# results_bipolar_vlaai

**Abandoned — chance probe.** No LODO / AASD / Carta scores.

## Method

Nearest-neighbour bipolar (each of 58 channels − its single anatomical NN), then causal EA with Fuglsang prior mapped `L R L^T` (`bipolar_eaP`), into `vlaai_ws` 10 s + sw2 + SPK w10_b3. C=58 so SPK shapes match.

Distinct from `lap_eaP` (ch − mean of ≤k NN), Hjorth stack, FIR-fb residual, denser decoder seg_switch.

## Gates

| Check | Result |
|---|---|
| Swap seed-search | **PASS** seed 4 Δ=**0.0113** (≥0.01) |
| AVGC chance @50 | **ABORT** mean loss **0.6886** ≥ 0.685 |

Earlier smoke (EA then on-the-fly bipolar, SPK) also aborted: loss **0.6919** @50 (swap seed 6 Δ=0.0166).

## Vs tip

No Acc scores. Tip remains SW2↔swmix **53.53** / linear_hp30 **53.12**.

## Code / artifacts

- `bipolar_vlaai_lodo.py`, `derive_bipolar_eap.py`
- `prepare_data.nearest_bipolar` / `bipolar_prior`; config prep `bipolar_eaP`
- Caches: `cache/bipolar_eaP` + external AASD/Carta/Fuglsang
- Log: `experiments/bipolar_vlaai_lodo.log`

## NEXT IDEA

Close further **spatial remaps → vlaai Acc** (lap / bipolar / Hjorth already chance or Acc≪60). Prefer non-front-end Acc: reliability / evidence calibration on frozen published streams, or the open `sph-lap-vlaai` / `wavelet-fb-vlaai` slots if they stay distinct. **Do not** denser-seg_switch decoder again.
