# Parent report — Performance H / external Acc push (2026-11-07)

## Verdict

**Target unmet.** Interpreted as Acc %: AASD Acc > 70 and Carta Acc > 60, with Performance H as the ranking metric.

| Set | Best Acc this session | Bar | Status |
|---|---:|---:|---|
| AASD | **53.80** (SR-Pool) / hist tip **54.05** | > 70 | **FAIL** (gap ≈ −16) |
| Carta2026 | **61.87** (BMMS3 mean4 + hp30) | > 60 | **PASS** |

Best Performance H (old scorer, preset HMM): BMMS hp30 Carta **0.6704**; AASD SR-Pool **0.6482**. Neither package jointly clears AASD Acc > 70.

**Hard ceiling (prior R2 diagnostics, confirmed):** AASD median dwell 8.9 s; oracle 5 s delay-cut Acc ≈ 61 %; linear per-subject ceiling ≈ 55 %; D0 found no M/M recording deficit. **AASD Acc > 70 is not reachable under current causal/data constraints** without a qualitatively stronger EEG↔stimulus coupling than any front-end on disk.

---

## Reused wins vs abandoned techniques

**Reused (kept as baselines / ingredients)**
- Prep `eaP_n5_t60`, DirectNet `vlaai_ws` 10 s × 3 blocks, `sw2`, SparrKULee init
- BMMS3 enhanced (correlation DensePath) — only Acc front-end that lifts Carta Acc
- External protocol: emissions from train only; HMM preset 0.994 / scale 0.8; `eval_external.score`
- `linear_hp30` readout (helps Carta Acc)

**Abandoned / closed this session**
| Method | Reason | Acc AASD / Carta |
|---|---|---|
| AADNet-corr (random BCE) | chance 0.6926 @50 steps | — |
| AADNet-MM (200 MM → BCE) | post-MM chance 0.6863 | — |
| Acc-fuse SW2↔BMMS w=0.5 | below tip Acc | 52.96 / 59.89 |
| FUT1 (LA=1.0 s) | Acc regress vs tip | 51.64 / 58.25 |
| SR-Pool tp=4 s | LODO Acc drop; Carta < 60 | 53.80 / 59.10 |

**Still closed from prior inventory (do not rerun):** tip Acc PP / emission-shift / decision stacking / DANN / VICReg-CPC / SpecViT / speech-xattn / Path A TRF / bipolar-vlaai / xcorr CNN / MM-TTA / SAA / PAD (R2 gates failed).

---

## New techniques (code / config paths)

| ID | Idea | Path |
|---|---|---|
| Acc-fuse | LODO-selected soft blend SW2 mean4 ↔ BMMS3 mean4 | `acc_fuse_sw2_bmms.py`, `results/research/acc_fuse_sw2_bmms.json`, `multi_agent_opt/results_acc_fuse_sw2_bmms.md` |
| AADNet-corr | Separate EEG/stim temporal encoders + cosine bottleneck | `aadnet_corr_lodo.py` (abandoned) |
| AADNet-MM | Label-free match–mismatch warm-start then BCE | `aadnet_mmwarm_lodo.py` (abandoned) |
| FUT1 | `causal_future_s=0.75` → legal LA=1.0 s | `fut1_sw2_lodo.py`, folds kept under `folds/direct_clf__eaP_n5_t60_w10b3_ptSPK_fut1_aug-sw2/` |
| SR-Pool | Training-free short `tw` re-pool + LODO-fast p_stay | `sr_pool_sw2.py`, `results/research/sr_pool_sw2.json` |
| BMMS×hp30 | CPU rescore mean4 BMMS with τ=30 hp | `multi_agent_opt/results_bmms_hp30_mean4.md` |

Claims: `multi_agent_opt/CLAIMS.md` (perf-h-agent rows 2026-11-07).

---

## Per-model metrics and deletions

| Run | LODO H/Acc | AASD H/Acc | Carta H/Acc | Deletions |
|---|---|---|---|---|
| Acc-fuse w=0.5 | 0.7947 / 72.64 | 0.6397 / 52.96 | 0.6482 / 59.89 | none (CPU) |
| AADNet / AADNet-MM | — | — | — | no ckpts written |
| FUT1 | 0.7885 / 71.67 | 0.6330 / 51.64 | 0.6386 / 58.25 | deleted `models/fut1_sw2/*.pt`, `all5_*fut1*.pt`; kept folds + JSON + evidence pkl |
| SR-Pool tp4 p0.994 | 0.7771 / 69.73 | 0.6482 / **53.80** | 0.6374 / 59.10 | none (inference-only) |
| BMMS mean4 hp30 | — | 0.6473 / 53.71 | 0.6704 / **61.87** | none |

---

## Surviving artifacts & reproduce

Python: `/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/.venv/bin/python`  
Workdir: `/root/disk_3tb/eeg-thao`  
`export GLIMPSE_THR=onorm`

```bash
# Acc-fuse (CPU)
python -u acc_fuse_sw2_bmms.py

# BMMS hp30 tip (CPU snippet in results_bmms_hp30_mean4.md)

# FUT1 (GPU; ~2 h) — already scored; re-run only if needed
python -u fut1_sw2_lodo.py

# SR-Pool (GPU inference; ~40 min)
python -u sr_pool_sw2.py
```

External data: `/root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset/experiments/external_data/{AASD,Carta2026}`  
Published SW2/BMMS evidence: `.../decoder_lodo/results/research/ext_clf_*_{w10b3sw2P,w10bmms3enhP}_evidence_s{0..3}.pkl`

---

## Blockers

1. **AASD Acc > 70 exceeds measured ceilings** (~55 % linear, ~61 % oracle delay-cut, ~0.60 steady 10 s evidence).
2. D0: AASD match–mismatch coupling is *not* weaker than in-domain → adaptation/TTA unlikely to unlock +16 Acc.
3. Single V100 shared with decoder_lodo M1 jobs; flock on `multi_agent_opt/gpu.lock` only coordinates eeg-thao agents.

---

## Next best experiment (if continuing)

**B1 wider-band prep from organiser raw 128 Hz** (0.5–16 or 1–32 Hz, 64 Hz, causal IIR) → retrain SW2 + BMMS3 seeds — only major unused SNR lever in `ACC_EXPERIMENT_PLAN.md`. Secondary: M3/M4 (AADNet-style *on top of vlaai features*, or CA-TCN using full 1 s) only after B1 screen Acc1 on AVGC/AASD. Do **not** chase further tip Acc PP or AADNet-from-scratch cosine bottlenecks.
