# results_hjorth_vlaai

**Abandoned at AVGC chance-probe** (no LODO / AASD / Carta).

## Method
Causal per-channel **Hjorth** activity / mobility / complexity (sliding win **0.5 s = 16** samples @32 Hz) stacked with raw `eaP_n5_t60` EEG → **C=232** into `vlaai_ws` 10 s classifier (`sw2`). Speech streams unchanged.

- SPK w10_b3 init: **shapes disallow** (SPK C=58 ≠ 232) → **random init**
- DISTINCT from lap_eaP / lp7 / FIR-fb residual / channel-mask / graph-adj / SE / multi-rate envelopes
- Code: `hjorth_vlaai_lodo.py`
- Log: `experiments/hjorth_vlaai_lodo_console.log`, `experiments/hjorth_vlaai_lodo.log`

## Gates
| Gate | Result |
|---|---|
| Swap-test Δ>0.01 | **PASS** seed 4 → Δ=**0.0133** (seeds 0–3 below floor; train_seed→20270105) |
| Antisymmetry | anti_err 0 |
| AVGC chance @50 steps | **FAIL** mean loss **0.6932** ≥ 0.685 (near ln2) |

## Scores
No LODO mean / AASD / Carta — aborted before fold score.

| Baseline | Acc |
|---|---:|
| Acc-stream SW2↔swmix (best) | 53.53 |
| linear_hp30 AASD | 53.12 |
| linear_hp30 Carta | 60.54 |
| **hjorth-vlaai** | n/a (chance) |

## NEXT IDEA
Stop stacking per-channel time-domain descriptor banks into `vlaai_ws` without SPK-compatible shapes (same fate as FIR-fb / WaveNet / Deep / S4 / TFDirect). Prefer Acc ideas on **frozen published logits** (reliability / evidence calibration / causal fusion) rather than another expanded-C front-end.
