# Next steps: five modelling experiments this EDA justifies

Each one names the finding it rests on (tables and figures are in `reports/eda_full/`).

1. **Integrator tuning on LODO evidence against the real switch schedule.**
   A 10-s moving average and a variance-normalised HMM already sit near the Acc-L frontier. They reach H of about 0.63–0.75 on SYNTHETIC switches, against about 0.49–0.60 for the sign-randomised null (`tables/switch_results_*`).
   - Before tuning further, fix the switch-spacing assumption. The official scorer charges a *perfect* decoder SDT = Inf on any switch followed within 50 s by a reverting switch (`tables/scorer_equivalence_tests.csv`, cases O*). That happens to 16% of the default synthetic switches.
   - Tune p_stay and the MA memory on a spacing distribution that matches the blind-test protocol once it is known, and select on the *median* held-out dataset.

2. **Raise SNR₁, not the integrator.**
   - LODO SNR₁ is 0.107 [0.094, 0.132] on KUL-AV-GC, against 0.185 [0.162, 0.205] within-subject (`tables/decoder_snr_law.csv`).
   - Accuracy follows Φ(k·SNR₁·√w) with k ≈ 0.74–0.98, so every 0.01 of SNR₁ is worth as much as a much longer window.
   - Try in this order:
     1. a nonlinear/DNN envelope decoder on the same c19 causal pipeline;
     2. CCA with a stimulus filter bank;
     3. test-time running Euclidean alignment. It removes much of the dataset fingerprint (99.7% → 87%) without labels.

3. **Leakage-proof validation by default.**
   - Use folds grouped by *envelope-identity cluster* (`tables/stim_identity_clusters.csv`), not by `streamWav` name. The KUL-AV-GC names are inconsistent with the audio for 29/30 audios.
   - KUL2016 and KUL-UHD share all their audio, so hold them out together.
   - Never score random windows from training trials. Spatial decoders gain up to 49 pp from trial fingerprints, and within-trial splits of 57–93% collapse to chance under LODO.

4. **Direction (spatial) decoding needs a new idea or should be dropped.**
   - FB-CSP and Riemannian classifiers are at chance under LODO (49–54%).
   - On the 99 genuine KUL-AV-GC direction switches they give Acc ≈ 51%, with L barely above the null.
   - The only robust physiological side marker is the USTC HEO channel (d = 0.43). The 1 Hz high-pass in KUL-AV-GC removes most EOG information.
   - Test whether the blind sets supply stream-position metadata. If they do, a speaker decoder plus the stream-azimuth mapping solves direction; if not, direction labels must come from speaker decoding plus priors.

5. **Harmonise the front end, then re-test the small pipeline effects.**
   - The causal 1–9 Hz pipeline with CAR is never worse than zero-phase filtering or the supplied 32 Hz data, and dropping CAR costs 0.3–1.4 pp (`tables/decoder_pipeline_comparison.csv`).
   - Add robust edge handling (UHD and USTC have transients in 100% of trials) and channel interpolation (UHD 5% bad channels).
   - Then check whether the gap to KUL2016/DTU closes for USTC, the weakest held-out dataset (LODO accuracy 52.8% at 1 s).
