---
name: full-corpus-eda
description: Full-corpus signal-processing EDA (phases A–E, H1–H12) for the ICASSP 2027 EEG-sAAD challenge in reports/eda_full — COMPLETE as of 2026-09-24; where outputs live and key findings
metadata:
  type: project
---
Completed 2026-09-24. Outputs are in `reports/eda_full/`:
- `EEG-sAAD_full_corpus_EDA_report.pdf` (the user asked for a PDF), plus README.md (same content) and NEXT_STEPS.md;
- tables/, figures/, scripts/ (`run_all.sh` driver; f2_report.py regenerates both reports from the tables).

**Why:** the user wanted H1–H12 from an n=1-trial EDA confirmed or refuted at corpus scale. The first attempt died because `EEG-sAAD Data Story.html` (1.6 MB) was attached whole; never read that file whole.

**How to apply:** reuse the caches and tables rather than recomputing. MATLAB/Octave are absent, so scripts/scorer.py is the verified Python port of the official scorer. Another session edited scripts during the run (d3 checkpointing, c1 FFT null). Never use `pkill -f` on script names: the queue's own `bash -c` wrappers match them too.

Key findings:
- Verdicts: H9 refuted (USTC columns not swapped); H2, and H7 for KUL-AV-GC, partly; all others confirmed.
- KUL-AV-GC: the attended speaker never switches; the attended direction flips at 300 s in all 99 trials.
- KUL-AV-GC streamWav names are inconsistent with the audio (29/30 envelope clusters carry several names). KUL-UHD audio ⊂ KUL2016.
- LODO SNR1 ≈ 0.07–0.13; accuracy ≈ Φ(0.9·SNR1·√w).
- Spatial decoders are at chance under LODO. The best integrator is a 10-s MA (H 0.63–0.75 on synthetic switches).
- Scorer quirk: a perfect decoder gets SDT=Inf when a reverting switch follows within 50 s.

2026-09-25: the user asked for one PDF combining both EDAs for new team members. It is `reports/onboarding/EEG-sAAD_onboarding_guide.pdf` (18 pages), rebuilt by `make_figures.py` then `build_guide.py`; tables are read live from reports/eda_full/tables. The f2_report placeholder bug and the Table 17 tuple formatting were fixed the same day.
