit blend** (`irm-synth-fuse`; frozen held-*/all5; w∈{0,0.25,0.5,0.75,1}; primary w=0.5; global_w by mean inner H; HMM 0.994/0.8; also blend+hp30; never AASD/Carta for w): LODO mean H **0.8790** Acc **83.11** (AVGC 0.8192/75.52, UHD **0.8614** recovered ≥0.85); AASD 0.6187/50.92 (Δ−2.20 vs linear_hp30 53.12); Carta 0.5497/49.09. Keeps/beats synth LODO tip; AASD Acc flat. Do not rerun IRM/synth weight blends or hp30-on-blend of these two. `multi_agent_opt/results_irm_synth_fuse.md`.
- **IRM↔stem-stochdepth logit blend** (`irm-stoch-fuse`; frozen held-*/all5; w∈{0,0.25,0.5,0.75,1}; primary w=0.5; global_w by mean inner H; HMM 0.994/0.8; also blend+hp30; never AASD/Carta for w; **≠** IRM↔synth/SWA/RoPE): LODO mean H **0.8722** Acc **82.61** (AVGC **0.8112/75.38** — below synth-fuse 0.8192/75.52; UHD 0.8560); AASD 0.6206/**51.05** (Δ−2.07 vs linear_hp30 53.12; Δ−2.48 vs SW2-swmix 53.53); Carta 0.5437/48.17. ≈SWA-family fuse; AVGC did not move. Do not rerun IRM/stochdepth weight blends. `multi_agent_opt/results_irm_stoch_fuse.md`, `stem_ms96_irm_stoch_fuse.py`.
- **SW2 as 5th raw logit seed** (`sw2-5th-seed`; stack published fuse_general a=0.5 SW2 with 4 clf seeds @DEC_OFFSET; mean5 / hp30_mean5 / mean_hp5 / soft_mean4↔SW2 w∈{0..1}; LODO Acc then H @0.994/0.8; **not** soft SW2↔already-formed linear_hp30): select→**`soft_mean4_sw2_w=0.5`**. LODO mean **0.7969/72.82**. AASD **0.6402/53.06** (Δ−0.06 vs linear_hp30 53.12; Δ−0.40 vs SW2-alone 53.46). Carta **0.6444/59.47**. mean5 LODO 0.7962/72.78 AASD 52.81; hp30_mean5 AASD 52.95 Carta 60.56 (illegal peek). No Acc>60. Do not rerun SW2-as-5th-seed equal-mean/HP/soft-mean4 grids. `multi_agent_opt/results_sw2_5th_seed.md`, `sw2_5th_seed.py`.
- **prep-diverse eaP+lap+lp7 logit bank** (`prep-diverse-seeds`; 4 eaP seeds + lap_eaP s0 + lp7_eaP s0; equal-mean N∈{4,5lap,5lp7,6} / hp30-on-mean / soft-w eaP4-mean|hp30 ↔ lap|lp7|both; LODO Acc then H @0.994/0.8; CPU; reuse folds+ext pickles; **not** SW2-as-5th / SW2↔hp30): select→**`soft_mean4_lp7_w=0.25`**. LODO mean **0.7958/72.91**. AASD **0.6366/52.51** (Δ−0.61 vs linear_hp30 53.12; Δ−0.95 vs SW2-alone 53.46). Carta **0.6445/59.11**. Ref hp30_mean4 AASD 52.90 Carta 60.46. No Acc>60. Do not re-grid prep-diverse mean/hp30/soft-w on this eaP+lap+lp7 bank. `multi_agent_opt/results_prep_diverse_seeds.md`, `prep_diverse_seeds.py`.

## newseed-hp30 (2026-10-29) — DO NOT REPEAT identical s4 bank expand without new frame recipe

Train one extra published-recipe vlaai_ws 10s classifier seed (s4, eaP_n5_t60, sw2, SPK w10_b3 init), then 5-seed linear_mean / linear_hp30 vs 4-seed AASD Acc 53.12. External emis: CACHE (repro 53.12) + s4 LODO col for 5-seed.

- AASD linear_hp30 Acc **53.27** (Δ **+0.15** vs 53.12) — flat; Carta hp30 60.65
- AASD linear_mean Acc 52.83 (Δ −0.08 vs 4-seed mean 52.91)
- LODO5 mean H/Acc 0.7943/72.68; hp30 0.7792/68.16; AVGC not chance
- Code: `newseed_hp30.py`, `experiments/run_newseed_hp30.sh`; folds `direct_clf__eaP_n5_t60_w10b3_ptSPK_s4_aug-sw2`; evidence `ext_clf_*_w10b3sw2P_evidence_s4.pkl`

- **xcorr-matched** 10 s DirectNet (`decoders.XCorrMatched`; multi-lag EEG↔env xcorr → CNN; eaP_n5_t60 sw2 light init). swapΔ=0.10467077791690826; loss 0.6906 still ~chance after 50 steps. `results_xcorr_matched.md`. Do not rerun identical xcorr bank. `multi_agent_opt/results_xcorr_matched.md`.
- **stem SW2-as-ms-channel** (`stem-sw2-ms`; concat published SW2 fuse a=0.5 + causal hp30(SW2) onto 24-d ms → **26-d**; DEC_OFFSET=3 align; expand SWA dw/proj/skip with **zero-init** new cols; EMA 0.995; ~1000 steps; **not** SW2↔hp30 blend / SW2-as-5th / stem fuse): LODO mean **0.8686/82.53** (AVGC 0.8054/75.29, UHD 0.8517); AASD **0.6197/50.87** (Δ−2.25 vs linear_hp30 53.12; Δ−2.59 vs SW2 53.46); Carta **0.5433/47.93**. Does not import SW2 external Acc into stem. Do not rerun SW2-channel concat / weight-expand grids on this stem. `multi_agent_opt/results_stem_sw2_ms.md`, `stem_ms96_sw2_ms.py`.

- **SW2 fuse ↔ technique-synth-switch (swmix) soft/hard blend** (`sw2-swmix-blend`; published fuse_general a=0.5 SW2 ↔ `direct_clf__…_aug-sw2_swmix` deltas; DEC_OFFSET=3; w∈{0..1} + causal |stream| gates; LODO Acc then H @0.994/0.8; CPU; **not** SW2↔hp30 / SW2-as-5th / prep-diverse / stem-sw2-ms): select→**`hard_abs_swmix|t1|s1`** (use swmix when |swmix|≥1 else SW2). LODO mean **0.7916/72.19**. AASD **0.6424/53.53** (Δ**+0.41** vs linear_hp30 53.12; Δ**+0.07** vs SW2 53.46) but Acc≪60. Carta **0.6460/59.71**. Soft-w Acc-monotone toward SW2; Acc-stream×Acc-stream stays mid-50s. Do not re-grid SW2↔swmix soft/hard under this selector. `multi_agent_opt/results_sw2_swmix_blend.md`, `sw2_swmix_blend.py`.
- **stem DropPath / stochastic depth** (`stem-stochdepth`; DropPath on abs-pos stem_ms96 transformer FFN+attn residuals from SWA; linear depth schedule; EMA 0.995; smoke p∈{0.05,0.1,0.2}→**0.1**; ~1000 steps; **≠** v4 `rel_ms` stochdepth / ALiBi / RoPE / denser synth-switch): LODO mean **0.8717/82.69** (AVGC **0.8149/75.38**, UHD 0.8528). AASD **0.6203/51.07** (Δ−2.05 vs linear_hp30 53.12; Δ−2.46 vs SW2-swmix 53.53); Carta **0.5413/48.19**. AVGC tip vs SWA, still short of 0.85/78 and IRM↔synth fuse 0.8192/75.52. Do not rerun DropPath/stochdepth grids on abs-pos stem_ms96 SWA. `multi_agent_opt/results_stem_stochdepth.md`, `stem_ms96_stochdepth.py`.
- **trf-lagbank-clf Path A** (`trf_lagbank_clf.py`; frozen train-only multivariate ridge TRF EEG→env, lags 0..400 ms @32 Hz; lag-bank + sum reconstruction scores as HMM emissions; no deep net; ≠ decoder CCA/FBCCA/WF package benchmark, ≠ xcorr CNN): AVGC probe acc1=**0.5418** snr=0.101 swapΔ=**0.194** (PASS). LODO mean **0.6622/60.37** (AVGC 0.6532/60.81); hp30 0.5469/53.94. AASD **0.6219/50.91** (Δ−2.21 vs linear_hp30 53.12; Δ−2.62 vs SW2 53.46); Carta **0.5984/54.58**. Above chance, far below stem/SW2 Acc. Early past-EEG TRF and CCA-spatial lag-bank stayed ~chance (acc1≈0.51/0.50). Do not rerun identical TRF lag-bank Path A / past-EEG / CCA-spatial lag-bank alone. `multi_agent_opt/results_trf_lagbank.md`, `results/trf_lagbank__eaP_n5_t60.json`.
- **TRF Path A ↔ SW2/swmix Acc-leader soft/hard blend** (`trf-sw2-fuse`; frozen `trf_lagbank__eaP_n5_t60_trf` ↔ Acc leader `hard_abs_swmix|t1` from SW2↔swmix; `t_end` align; w∈{0..1} + causal |leader|/|trf| gates; LODO Acc then H @0.994/0.8; CPU; **not** TRF Path A alone / SW2↔swmix re-grid / SW2↔hp30 / stem-sw2-ms): select→**`abs_leader|t1|s4`**. LODO mean **0.7910/72.62** (Acc tip vs leader 72.19). AASD **0.6369/52.88** (Δ**−0.65** vs leader 53.53; Δ**−0.24** vs hp30 53.12; Δ**−0.58** vs SW2 53.46); Carta **0.6417/59.67**. Acc≪60. Soft-w Acc-monotone toward leader; gating TRF into weak-|leader| frames hurts external Acc. Do not re-grid TRF↔SW2/swmix soft/hard under this selector. Prefer band-limited TRF if Path A Acc is revisited. `multi_agent_opt/results_trf_sw2_fuse.md`, `trf_sw2_fuse.py`.

- **hjorth-vlaai** (causal Hjorth act/mob/comp win 0.5s stacked with eaP → C=232 into vlaai_ws 10s sw2; random init, SPK shapes disallow; ≠ lap/lp7/FIR-fb/chanmask/graph/SE/mrate): swap seed-search PASS Δ=**0.0133** (seed 4). Abandoned: AVGC chance-probe mean loss **0.6932 ≥ 0.685** @50 steps. No LODO/ext. Do not rerun Hjorth-stack → vlaai_ws. `multi_agent_opt/results_hjorth_vlaai.md`, `hjorth_vlaai_lodo.py`.

- **band-trf** (`band_trf_lagbank.py`; causal Butter 1–8 Hz order 4 on EEG+env then Path A ridge TRF lag-bank; ≠ broadband TRF Path A LODO 0.6622/AASD 50.91 / trf-sw2-fuse / lp7_eaP): AVGC probe acc1=**0.5465** swapΔ=0.1924 PASS. LODO mean **0.6655/60.36**. AASD **0.6246/51.02** (Δ-2.51 vs swmix 53.53; Δ-2.10 vs hp30 53.12); Carta **0.5970/54.33**. Do not rerun identical 1–8 Hz Path A. `multi_agent_opt/results_band_trf.md`, `/root/disk_3tb/eeg-thao/results/band_trf_lagbank__eaP_n5_t60_b1_8.json`.

- **stem layer-wise LR** (`layerwise-lr-stem`; AdamW head/skip @ 8e-4, dilated stem+transformer backbone @ **0.1×**; SWA init; abs pos + ms/w96; EMA 0.995; ~1000 steps; **≠** DropPath/stochdepth / IRM↔* grids / ALiBi/RoPE): LODO mean **0.8690/82.45** (AVGC **0.8027/74.95** ≈SWA 0.8015/74.91, below IRM↔synth fuse 0.8192/75.52; UHD 0.8523). AASD **0.6196/51.03** (Δ−2.09 vs linear_hp30 53.12; Δ−2.50 vs SW2-swmix 53.53); Carta **0.5424/48.46**. AVGC did not move toward 0.85/78. Do not rerun backbone×0.1 head LR grids on abs-pos stem_ms96 SWA. `multi_agent_opt/results_layerwise_lr.md`, `stem_ms96_layerwise_lr.py`.

- **dual-band-trf** (`dual_band_trf.py`; causal Butter 1–4+4–8 Hz order 4 Path A ridge TRF banks → train-only LDA/equal/soft-w fuse select LODO Acc then H; ≠ broadband TRF Path A / band-trf 1–8 / trf-sw2-fuse): AVGC equal-fuse acc1=**0.5472** swapΔ=0.1411 PASS. Select **`lda`**. LODO mean **0.6816/61.53**. AASD **0.6252/51.09** (Δ-2.44 vs swmix 53.53; Δ-2.03 vs hp30 53.12; Δ+0.07 vs band-trf 51.02); Carta **0.6021/54.97**. Do not re-grid identical dual-band soft-w/LDA. `multi_agent_opt/results_dual_band_trf.md`, `/root/disk_3tb/eeg-thao/results/dual_band_trf__eaP_n5_t60_b14_48.json`.

- **3-way Acc soft/hard blend SW2×swmix×linear_hp30** (`tri-acc-blend`; simplex w_sw2/w_swmix/w_hp30 on {0,0.25,0.5,0.75,1}+interiors + causal |stream| gates; LODO Acc then H @0.994/0.8; CPU; reuse SW2/swmix/hp30 caches; **≠** pairwise SW2↔swmix / SW2↔hp30 / TRF↔SW2 / SW2-as-5th): select→**`soft_w=(1,0,0)`≡pure SW2**. LODO mean **0.7914/72.09**. AASD **0.6439/53.46** (Δ**−0.07** vs SW2↔swmix leader 53.53; Δ**+0.34** vs hp30 53.12). Carta **0.6470/59.60** (Δ−0.94 vs hp30 60.54). Acc≪60. Mixing toward swmix/hp30 loses mean LODO Acc vs pure SW2. Do not re-grid 3-way SW2↔swmix↔hp30 soft/hard under this selector. `multi_agent_opt/results_tri_acc_blend.md`, `tri_acc_blend.py`.

- **multi-τ causal one-pole HP on SW2 fuse alone** (`sw2-multi-tau`; τ∈{10,15,20,25,30,35,40,45,50}; `hp{τ}` / `soft_hp{τ}`=0.5·raw+0.5·hp / `bank_med` median 2–3 τ; LODO Acc then H @0.994/0.8; CPU; **≠** 4-seed `multi-tau-hp` / Acc-stream blends SW2↔swmix/hp30/tri): select→**`soft_hp50`**. LODO mean **0.7956/71.25**. AASD **0.6447/53.51** (Δ**−0.02** vs leader 53.53; Δ**+0.39** vs hp30 53.12). Carta **0.6498/60.17** (Δ−0.37 vs hp30 60.54). Acc≪60. Illegal best AASD Acc `soft_hp15` 53.67 / `bank_med_10_30` 53.65 not LODO-selected. Same qualitative failure as 4-seed multi-τ on a different stream. Do not re-grid multi-τ HP on SW2 fuse. `multi_agent_opt/results_sw2_multi_tau.md`, `sw2_multi_tau.py`.
- **decoder denser synthetic switches** (`decoder-sw3`; vlaai_ws 5s recon from SPK pretrain; seg_switch=0.75 + multi-switch seg_switch2=0.6 post=2.5; 4 seeds LODO+all5; fuse a=0.5 with published clf-sw2; HMM 0.994/0.8; **≠** SW2↔hp30/swmix soft blends / stem denser-synth / seq-loss A1): decoder-alone LODO mean H ~0.41 (collapsed vs published ~0.76). Fuse LODO **0.7894/71.88**. AASD **0.6381/52.37** (Δ**−1.16** vs Acc leader 53.53); Carta **0.6493/58.67**. Denser switches hurt reconstruction Acc; fuse rides clf. Do not re-grid denser seg_switch on this recipe. `multi_agent_opt/results_decoder_sw3.md`, `decoder_sw3.py`.



- **dss-trf-clf Path A** (`dss_trf_clf.py`; train-only envelope-informed DSS spatial filters → frozen K-dim ridge TRF recon lag-bank; fuse=`lda`; SSD→comp-TRF weaker probe 0.521; ≠ full-channel TRF Path A/band/dual-band, ≠ decoder CCA package): AVGC probe acc1=**0.5385** swapΔ=**0.1860**. LODO mean **0.6891/62.75**. AASD **0.6342/52.09** (Δ-1.44 vs swmix 53.53; Δ-1.03 vs hp30 53.12); Carta **0.6034/55.55**. Do not rerun identical DSS/SSD→comp-TRF Path A. `multi_agent_opt/results_dss_trf.md`, `/root/disk_3tb/eeg-thao/results/dss_trf__eaP_n5_t60.json`.

- **bipolar-vlaai** (bipolar_eaP NN→EA C=58 SPK; ≠ lap/Hjorth/FIR): abandoned: ABORT_CHANCE: mean loss 0.6886 >= 0.685 after 50 steps (near chance). `results_bipolar_vlaai.md`.

- **sph-lap-vlaai** (Perrin CSD sph_lap_eaP C=58 SPK; ≠ NN lap/bipolar): abandoned: ABORT_CHANCE: mean loss 0.7122 >= 0.685 after 50 steps (near chance). `results_sph_lap_vlaai.md`.

- **AVGC-hard curriculum stem** (`avgc-curriculum-stem`; IRM init; frozen IRM↔synth w=0.5 teacher; BCE hardness upsample + staged AVGC batch frac→0.45 after warm-up; EMA 0.995; ≠ avgc-upweight λ): LODO mean **0.8244/79.20** (AVGC **0.7499/72.37** Δ**−0.0693 H / −3.15 Acc** vs IRM↔synth fuse 0.8192/75.52; UHD **0.8015** lost ≥0.85). AASD **0.6194/50.72**; Carta 0.5323/47.23. Hard mining on non-AVGC when held-AVGC destroys transfer. Do not reopen AVGC loss-λ or hard-window/batch-frac curricula on stem_ms96. `multi_agent_opt/results_avgc_curriculum.md`, `stem_ms96_avgc_curriculum.py`.

- **online-moment-sw2** (`online_moment_sw2.py`; past-only causal EWMA μ/σ affine-to-train / scale-only / whiten on Acc tip streams: published SW2 fuse_general a=0.5 **and** Acc leader `hard_abs_swmix|t1`; τ∈{5,15,30,60}+scale@30; LODO Acc then H @0.994/0.8; CPU; reuse sw2_swmix_blend caches; **≠** online-moment-hp30 / Acc-stream soft-hard re-grids): select→**`tip_raw`**. LODO mean **0.7916/72.19**. AASD **0.6424/53.53** (Δ**0.00** vs tip; Δ**+0.41** vs hp30 53.12). Carta **0.6460/59.71** (Δ−0.83 vs hp30 60.54). Acc>60=**NO**. All affine/whiten regress LODO Acc (best adapted tip_affine_t30_scale Acc 71.92). Same flatness as online-moment-hp30. Do not rerun EWMA moment match on frozen SW2 / tip Acc streams. `multi_agent_opt/results_online_moment_sw2.md`.

- **domain emission-LL gate Acc tip↔linear_hp30** (`domain-emis-acc-tip`; `domain_emis_acc_tip.py`; train-fit Gaussian emissions per stream on LODO train folds; causal frame LL under 2-state ±μ or IID; soft EWMA/raw sigmoid or hard cum/EWMA of LL_tip−LL_hp30; select mean LODO Acc then H @0.994/0.8; CPU; tip=`hard_abs_swmix|t1` + DEC_OFFSET-aligned hp30; **≠** domain-emis-ll fuse↔hp30 / Acc soft-w re-grids / online-moment): select→**`const_tip`** ≡ tip LODO **0.7916/72.19**. AASD **0.6424/53.53** (Δtip **0.00**; Δhp **+0.41**). Carta **0.6460/59.71** (Δ−0.83 vs hp30 60.54). Acc>60=**NO**. Soft LL ≤71.15 Acc; hard LL ≤71.21; const_hp30 68.15. Same collapse-to-const as fuse↔hp30 domain-emis. Do not rerun framewise emission-LL tip↔hp30. `multi_agent_opt/results_domain_emis_acc_tip.md`.

- **platt-acc-tip** (`platt_acc_tip.py`; train-fold-only Platt a·x+b / isotonic logit∘iso on Acc tip `hard_abs_swmix|t1` **and** SW2 fuse a=0.5; LODO Acc then H @0.994/0.8; CPU; reuse sw2_swmix_blend caches; **≠** stem `results_calib.md` Platt / online-moment-sw2 / Acc soft-w / domain-emis-acc-tip): select→**`tip_raw`**. LODO mean **0.7916/72.19**. AASD **0.6424/53.53** (Δtip **0.00**; Δhp **+0.41**). Carta **0.6460/59.71**. Acc>60=**NO**. All Platt/iso regress LODO Acc (best calib sw2_isotonic 71.48). Do not rerun train-fold Platt/isotonic on frozen SW2/tip Acc streams. `multi_agent_opt/results_platt_acc_tip.md`.

- **wavelet-fb-vlaai** (causal complex-Morlet real+imag FIR at f0∈{1.5,3,5,7.5} Hz, n_cycles=5, taps=97 → zero-init mix residual → vlaai_ws 10s SPK remap; ≠ Butterworth 1–3/3–5/5–7/7–9 FIR `fb_vlaai`; swapΔ=0.0057; chance 0.6735): LODO mean H 0.7569 Acc 69.08. AASD 0.6405/52.79 (hp30 53.12) vs tip 53.53 / FIR fb 52.84. Carta 58.48. Do not rerun Morlet/CWT residual→vlaai. `multi_agent_opt/results_wavelet_fb_vlaai.md`.

- **tsception-aad** (`tsception_lodo.py`, `decoders.TSception`; Ding multi-scale temporal inception kerns ~0.5/0.25/0.125 s + parallel pointwise S-mix → dense (B,W,d); eaP_n5_t60 sw2 Fuglsang; ≠ ATCNet/Conformer/Shallow/Deep/WaveNet/EEG-TCNet/Mamba/vlaai/FIR/Morlet residual): swapΔ=**0.0318** PASS; abandoned: AVGC chance-probe mean_loss **0.6914 ≥ 0.685** @50 steps. No LODO/ext. Do not rerun this TSception T/S inception DirectNet. `multi_agent_opt/results_tsception.md`.

- **fbcnet-spk** (`fbcnet_spk_lodo.py`, `decoders.FBCNetSPK`; Mane-style frozen FIR 1–3/3–5/5–7/7–9 Hz → per-band Spatial(id)+Temporal(delta) → shared VLAAI soft-mix; SPK remap `bb.*`→`bb.bb.*`; eaP_n5_t60 sw2 Fuglsang; ≠ FIR/Morlet residual→vlaai, ≠ C=232 stack, ≠ TSception/EEGNet/ATCNet/Conformer random-init): swapΔ=**0.0077** PASS; abandoned: AVGC chance-probe mean_loss **0.7034 ≥ 0.685** @50 steps. No LODO/ext. Bandpassed multi-view ≠ SPK raw-EEG training dist (residual FIR starts as identity; this does not). Do not rerun FBCNet parallel-band→shared-vlaai soft-mix. `multi_agent_opt/results_fbcnet_spk.md`.

- **speech-resid-spk** (`speech_resid_spk_lodo.py`, `decoders.SpeechResidVLAAI`; speech-only zero-init FiLM→vlaai_ws 10s SPK remap; eaP_n5_t60 sw2; ≠ FIR/Morlet/FBCNet/speech-adaln/MixStyle; swapΔ=0.0073: LODO 0.7699/70.69; AASD 0.6404/52.87 (Δ-0.66 vs tip); Carta 59.64. `multi_agent_opt/results_speech_resid_spk.md`.)

- **temp-attn-spk** (`temp_attn_spk_lodo.py`, `decoders.TempAttnVLAAI`; zero-init causal temp MHSA residual→vlaai_ws 10s SPK remap; eaP_n5_t60 sw2; ≠ speech-FiLM/SE/graph/Conformer/FIR/Morlet/FBCNet/stem-ALiBi-RoPE; swapΔ=0.0102: LODO 0.7634/70.03; AASD 0.6416/52.92 (Δ-0.61 vs tip); Carta 59.50. `multi_agent_opt/results_temp_attn_spk.md`.)
- **confidence-gated residual on IRM↔synth fuse** (`fuse-conf-resid`; frozen w=0.5; affine/MLP/GRU zero-init resid trained on LODO train only; apply only when past-only EWMA `|fuse|` τ=4 s < thr; select Acc then H @0.994/0.8; **≠** conf-gate-xstream fuse→hp30σ): select→**affine_a1.0_t1.0**. LODO mean **0.8715/83.66** (AVGC **0.8080/76.62** Δ**−0.0112 H / +1.10 Acc** vs tip 0.8192/75.52). AASD **0.6249/51.04** (Δ+0.12 vs fuse 50.92; Δ−2.08 vs hp30 53.12); Carta 0.5495/48.88. Acc↑ trades H↓; miss 0.85/78. Do not rerun this gated residual grid on IRM↔synth fuse. `multi_agent_opt/results_fuse_conf_resid.md`, `fuse_conf_resid.py`.

- **IRM↔synth agree-gate** (`irm-synth-agree-gate`; frozen IRM+synth caches; same-sign→mean; disagree→max-|logit|/IRM/synth/soft_|logit|; LODO Acc then H @0.994/0.8; **≠** fixed-w fuse / MoE fuse↔hp30 / conf-resid): select→**agree_soft_abs**. LODO mean **0.8790/83.08** (AVGC **0.8187/75.57** Δ**−0.0005 H / +0.05 Acc** vs tip 0.8192/75.52). AASD **0.6209/50.96**; Carta 0.5528/49.23. ≈ identity with tip (≈88–93% agree). Miss 0.85/78. Do not re-grid this agree-gate family. `multi_agent_opt/results_irm_synth_agree_gate.md`, `irm_synth_agree_gate.py`.

- **fuse_general a-grid on published SW2** (`fuse-general-a-grid`; `fuse_general_a_grid.py`; a∈{0,0.25,0.5,0.75,1} on 4dec+4clf SW2 streams + optional fixed tip `hard_abs_swmix|t1`; LODO Acc then H @0.994/0.8; CPU; reuse sw2 caches; **≠** soft-w SW2↔swmix/hp30 / tip Platt/domain-emis / IRM↔synth): select→**`a=0.75`**. LODO mean **0.7972/72.80**. AASD **0.6427/53.12** (Δtip **−0.41**; Δhp **0.00**). Carta **0.6467/59.50**. Acc>60=**NO**. a=0.75 lifts LODO vs published a=0.5 but loses tip Acc 53.53. Do not re-grid fuse_general a on this SW2 stack under LODO Acc/H. `multi_agent_opt/results_fuse_general_a_grid.md`.
- **SW2 decoder↔clf disagree-gate** (`dec-clf-disagree-gate`; `dec_clf_disagree_gate.py`; same-sign→fuse_general a∈{0.5,0.75}; disagree→clf|dec|maxabs|soft_|logit|; LODO Acc then H @0.994/0.8; CPU; reuse fuse_general_a_grid/sw2 caches; **≠** fuse-general-a-grid fixed-a / SW2↔swmix soft-w / tip Platt/domain-emis / IRM↔synth agree-gate): select→**`agree_a075_clf`**. LODO mean **0.7961/72.63**. AASD **0.6407/52.93** (Δtip **−0.60**; Δhp **−0.19**). Carta **0.6474/59.37**. Acc>60=**NO**. Sign-agree ≈61–64%; disagree→clf tracks a=0.75 LODO but loses tip Acc. Do not re-grid this dec↔clf disagree-gate family. `multi_agent_opt/results_dec_clf_disagree_gate.md`.

- **aux-env-spk** (`aux_env_spk_lodo.py` AuxEnvDirectNet; SPK vlaai_ws multitask BCE+EEG-trunk env MSE λ=0.1 (smoke [0.1, 0.3, 1.0]); eaP_n5_t60 sw2; ≠ speech-FiLM/temp-MHSA/hardswitch/FIR/Morlet; swapΔ=0.4149: LODO 0.7684/70.21; AASD 0.6399/52.93 (Δ-0.60 vs tip); Carta 59.49. `multi_agent_opt/results_aux_env_spk.md`.)

- **clf denser-switch FT from published SW2** (`clf-hardswitch-ft`; p_switch=0.75/clf_p_switch2=0.6; held-*/all5 init; 4×250; seeds [0, 1, 2, 3]; ≠ decoder-sw3 / swmix mix820 / vlaai identity residuals): LODO 0.7896/71.61 via sw2_fuse_a0.5. AASD 0.6462/53.44 (Δ-0.09 vs tip 53.53). Carta 0.6435/59.52. Acc>60=NO. `/root/disk_3tb/eeg-thao/multi_agent_opt/results_clf_hardswitch_ft.md`. `multi_agent_opt/results_clf_hardswitch_ft.md`, `clf_hardswitch_ft_lodo.py`.

- **subj-adapt-emis-tip** (`subj_adapt_emis_tip.py`; past-only EWMA adaptive HMM emission μ±,σ on Acc tip `hard_abs_swmix|t1` **and** SW2 fuse a=0.5; modes abs_μ / m2_ratio / soft_μ / soft_μ+σ; τ∈{15,30,60}; trial or subject persist; evidence frozen; LODO Acc then H @0.994/0.8; CPU; reuse sw2_swmix_blend caches; **≠** online-moment logit affine, ≠ tip Platt, ≠ domain-emis tip↔hp30, ≠ §C6 bank): select→**`sw2_abs_mu_subj_t30`**. LODO mean **0.7977/73.08** (LODO lift vs tip_raw 72.19). AASD **0.6387/52.90** (Δtip **−0.63**; Δhp **−0.22**). Carta **0.6413/59.27**. Acc>60=**NO**. abs_μ lifts LODO, hurts external Acc; soft_μ illegal AASD peek ≤53.79. Do not rerun adaptive emission μ/σ EWMA on frozen SW2/tip Acc streams. `multi_agent_opt/results_subj_adapt_emis_tip.md`.

- **contrast-aad-spk** (`contrast_aad_spk_lodo.py` ContrastAADDirectNet; SPK vlaai_ws multitask BCE+InfoNCE att/unatt λ=1.0 τ=0.1 (smoke [0.1, 0.3, 1.0]); eaP_n5_t60 sw2; ≠ aux-env MSE/speech-FiLM/temp-MHSA/hardswitch/FIR/Morlet; swapΔ=0.4149: LODO 0.7629/70.16; AASD 0.6388/52.78 (Δ-0.75 vs tip); Carta 59.89. `multi_agent_opt/results_contrast_aad_spk.md`.)

- **focal-stem-avgc** (`stem_ms96_focal_avgc.py`; IRM init; focal BCE γ smoke∈{1,2}×α∈{None,0.25}→**γ=1.0 α=None**; EMA 0.995; 800 steps; ≠ AVGC λ/curriculum/soft-H/fuse-conf-resid/PCGrad/SAM): pure focal LODO **0.8677/82.27** (AVGC **0.8065/75.16**). Fuse w=0.5 focal↔synth LODO **0.8821/83.36** (AVGC **0.8226/75.54** Δtip **+0.0034 H / +0.02 Acc** vs IRM↔synth 0.8192/75.52; UHD 0.8587). AASD **51.05/51.06**. Miss ≥0.85/78. New LODO/AVGC tip via fuse. Do not re-grid this focal γ/α family on IRM stem. `multi_agent_opt/results_focal_stem_avgc.md`.

- **focal-synth-adapt-readout** (`focal_synth_adapt_readout.py`; past-only per-subj/trial causal temp/scale/affine + train Platt/scale on frozen **focal↔synth fuse w=0.5**; LODO Acc then H @0.994/0.8; CPU; reuse fuse_conf_resid synth caches; **≠** fuse-conf-resid gated resid, ≠ irm-synth-agree-gate, ≠ online-moment Acc tip, ≠ tip Acc PP): select→**`platt_then_trial_temp_t30`**. LODO mean **0.8757/84.00** (AVGC **0.8105/76.65** Δtip **−0.0121 H / +1.11 Acc** vs 0.8226/75.54). AASD **50.92**; Carta 49.43. Label-free `trial_temp_t15` mild AVGC H↑ (0.8247) but Acc selector prefers Platt. Miss ≥0.85/78. Do not re-grid this temp/scale/affine/Platt family on focal↔synth fuse. `multi_agent_opt/results_focal_synth_adapt_readout.md`.

- **sw2-feat-gru** (`sw2_feat_gru_lodo.py`; freeze published SW2 vlaai_ws held/all5; NEW causal gru head on penultimate feats; eaP_n5_t60; ≠ hardswitch FT / vlaai residuals / aux-InfoNCE / tip PP / FIR-Morlet-FBCNet): LODO 0.7848/71.55. AASD 0.6343/52.20 (Δtip -1.33). Carta 58.84. Acc>60=NO. `multi_agent_opt/results_sw2_feat_gru.md`.

- **clf longer-tailed switch-duration FT from published SW2** (`clf-switch-dur-ft`; p_switch=0.5/clf_p_switch2=0.5 same rate as sw2; LODO-train lognormal hold prior soft-map τ; held-*/all5 init; 4×250; seeds [0, 1, 2, 3]; ≠ hardswitch denser p_switch / decoder-sw3 / frozen-trunk GRU / tip PP): LODO 0.7887/72.03 via sw2_fuse_a0.5. AASD 0.6437/53.20 (Δ-0.33 vs tip 53.53). Carta 0.6443/59.62. Acc>60=NO. `multi_agent_opt/results_clf_switch_dur_ft.md`, `clf_switch_dur_ft_lodo.py`.

- **tip↔sdur abs-gate** (`tip-sdur-abs-gate`; `tip_sdur_abs_gate.py`; causal `|logit|` hard/soft gate Acc tip `hard_abs_swmix|t1` ↔ clf-switch-dur fuse a=0.5; LODO Acc then H @0.994/0.8; CPU; reuse sw2_swmix_blend + sdur caches; **≠** SW2↔swmix soft-w / tip Platt/domain-emis / fuse_general a-grid): select→**`hard_abs_tip|t1.5|s1`**. LODO mean **0.7933/72.32**. AASD **0.6406/53.06** (Δtip **-0.47**). Carta **0.6466/59.66**. Acc>60=**NO**. `multi_agent_opt/results_tip_sdur_abs_gate.md`.

- **mixup-stem-avgc** (`stem_ms96_mixup_avgc.py`; focal-IRM init; input/label Mixup `lam~Beta(α,α)` smoke α∈{0.2,0.4}→**α=0.4**; EMA 0.995; 800 steps; ≠ MixStyle-vlaai / AVGC λ/curriculum / focal γ re-grid / soft-H / adapt-readout): pure mixup LODO **0.8755/82.11** (AVGC **0.8162/74.70**). Fuse w=0.5 mixup↔synth LODO **0.8815/83.12** (AVGC **0.8213/75.29** Δtip **−0.0013 H / −0.25 Acc** vs 0.8226/75.54; UHD 0.8573). AASD **50.84/50.96**. Miss ≥0.85/78. Do not re-grid Mixup α on tip stem. `multi_agent_opt/results_mixup_stem_avgc.md`.


## CLOSED: Mixup / BCE-family / stem re-tunes from tip inits (2026-10-30)

Do **not** re-grid Mixup α, focal γ, AVGC λ/hard curricula, or other BCE-family stem FT from tip (focal-IRM/SWA/IRM) inits. Tip was focal↔synth fuse w=0.5 (**0.8821/83.36; AVGC 0.8226/75.54**) at Mixup close; later **causal-ds-temp** `held_T_nll_med` moves AVGC to **0.8245/75.64**. See `results_mixup_stem_avgc.md`.

- **multi-stem soft-attn pool** (`multi-stem-attn-pool`; causal EWMA feats of frozen focal+synth+IRM → MLP softmax gates; train BCE on LODO train only; ≠ fixed-w / agree-gate / conf-resid / MoE fuse↔hp30 / Mixup/focal): Select soft_attn. LODO 0.8650/82.28 (AVGC 0.8057/75.10 Δtip -0.0169 H). AASD 0.6196/50.99. `/root/disk_3tb/eeg-thao/multi_agent_opt/results_multi_stem_attn_pool.md`. `multi_agent_opt/results_multi_stem_attn_pool.md`, `multi_stem_attn_pool.py`.
- **csp-mdm-fuse** (`csp_mdm_fuse.py`; joint EEG+env SPD CSP→numpy Riemannian MDM + CSP→env; fuse↔Acc tip planned): abandoned AVGC chance-probe (joint_mdm acc1 0.504 / csp_env 0.501; swapΔ≪0.01). ≠ EEG-only FBCSP/Riemann spatial decode. Do not rerun identical joint-SPD CSP/MDM windows. `multi_agent_opt/results_csp_mdm_fuse.md`.

- **hilbert-coh-fuse** (`hilbert_coh_fuse.py`; 1–8 Hz Hilbert plv EEG↔env Path A + fuse↔Acc tip): abandoned AVGC chance-probe. ≠ TRF/DSS/CSP-MDM. `multi_agent_opt/results_hilbert_coh_fuse.md`.

- **hilbert-coh-fuse** (`hilbert_coh_fuse.py`; 1–8 Hz Hilbert plv EEG↔env Path A + fuse↔Acc tip): abandoned AVGC chance-probe. ≠ TRF/DSS/CSP-MDM. `multi_agent_opt/results_hilbert_coh_fuse.md`.

- **clf shorter-hold switch curriculum FT from published SW2** (`clf-shorthold-ft`; p_switch=0.5/clf_p_switch2=0.5 same rate as sw2; LODO-train short-half hold prior ICDF→short-side τ/gaps; held-*/all5 init; 4×250; seeds [0, 1, 2, 3]; ≠ switch-dur longer holds / hardswitch denser p_switch / tip-sdur / decoder-sw3 / frozen-trunk GRU / tip PP): LODO 0.7898/71.92 via sw2_fuse_a0.5. AASD 0.6466/53.44 (Δ-0.09 vs tip 53.53). Carta 0.6356/59.52. Acc>60=NO. `/root/disk_3tb/eeg-thao/multi_agent_opt/results_clf_shorthold_ft.md`. `multi_agent_opt/results_clf_shorthold_ft.md`, `clf_shorthold_ft_lodo.py`.

## CLOSED: Switch-schedule classifier FT family (2026-10-30)

Do **not** re-grid denser p_switch, longer-hold duration priors, or shorter-hold curricula on published SW2 clf. Acc tip remains SW2↔swmix hard_abs_swmix|t1 **53.53**. See `results_clf_shorthold_ft.md`.

- **dann-grl-stem** (`stem_ms96_dann_grl.py`; GRL + domain CE on pooled stem features; domains=LODO train dataset ids; init focal-IRM; smoke λ_max∈{0.1,0.5,1.0}→**0.1**; EMA; fuse w=0.5↔synth; ≠ IRM/PCGrad/Mixup/focal): pure LODO **0.8620/81.85** (AVGC **0.8042/74.92**). Fuse LODO **0.8770/82.97** (AVGC **0.8199/75.46** Δtip **−0.0027 H / −0.08 Acc** vs 0.8226/75.54; UHD 0.8582). AASD **50.87/50.96**. Miss ≥0.85/78. Tip unchanged. Do not re-grid DANN λ on this stem. `multi_agent_opt/results_dann_grl_stem.md`.

- **vicreg-stem-pretrain** (`stem_ms96_vicreg_pretrain.py`; VICReg label-free aug-invariant pretrain on stem_ms96 ms windows → LODO BCE FT; PRE=400 FT=600; expander 256; sim/std/cov 25/25/1; scratch encoder; EMA; fuse w=0.5↔synth; ≠ DANN/IRM/Mixup/focal/ContrastAAD/CPC): smoke AVGC 0.7525/70.24 OK. Pure LODO **0.8556/82.03** (AVGC **0.7875/73.66**). Fuse LODO **0.8727/83.06** (AVGC **0.8073/74.78** Δtip **−0.0153 H / −0.76 Acc** vs 0.8226/75.54). AASD **51.44/51.19**. Miss ≥0.85/78. Tip unchanged. Do not re-grid VICReg coeffs on this stem. `multi_agent_opt/results_vicreg_stem_pretrain.md`.

- **temporal-cpc-stem** (`stem_ms96_cpc_pretrain.py`; temporal CPC label-free context-GRU→future latent InfoNCE on stem_ms96 ms windows → LODO BCE FT; PRE=400 FT=600; k∈{2,4,8,12,16,24} temp 0.1 stride 4; scratch encoder; EMA; fuse w=0.5↔synth; ≠ VICReg/Barlow/DANN/IRM/Mixup/focal/ContrastAAD): smoke AVGC 0.7477/69.92 OK. Pure LODO **0.8561/81.82** (AVGC **0.7832/73.17**). Fuse LODO **0.8743/82.97** (AVGC **0.8031/74.49** Δtip **−0.0195 H / −1.05 Acc** vs 0.8226/75.54). AASD **51.59/51.19**. Miss ≥0.85/78. Tip unchanged. Do not re-grid CPC offsets/temp on this stem. `multi_agent_opt/results_temporal_cpc_stem.md`.

## CLOSED: stem SSL pretrain family (VICReg / CPC) (2026-10-30)

Do **not** re-grid VICReg coeffs, Barlow twins, or temporal-CPC offsets/temp on `stem_ms96` ms windows. Both SSL→FT runs sit below tip fuse AVGC (VICReg −0.0153; CPC −0.0195). Tip later moved by causal-ds-temp held_T_nll_med to AVGC **0.8245/75.64** (was fuse raw 0.8226/75.54).

- **spectrogram-vit-acc** (`spectrogram_vit_lodo.py` / `decoders.SpecViT`; EEG→causal Mel-STFT n_fft=16 n_mels=12 n_comp=8 → tiny ViT L=2 d=64 → antisymmetric DirectNet 10s eaP sw2; ≠ TFDirect/vlaai/FBCNet/TSception/SSL): swapΔ=0.0752 PASS; AVGC chance-probe mean_loss **0.6962≥0.685**. No LODO/ext. Same chance fate as TFDirect/WaveNet/Mamba. Do not re-grid Mel/STFT SpecViT. `multi_agent_opt/results_spectrogram_vit_acc.md`.

- **subj-cond-calib** (`subj_cond_calib.py`; subject-conditional ridge/Platt on frozen Acc tip hard_abs_swmix|t1 + SW2; train-subject Platt bank + unlabeled logit-moment ridge→(a,b); ≠ global Platt/online-moment/emis-adapt): Select **sw2_ridge_ab** LODO **0.7525/72.58**. AASD **53.01** (Δtip **−0.52**). Carta 59.46. tip_raw still AASD 53.53. Tip unchanged. Do not re-grid subj ridge/Platt on tip streams. `multi_agent_opt/results_subj_cond_calib.md`.

## CLOSED: SpecViT Mel Acc front-end + tip subj-cond calib (2026-10-30)

Do **not** rerun Mel/STFT SpecViT Acc DirectNet or subject-conditional ridge/Platt on Acc tip SW2/`hard_abs_swmix|t1`. Acc tip remains SW2↔swmix hard_abs_swmix|t1 (**53.53**); Carta tip linear_hp30 (**60.54**).

- **trf-pathb-sw2** (`trf_pathb_sw2_lodo.py`, `decoders.TRFPathBVLAAI`; frozen train-only envelope TRF gated (init=0.25) corr-proxy side→SPK vlaai_ws 10s; eaP_n5_t60 sw2; ≠ speech-resid FiLM@0 / Path A TRF / trf-sw2-fuse; swapΔ=0.0079: LODO 0.7717/70.22; AASD 0.6391/52.68 (Δ-0.85 vs tip); Carta 59.49. `multi_agent_opt/results_trf_pathb_sw2.md`.)

## CLOSED: emission-shift Acc PP family (domain / subj-adapt / proto-mem) (2026-10-30)

Do **not** rerun tip↔hp30 emission-LL gates, past-only EWMA adaptive HMM μ/σ, or cross-dataset prototype-memory soft-kNN/Sinkhorn emission blending on frozen SW2/`hard_abs_swmix|t1`. All stay mid-50s AASD Acc; tip Acc remains **53.53**. Conformer Acc front-end already scored — do not reopen.

- **proto-mem-emis** (`proto_mem_emis.py`; LODO-train per-subj/ds/k-means prototypes of unlabeled logit moments + labeled signed (μ,σ); hard-NN / soft-kNN / Sinkhorn match → HMM emission shift; evidence frozen; ≠ domain-emis tip↔hp30 / subj-adapt EWMA): Select **`sw2_hard_subj`**. LODO **0.7984/72.96**. AASD **0.6418/53.17** (Δtip **−0.36**). Carta **59.53**. Prototype diversity AASD n_unique=**71** is_const=**False** (≠ domain-emis collapse-to-const). Acc>60=**NO**. Tip unchanged. `multi_agent_opt/results_proto_mem_emis.md`.

## CLOSED: EEG noise / SpecAugment clf FT (2026-10-30)

- **clf EEG noise/channel-drop/SpecAugment FT from published SW2** (`clf-noise-aug-ft`; noise=0.25/p_chan=0.35/p_cluster=0.5/tilt+time+freq masks; p_switch=0.5; held-*/all5 init; 4×250; seeds [0,1,2,3]; ≠ hardswitch/shorthold/switch-dur / MixStyle / tip PP / emis-shift): Select **sw2_fuse_a0.5**. LODO **0.7900/71.74**. AASD **0.6439/53.20** (Δtip **−0.33**). Carta **59.45**. Acc>60=**NO**. Tip unchanged. Family CLOSED (flat Acc envelope). `multi_agent_opt/results_clf_noise_aug_ft.md`, `clf_noise_aug_ft_lodo.py`.

- **causal-ds-temp** (`causal_ds_temp.py`; held-only / per-dataset causal T on frozen **focal↔synth fuse w=0.5**; NLL T_d mean/med/nested-pick + causal T grid + per-ds emis T; LODO Acc then H @0.994/0.8; CPU; reuse adapt_readout caches; **≠** adapt-readout per-subj/trial EWMA, ≠ multi-τ, ≠ global train_temp HMM-invariant): select→**`held_T_nll_med`**. LODO mean **0.8844/83.38** (AVGC **0.8245/75.64** Δtip **+0.0019 H / +0.10 Acc** vs raw 0.8226/75.54; UHD 0.8630). AASD bookkeeping 51.17. Miss ≥0.85/78. Small tip move. Do not re-grid held-only/per-ds NLL or nested-grid T on this fuse. `multi_agent_opt/results_causal_ds_temp.md`.

## CLOSED: decision-layer causal temperature on tip fuse (2026-10-30)

Do **not** re-grid held-only / per-dataset NLL temperature or nested Acc·H T grids on frozen focal↔synth fuse w=0.5. Global T is HMM-invariant; held-only T gives only tiny AVGC lift (+0.0019). Tip now **held_T_nll_med** on fuse (**0.8844/83.38; AVGC 0.8245/75.64**). Isotonic + stacking also CLOSED — prefer **group-dro-stem-avgc** — not Mixup/SSL/DANN/adapt-readout/multi-τ/isotonic/stack.


- **isotonic-ds-fuse** (`isotonic_ds_fuse.py`; train-only isotonic monotone map tip fuse scores→probs→logit on frozen **focal↔synth fuse w=0.5**; global_iso_all / held_iso_global / held_iso_per_ds_pick / per_ds_iso_emis_gheld; LODO Acc then H @0.994/0.8; CPU; reuse adapt_readout caches; **≠** causal-ds-temp T, ≠ platt-acc-tip Acc streams, ≠ adapt-readout): select→**`held_iso_per_ds_pick`**. LODO mean **0.8761/83.95** (AVGC **0.8123/76.60** Δtip **−0.0122 H / +0.96 Acc** vs tip 0.8245/75.64). All iso variants raise AVGC Acc ~+1 but cut H ~−0.01. Tip unchanged. Do not re-grid train-only isotonic on this fuse. `multi_agent_opt/results_isotonic_ds_fuse.md`.

## CLOSED: isotonic-on-tip-fuse family (2026-10-30)

Do **not** re-grid train-only isotonic (global / held-only / per-ds nested-pick / per-ds-emis) on frozen focal↔synth fuse w=0.5. Acc↑/H↓ tradeoff; tip remains **held_T_nll_med** (**0.8844/83.38; AVGC 0.8245/75.64**). Prefer **group-dro-stem-avgc** — not Mixup/SSL/DANN/T/isotonic/stack.

- **evidential-tip-stack** (`evidential_tip_stack.py`; train-only NNLS/simplex + logistic + Dirichlet-evidential stack on frozen **focal+synth+IRM** tip stem logits; leave-held fit without held labels; LODO **H then Acc** @0.994/0.8; CPU; ≠ multi-stem soft-attn EWMA→MLP gates, ≠ T/iso on tip fuse): select→**`dirichlet_evid`**. LODO mean **0.8700/82.34** (AVGC **0.8084/75.23** Δtip **−0.0161 H / −0.41 Acc** vs tip 0.8245/75.64). nnls 0.8113/75.15; logistic 0.7897/74.75; equal_mean3 0.8164/75.61 — all below tip. Tip unchanged. `multi_agent_opt/results_evidential_tip_stack.md`.

## CLOSED: decision-layer stacking on tip stems (2026-10-30)

Do **not** re-grid train-only NNLS/simplex / logistic / Dirichlet-evidential (or similar linear/evidential) stacks on frozen focal+synth+IRM tip logits. All hurt vs tip held_T_nll_med (**0.8844/83.38; AVGC 0.8245/75.64**). Decision-layer PP on tip fuse saturated: **no T / iso / stack / attn / agree**. Prefer qualitatively new **train-time AVGC** (e.g. Group-DRO / CVaR worst-dataset stem) — not Mixup/SSL/DANN/decision-PP.

- **speech-xattn-dec** (`speech_xattn_dec_lodo.py`, `decoders.SpeechXAttnVLAAI`; speech-env KV × EEG-latent Q MHCA residual→vlaai_ws 10s SPK; eaP_n5_t60 sw2; ≠ temp-attn-spk/speech-resid/aux-env/env_xattn; swapΔ=0.0106: LODO 0.7671/70.21; AASD 0.6423/53.75 (Δ+0.22 vs tip; head=tip_fuse_a0.50); Carta 59.49. Tip moved=YES. `multi_agent_opt/results_speech_xattn_dec.md`.)
- **xattn-sw2-emis-fuse** (2026-10-30): speech-xattn↔hard_abs_swmix a=0.5 under SW2 a=0.5 emis → AASD Acc **54.05** (new Acc tip), Carta **60.60**. STOP further a-grid / emis-choice tip soft-fuse PP on these two streams.

- **V-REx stem** (`vrex_stem_20261106`; environment-risk variance penalty on
  the IRM stem init; λ=1.0, 800 steps): LODO mean **0.8652/82.10**, synth fuse
  **0.8794/83.03**; AASD **0.6196/51.10**; Carta **0.5412/47.99**. LODO is
  **SUSPECT column-0 stem** under GAP-01 and external performance regressed.
  Closed as non-improving. Do not rerun V-REx/environment-risk variance on
  this stem. The original 79-entry forbidden inventory is unchanged; this is
  a post-inventory addendum. `experiments/multi_agent_opt/HANDSHAKE_VREX_DONE.md`.

- **T1 latency diagnostic / LagBank-Sel gate** (`t1_latency_20261106`):
  ±500 ms on frozen 4-seed SW2. AASD peaks at 0 ms; Carta −31.25 ms gain is
  below 2 subject-bootstrap SE. No conversion-offset evidence, so Step B was
  not run. Do not repeat this external-label latency scan.

- **T2 ESR-stem** (`t2_esr_20261106`; exactly odd, position-free,
  random-start resynthesis, 1500 steps, K=8): gates pass (odd error 0,
  position agreement 1.0), but LODO mean **0.7489/71.91**, AASD
  **0.6073/50.68**, Carta **0.5429/50.35**. Honest but non-improving; do not
  rerun this ESR recipe. `results/research/stem_esr.json`.

- **T1 stimulus-latency diagnostic** (`t1_latency_20261106`; frozen 4-seed
  SW2, −16…+16 samples @32 Hz): AASD peak is exactly 0 ms, acc1 0.53977.
  Carta peak −31.25 ms improves acc1 only +0.00532 with subject-bootstrap
  SE 0.00447 (<2 SE). No fixed conversion offset; Step B causal lag-bank was
  correctly gated off. Do not rerun this diagnostic grid.

- **T2 ESR-stem AVGC screen** (`t2_esr_20261106`; exact odd wrapper,
  no position embedding, replicate causal padding, random-start resynthesis,
  K=8, 1500 steps): structural oddness error 0 and position agreement 1.0,
  but held AVGC H **0.7036** / Acc **68.50** (acc1 0.7151). Stopped before
  five-fold/external. Do not rerun this exact ESR recipe; a materially new
  evidence/resynthesis design would be required.

- **T1 fixed stimulus-latency diagnostic** (`t1_latency_20261106`; frozen
  4-seed SW2, AASD 18 subjects/900 trials, δ=−16…+16 samples @32 Hz):
  peak **δ=0 / 0 ms**, subject-mean acc1 **0.53977** (median 0.54103).
  Every nonzero lag is worse and no nonzero peak exceeds 2 bootstrap SE.
  Therefore no conversion offset was found and causal lag-bank Step B was
  not run. Do not repeat this fixed-lag AASD sweep without new alignment
  evidence. `results/research/lagbank_diag.{json,csv}`.

- **T3 CF-Null on honest SW2** (`t3_cfnull_diag_20261107` +
  `t3_cfnull_train_20261107`): the diagnostic gate passed (rho AASD
  **0.9765**, Carta **0.8652**), but lambda 0.1 seed-0 fine-tuning for
  4x250 steps across all five LODO folds plus all5 external gave LODO mean
  **0.7880/71.74**, AASD **0.6367/52.35**, Carta **0.6428/58.57**.
  External Acc regressed and remains far below target, so seeds 1-3 were
  not spent. Do not rerun this CF-Null loss/20 s null construction without
  a materially different front-end hypothesis.

- **weight-soup-hp30 / T7** (`weight_soup_hp30.py`; parameter avg of 4 SW2 fold/all5 weights + causal τ=30 HP; ≠ logit equal-mean / stem SWA): T7 gate FAIL (frame_acc Δ **−22.65** pt vs equal-mean). Soup logits collapsed (std≈0.03 vs seed≈1.1; corr≈−0.09). LODO mean **0.5476/51.69**; AASD **0.6092/49.58**; Carta **0.5635/52.69**. Cross-seed weight averaging is not mode-connected for vlaai_ws. Do not rerun cross-seed SW2 weight soups. Bulky folds/ckpts/evidence deleted; kept `results_weight_soup_hp30.md` + JSON.

- **hp30-center-ft** (`hp30_center_ft_lodo.py`; FT published SW2 so BCE is on causal τ=30 HP of trial-streamed logits; infer with same HP; inherit linear_hp30; ≠ multi-τ/rel_hp/online-moment/soup): SCREEN seed0 held-AVGC **FAIL** — loss 0.48→2.62 over 250 trial-steps; AVGC H **0.6587**/Acc **60.93** (pub≈0.726, Δ−0.067). Training on HP-logits destabilises the published net. Do not rerun this HP-as-train-target FT. Bulky ckpts deleted; `results_hp30_center_ft.md`.

- **b1-hp30-semimarkov** (`b1_hp30_semimarkov.py`; Weibull semi-Markov B1 on linear_hp30 and BMMS3-mean4+hp30; LODO Acc then H select; ≠ SW2-fuse B1 / msm.py): primary **linear_hp30 m40 k0.6** LODO **0.7802/68.43**; AASD **0.6402/52.86**; Carta **0.6551/60.48**. BMMS-hp30 m40 k0.6 LODO 0.7667/67.32; AASD 53.48; Carta **61.95**. LODO Acc stays <70; AASD Acc flat mid-50s. Do not re-grid B1 dwell on these hp30 streams. CPU-only; no bulky artifacts. `results_b1_hp30_semimarkov.md`.

- **t5-distractor-hp30** (`t5_distractor_hp30.py`; in-trial past-envelope distractor ranking on frozen SW2 + causal τ=30 HP; AASD frame-Acc screen n=80 K=2 shift=8s): SCREEN **FAIL** — base 0.5256 → t5_hp30 0.5294 (Δ **+0.38** pt < +1 gate). No full LODO/ext. Do not rerun this distractor K/shift grid on frozen SW2+hp30. No bulky artifacts. `results_t5_distractor_hp30.md`.
- **hier-dwell-hp30 / H1** (`hier_dwell_hp30.py`; dual geo dwell long+short + causal volatility gate on frozen linear_hp30 τ=30; LODO Acc then H; ≠ B1 single prior / MoE fuse↔hp30 / multi-τ): selected **L0.998_S0.9_t5_b5_causal_runmed** LODO **0.7801/68.28**; AASD **0.6412/53.06**; Carta **0.6546/60.58**. LODO Acc stays <70 / H<0.80; Carta Acc barely held. Do not re-grid this volatility-gate dual-dwell family on hp30. CPU-only; no bulky artifacts. `results_hier_dwell_hp30.md`.
- **esr-hp30 / H2 residual** (`esr_hp30.py`; frozen T2 ESR held-/all5 logits + causal τ=30 HP; LODO Acc then H @ p_stay 0.994/0.998; ≠ ESR retrain / exact T2): selected p0.994 LODO **0.8138/74.96**; AASD **0.6141/50.83**; Carta **0.5753/52.79**. LODO clears priority-B train bar but Carta Acc/H collapse vs linear_hp30. Do not rerun ESR+hp30 readout on these ckpts. No new bulky artifacts (reused `models/stem_esr`). `results_esr_hp30.md`.
- **h3-lagbank-skip** (sequence slot H3): T1 LagBank Step A already closed (AASD peak δ=0 ms); Step B causal lag-bank remains gated off without new alignment evidence. No run. Do not reopen without new converter/latency evidence.
- **sr-pool-hp30 / H4 residual** (`sr_pool_hp30.py`; SR-Pool short tw → mean4 → causal τ=30 HP + LODO-fast p_stay grid; ≠ sr_pool_sw2 alone / multi-τ / MoE): selected tp=4.0 p0.994 LODO **0.7390/64.64**; AASD **0.6488/53.67**; Carta **0.6534/59.72**. Below linear_hp30 LODO and Carta Acc; AASD Acc tip-ish but fails priority B. Do not rerun SR-Pool+hp30 on published SW2. No new bulky ckpts (inference-only). `results_sr_pool_hp30.md`.

- **ablate-hp30-a1_bandpass** (`ablate_hp30_oneknob.py`; bp_15_45): LODO **0.7854/73.26**; AASD **0.6017/49.29**; Carta **0.5735/52.12**. Priority-B FAIL. Do not rerun this exact one-knob grid. `results_ablate_hp30_a1_bandpass.md`.

- **ablate-hp30-a2_cmrem** (`ablate_hp30_oneknob.py`; cm_off_hp30): LODO **0.7797/68.15**; AASD **0.6410/52.90**; Carta **0.6560/60.46**. Priority-B FAIL. Do not rerun this exact one-knob grid. `results_ablate_hp30_a2_cmrem.md`.

- **ablate-hp30-a3_mad** (`ablate_hp30_oneknob.py`; mad_t60): LODO **0.7810/68.15**; AASD **0.6387/52.85**; Carta **0.6541/60.42**. Priority-B FAIL. Do not rerun this exact one-knob grid. `results_ablate_hp30_a3_mad.md`.

- **ablate-hp30-a4_b3** (`ablate_hp30_oneknob.py`; b3_e10_g0.5): LODO **0.8048/73.21**; AASD **0.6144/50.21**; Carta **0.6154/56.36**. Priority-B FAIL. Do not rerun this exact one-knob grid. `results_ablate_hp30_a4_b3.md`.

- **b3_multiscale_resid_hp30** (`b3_multiscale_resid_hp30.py`; banke4_e10_e30_g0.5_a1_lam0.01): LODO **0.8069/75.35**; AASD **0.6160/50.96**; Carta **0.5899/54.78**. PASS-gate FAIL (ext H<0.65 and/or LODO bar). Do not rerun this exact multi-τ bank + α-resid grid. `results_b3_multiscale_resid_hp30.md`.

## lead-ml-hp30 addenda (2026-11-08) — linear_hp30 structure push

Root-cause note: `linear_hp30` wins on real-switch **H** because causal τ=30 HP removes slow
signed bias that synthetic-LODO transformers overfit; Acc-fuse raw lifts LODO Acc but alone
puts Carta H just under 0.65. Soft α=0.5 mix keeps the HP basin while clearing LODO bars.

- **dual-hp30-sw2-bmms** (`dual_hp30_sw2_bmms.py`; soft-w SW2-linear_hp30 ↔ BMMS-hp30; ≠ Acc-fuse raw / SW2↔hp30): select w=0.5 LODO **0.7838/68.55**; AASD **0.6413/53.09**; Carta **0.6600/61.37**. LODO Acc<70. Do not re-grid this dual-hp30 soft-w. `results_dual_hp30_sw2_bmms.md`.

- **capped-odd-resid-hp30** (`capped_odd_resid_hp30.py`; hard-capped α∈{0.05,0.10,0.15} odd EMA ridge on linear_hp30; ≠ B3 free-α): select a0.15_e15_g1 LODO **0.7834/68.70**; AASD **0.6396/52.84**; Carta **0.6537/60.38**. Tiny LODO lift. Do not re-grid this capped odd residual. Script deleted after log. `results_capped_odd_resid_hp30.md`.

- **capped-sw2-hp30-blend** (`capped_sw2_hp30_blend.py`; hard-capped α_hp∈{0.1..0.3} Acc-first on SW2/AccFuse↔linear_hp30; ≠ free SW2↔hp30): Acc-first → accfuse α=0.10 LODO **0.7966/72.54**; AASD **0.6408/52.94**; Carta **0.6449/60.12**. Peak LODO H **0.7995**@α0.20. Superseded by H-select. Script deleted. `results_capped_sw2_hp30_blend.md`.

- **accfuse-hp30-hselect** (`accfuse_hp30_hselect.py`; AccFuse↔linear_hp30 / BMMS-hp30 finer α; H-max among Acc>70): tip **accfuse_x_linear_hp30 α=0.50** LODO **0.7998/71.30**; AASD **0.6427/53.08**; Carta **0.6521/60.60**. Fixed-0.994 H short by 0.0002. Kept as helper. `results_accfuse_hp30_hselect.md`.

- **accfuse-hp30-pstay** (`accfuse_hp30_pstay.py`; nested p_stay on α=0.50): nested LODO **0.8041/72.13**; EXT@0.994 AASD **0.6427**; Carta **0.6521**. Demonstrated nested clears LODO. Script deleted after log. `results_accfuse_hp30_pstay.md`.

- **agree-gate-dual-hp30** (`agree_gate_dual_hp30.py`; sign-agree SW2-hp30↔BMMS-hp30; ≠ IRM-synth agree-gate): select disagree_softabs LODO **0.7840/68.44**; AASD **0.6424/53.22**; Carta **0.6617/61.41**. LODO Acc fail; Carta H tip. Script deleted. `results_agree_gate_dual_hp30.md`.

- **accfuse-hp30-final** (`accfuse_hp30_final.py`; AccFuse↔linear_hp30 α=0.50 + global p_stay grid): **KEPT WINNER**. Acc→H select **p=0.998 sc=1.0** LODO **0.8041/72.13**; Carta **0.6523/60.27**; AASD 0.6365/52.34. Milder LODO-pass **p=0.996 sc=0.8** LODO **0.8011/71.43**; Carta **0.6522/60.55**; AASD **0.6421/53.01**. Preset 0.994/0.8 LODO 0.7998/71.30 (H −0.0002). Do not re-grid this exact AccFuse↔linear_hp30 α=0.50 p_stay table. `results_accfuse_hp30_final.md`, `results/research/accfuse_hp30_final.json`.

## lead-ml stricter targets push (2026-11-08) — AccFuse↔hp30 → soft_hp50 near-miss

New bars: LODO **H>0.85 Acc>75** and test **H≥0.65**. Prior tip AccFuse↔linear_hp30 α=0.50 @p=0.996/0.8: LODO **0.8011/71.43**, Carta **0.6522**, AASD **0.6421**. Gap ≈ +0.05 H / +3.6 Acc on LODO.

Root cause unchanged: stem-family streams clear new LODO bars but carry slow bias that kills real-switch H; τ=30 HP strips bias on 4-seed logits. Soft-mixing HP'd stem into AccFuse tip clears LODO but Carta saturates ~0.643 (Pareto wall).

- **stem-hp30-accfuse** (`stem_hp30_accfuse.py`; HP τ=30 on IRM↔synth fuse then soft-w with AccFuse tip / af_raw / lin_hp; ≠ MoE fuse↔hp30 / IRM↔SWA blend+hp30 / esr-hp30): LODO-select wrongly picked `tip_x_stem_raw` w=0.5 → LODO **0.8845/83.46**; AASD **0.6365/51.87**; Carta **0.5926/53.85**. Useful baselines: stem_hp_alone **0.8354/74.29**; tip_x_stem_hp@0.5 **0.8504/76.53**. Kept as discovery helper. `results_stem_hp30_accfuse.md`.

- **stem-hp30-accfuse-ext** (`stem_hp30_accfuse_ext.py`; external gate for tip↔stem_hp / af↔stem_hp Acc>75 grid): **no joint pass**. Best LODO-pass Carta among τ=30 blends: af_raw_x_stem_hp w=0.25 Carta **0.6241**. tip_x_stem_hp w=0.35 LODO **0.8513/76.94** Carta **0.6216**. Script deleted after log. `results_stem_hp30_accfuse_ext.md`.

- **volgate-accfuse-hp30** (`volgate_accfuse_hp30.py`; past-only EWMA |ΔAccFuse| volatility gate AccFuse↔lin_hp30; ≠ hier-dwell / MoE |fuse| / domain-emis): select soft_sigmoid t4 thr0.2 ss2 → LODO **0.8015/71.37**; AASD **0.6420/53.01**; Carta **0.6526/60.57**. ≈ tip; no LODO lift. Script deleted. `results_volgate_accfuse_hp30.md`.

- **multitau-stem-hp-accfuse** (`multitau_stem_hp_accfuse.py`; multi-τ / soft_hp / bank on IRM↔synth stem then tip soft-w; ≠ 4-seed multi-τ / SW2 multi-τ / τ=30-only): **no joint pass**. Best Carta among LODO-pass: **soft_hp50@w=0.15** LODO **0.8542/77.96** Carta **0.6427** AASD **0.6421**. Kept as discovery. `results_multitau_stem_hp_accfuse.md`.

- **capped-stemhp-resid-tip** (`capped_stemhp_resid_tip.py`; tip+α·clip(stem_hp−tip,±cap); ≠ capped-odd on lin_hp30 / soft-w): **no joint pass**; never cleared LODO H>0.85 (peak **0.8452/76.14** @a0.5 cap2 Carta **0.6212**). Best Carta near Acc>74: a0.2 cap1.5 Carta **0.6451**. Script deleted. `results_capped_stemhp_resid_tip.md`.

- **soft-hp35-tip-refine** (`soft_hp35_tip_refine.py`; dense w×p_stay on soft_hp{30,35,40}↔tip): **no joint pass**. Best Carta among LODO-pass: **soft_hp40@w=0.14** LODO **0.8510/77.93** CartaMax **0.6443**. Kept as near-miss helper. `results_soft_hp35_tip_refine.md`.

- **agree-tip-softhp50** (`agree_tip_softhp50.py`; sign-agree tip↔soft_hp50; ≠ IRM-synth agree-gate): select agree_mean_tip LODO **0.8354/75.79** Carta **0.6420** (H short); soft disagree → Carta collapse. Script deleted. `results_agree_tip_softhp50.md`.

- **light-hp-anchor-resid** (`light_hp_anchor_resid.py`; tiny GRU resid on [tip,stem_hp] + λ‖out−tip‖²; smoke λ→0.1; ≠ fuse_conf_resid / stem_ms96): LODO MEAN **0.8536/83.45** (clears new LODO bars); AASD **0.6404/51.56**; Carta **0.6070/55.49**. External H collapses like stem family. Script+ckpts deleted. `results_light_hp_anchor_resid.md`.

- **soft-hp50-accfuse-final** (`soft_hp50_accfuse_final.py`; AccFuse↔lin_hp30 α=0.50 tip × soft_hp50(IRM↔synth) **w=0.15**): **KEPT NEAR-MISS TIP** for stricter LODO bars. Operating **p=0.996/0.8** LODO **0.8537/77.79**; AASD **0.6431/53.11**; Carta **0.6432/59.00**. Clears LODO H>0.85 Acc>75; Carta short by ≈0.0068. Do not re-grid this exact recipe without a new structural hypothesis. `results_soft_hp50_accfuse_final.md`.

### Ceiling note (stricter joint target)

Across soft-w / capped-resid / agree-gate / light-GRU of AccFuse-tip ↔ HP'd stem, every config that clears **LODO H>0.85 ∧ Acc>75** tops out at **Carta H ≤ ~0.644** (best: soft_hp50@0.15 → 0.6427; soft_hp40@0.14 → 0.6443@0.996). Tip alone keeps Carta **0.652** but LODO only **0.801/71.4**. Soft linear mixing of stem Acc into the HP tip appears to hit a Pareto wall ~0.007 below test H≥0.65 once LODO clears the new bars. Further progress needs a qualitatively different bias-control mechanism than soft/hard stem mixing into AccFuse↔hp30.

## lead-optimizer session (2026-11-09) — DO NOT REPEAT

Why-H analysis: `multi_agent_opt/results_why_assd_carta_h.md`. Reference tip `soft_hp50_accfuse_final` @**p=0.998/sc=1.0** LODO **0.8549/78.43** (op **0.996/0.8** AASD **0.6431** Carta **0.6432**).

- **ac-transplant-tip-stem** (`ac_transplant_tip_stem.py`; tip+α(stem_hp−tip_hp) / LP-split / midband; ≠ soft-w / capped tip+α(stem_hp−tip)): LODO-max top clears **0.8747/80.26** but AASD/Carta **~0.61/0.58**. Mild rescore (`ac_transplant_mild_ext.py`): LODO-pass needs α≳0.15–0.20 → Carta ≤**0.6387** (best mild LODO-pass `acresid_t50_a0.18`). Midband collapsed. No joint pass. `results_ac_transplant_tip_stem.md`, `results_ac_transplant_mild_ext.md`.
- **emis-sigma-gate-tip** (`emis_sigma_gate_tip.py`; fixed tip/soft_hp50; past-only EWMA |d|/|Δd| → σ multiplier via evidence÷sched; ≠ volgate blend / domain-emis): chosen ≈ softhp50 noscale; tip gates keep Carta≥0.65 but LODO~0.80; softhp50 σ-gates do not lift ext H. No joint pass. `results_emis_sigma_gate_tip.md`.
- **dyn-pstay-softhp50** (`dyn_pstay_softhp50.py`; switch-amplify proxy boost·sign(Δ)·gate on frozen soft_hp50 tip; intent=lower p_stay at swings; ≠ hier-dwell / volgate): base L_A≈0.813 L_C≈0.697; boosts flat. No joint pass. `results_dyn_pstay_softhp50.md`.
- **light-hp-resid-tf** (`light_hp_resid_tf.py`; out=tip+HP_30(TinyCausalTF[tip,stem,stem_hp,tip_hp]); 500 steps/fold; hard HP resid; ≠ GRU soft-anchor / v1–v4 logit TF): LODO **0.8759/82.66**; AASD **0.6284**; Carta **0.6084**. Failing ckpts deleted. `results_light_hp_resid_tf.md`.
- **elm-hp-resid** (`elm_hp_resid.py`; tip+α·HP_30(ELM tanh ridge on 4-d feats); CPU; ≠ light TF): α=0.5 LODO **0.8517/78.97** Carta **0.6287**; α=0.25 Carta **0.6423** LODO H short. Same Pareto. `results_elm_hp_resid.md`.
- **ac-transplant-softhp50** (`ac_transplant_softhp50.py`; tip+α(HP(soft_hp50)−tip_hp) / tip+α·HP(soft_hp50); ≠ soft-w tip↔soft_hp50 / stem_hp AC-transplant): never jointly clears; LODO-pass α≳0.20–0.30 Carta ≤**0.632**; tiny α keeps Carta≥0.65 but LODO H≪0.85. `results_ac_transplant_softhp50.md`.
- **decision-fuse-tip-softhp** (`decision_fuse_tip_softhp.py`): abandoned — LODO AccFuse evidence lacks `n128`/`att128` for official expand path. Superseded by evidence |Δ|-gate below. No metrics.
- **delta-gate-tip-softhp** (`delta_gate_tip_softhp.py`; soft_hp50 default, tip when EWMA(|Δsoft|)≥thr; hard/soft sigmoid; ≠ volgate AccFuse↔hp30): best LODO-pass `hard_switch_tip@p75` LODO **0.8525/77.83** AASD **0.6424** Carta **0.6429**; soft_sigmoid lifts Carta to **0.6485** but LODO H **0.835**. No joint pass. Same ~0.007 Pareto wall. `results_delta_gate_tip_softhp.md`.

### Session ceiling update (2026-11-09)

Still **no** config clears **LODO H>0.85 ∧ Acc>75 ∧ Assd H≥0.65 ∧ Carta H≥0.65**. Best near-miss remains **soft_hp50_accfuse_final** @**p=0.996/0.8**: LODO **0.8537/77.79**, AASD **0.6431**, Carta **0.6432** (also **p=0.998/sc=1.0** LODO **0.8549/78.43**). AC-transplant / ELM / light-TF / σ-gate / |Δ|-gate / switch-amplify do not break the wall. Further work needs a new front-end or train-time bias control, not another tip↔stem readout mix.



