# Full-corpus signal-processing EDA: ICASSP 2027 EEG-sAAD Grand Challenge training data

Scope: every trial of the five local training datasets (1979 trials, 95 subjects, 77.0 h of EEG at 128 Hz). A previous EDA based on one example trial per dataset proposed hypotheses H1-H12; this report confirms or refutes each at corpus scale with subject-level statistics (medians with 95% bootstrap CIs over subjects, 2,000 resamples; Wilcoxon signed-rank tests across subjects; FDR where many tests are made). Every number was computed by the scripts in reports/eda_full/scripts and is traceable to a CSV in reports/eda_full/tables.

## 1. Executive summary

- Labels (H6 confirmed): the corpus contains no within-trial speaker switch. The only genuine switches are 99 direction switches in KUL-AV-GC, all at exactly t = 300 s, where the attended speaker stays the same and moves side. The trial-level label is wrong for 50% of KUL-AV-GC direction samples. A cross-half decoder test agrees: 84.4 [79.7, 89.1]% at 60 s if the speaker is constant, vs 20.3 [12.5, 28.1]% if it switched.
- Recording 'dialects' (H1, H3 confirmed): median channel SD spans 16.5x across datasets; DTU, KUL-UHD and USTC are not high-passed (RMS/SD 211, 51.4, 97.6) and have edge transients in 100% (KUL-UHD) and 100% (USTC) of trials. KUL-AV-GC is band-limited to about 1-41 Hz and notched at 50 Hz.
- Neural tracking is present in every dataset. The envelope-EEG correlation peaks at 93.8-188 ms, and the attended envelope is tracked more strongly than the ignored one in all five datasets (Wilcoxon p <= 0.01).
- SNR law (H4 confirmed): accuracy follows Phi(k*SNR1*sqrt(w)) with k = 0.74-0.98. KUL-AV-GC 1-s SNR1 is 0.185 within-subject and 0.107 leave-one-dataset-out (LODO), close to the 0.099 implied by the official baseline's 53.94% at 1 s.
- Cross-dataset (LODO) accuracy at 1 s is 52.8%-55.6%, and at 10 s it is 56.4%-66.5%. USTC (NeuroScan, Mandarin) is the hardest held-out dataset.
- Pipeline (H5 confirmed): causal filtering is never worse than zero-phase filtering or the supplied 32 Hz data under LODO, and common-average referencing helps. Montage size (57/58/62 channels) and lag range (0-250 vs 0-500 ms) make no reliable difference.
- Shortcuts: the label is locked to trial-ID parity in KUL-UHD and USTC (H8, 100%). In KUL2016 an envelope-statistics classifier picks the attended column at 82.5% under LOSO (H7). The covariance identifies the dataset at 99.7% (H10), and alignment only reduces this to 84.3% (EA) or 87.1% (causal running EA).
- Spatial decoders learn trial fingerprints (H11 confirmed). With 5-s windows, the trial ID is predicted at 63%-89%. Going from a within-trial split to held-out trials costs a median 22.4 pp, and LODO accuracy is at chance (49.0%-54.2%).
- USTC columns are not swapped (H9 refuted): column 1 wins only 25.3% of USTC trials under LOSO, and no subject shows a significant column-1 majority.
- Stimulus reuse: all KUL-UHD audio also occurs in KUL2016. In KUL-AV-GC the streamWav names do not match the audio: 29/30 envelope-identity clusters carry several names. Build story-disjoint folds from envelope identity, not names.
- Switch detection (Phase E, SYNTHETIC switches, LODO evidence): the best integrator on the median held-out dataset is ma_10s, with H = 0.635-0.75 against a sign-randomised null of 0.493-0.597. On the 99 genuine KUL-AV-GC direction switches, spatial LODO evidence gives Acc = 51.7% and L = 0.604 (null L 0.451): direction is essentially not decodable across datasets.
- Scorer: MATLAB is not installed, so the official scorer was ported to Python and verified (34/34 tests plus 3,000 random cases against a brute-force reference). The port exposed a property of the metric: a perfect decoder scores SDT = Inf on a switch that is followed within 50 s by a switch back (414/3113 synthetic switches). The full causal pipeline passes the lookahead perturbation test (21/21).

## 2. How to read this report

Colours: every figure uses one fixed colour per dataset - KUL-AV-GC red, DTU blue, KUL2016 green, KUL-UHD purple, USTC orange. Grey marks a baseline or null. Numbers are summarised per subject first and then across subjects, so a subject with many trials counts once. Unless stated otherwise, accuracies are medians over subjects with 95% bootstrap CIs in brackets, and n (subjects, trials or windows) is given in every caption. Chance level is 50% for every two-way decision. Each figure's plotted values are also saved as a CSV next to the PNG in figures/.

Every figure and table is followed by a shaded 'How to read this' note explaining its axes, colours and what to look for. Terms used throughout:

*Glossary.*

| term | meaning |
|---|---|
| Trial | One continuous recording (50 s to 10 min) in which a listener hears two talkers and attends to one. |
| Column 0 / column 1 | The two envelope streams stored per trial. Column 0 is the stream attended for most of the trial (converter convention); column 1 is the other one. |
| Envelope | The slow loudness contour of a talker's speech (roughly 1-9 Hz). The decoder tries to reconstruct it from EEG. |
| Backward decoder | A linear filter that maps the last 0-500 ms of EEG after each moment onto the speech envelope at that moment. The talker whose envelope correlates better with the reconstruction is predicted as attended. |
| r(att), r(ign), d | Correlation of the reconstruction with the attended and ignored envelope over a window; d = r(att) - r(ign). d > 0 means the window was decoded correctly. |
| SNR1 | mean(d) / SD(d) over 1-s windows: how strong the 1-s evidence is compared with its noise. Accuracy follows Phi(k * SNR1 * sqrt(window)) (the SNR law, H4). |
| LOTO | Leave-one-trial-out within a subject: train on the subject's other trials, test on the held-out trial. |
| LOSO | Leave-one-subject-out within a dataset: train on the other subjects of the same dataset. |
| LODO | Leave-one-dataset-out: train on four datasets, test on the fifth. This is the closest proxy for the blind test sets (new subjects, new recording setup). KUL-AV-GC is the main proxy fold. |
| Within-trial split | Random windows split into train/test with windows of the same trial on both sides. Leaky; shown only to measure how much a decoder memorises trials. |
| Median [95% CI] | The median over subjects, with a 95% bootstrap confidence interval (2,000 resamples of subjects). If two CIs do not overlap, the difference is very unlikely to be chance. |
| p, q | p = Wilcoxon signed-rank test across subjects (or binomial test across trials). q = p corrected for many tests (Benjamini-Hochberg FDR). Below 0.05 is conventionally 'significant'. |
| Cohen's d | Effect size: difference of means divided by the pooled SD. About 0.2 is small, 0.5 medium, 0.8 large. |
| Acc, SDT, L, H | The official metrics. Acc = % of 128 Hz samples labelled correctly. SDT = switch detection time in s (Inf = missed). L = median over switches of exp(-SDT/60), averaged over subjects (1 = instant, 0 = missed). H = harmonic mean of Acc and L; the headline score (official baseline H = 0.70). |
| SYNTHETIC switch | A switch imposed on a fixed-attention trial by swapping which stream counts as attended from a random time on (same rule as the official generateSyntheticSwitches.m). It measures reaction speed on data where no real switch happened. |
| Integrator | A causal post-processor that turns the noisy 1-s evidence into stable labels: argmax (none), moving average, HMM or a state ladder. More smoothing = higher Acc but later switch detection. |
| Null (sign-randomised) | The same pipeline fed with evidence whose sign is randomised: what a decoder with zero information scores. |
| CAR | Common average reference: subtracting the mean of all channels at each sample. |
| EA / running EA | Euclidean alignment: whitening each recording by its own mean covariance so recordings look alike. Running EA uses only the past (usable at test time). |
| Causal / zero-phase | Causal filters use only past samples (required by the 1-s lookahead rule); zero-phase filters also use future samples (not allowed at test time). |
| Montage 57 / 58 / 62 | Channel sets matched by name: 57 common to all five datasets, 58 = BioSemi-64 ∩ USTC, 62 = BioSemi-64 ∩ KUL-UHD. |

## 3. Most interesting findings

These are the results that most change how the data should be used. Each one says what was found, how strong the evidence is, and why it matters for building and evaluating a decoder.

**Finding 1. The official scorer can mark a perfect decoder as having missed a switch.**

When a switch is followed within 50 s by a switch back, the scorer's split search prefers 'old label until the end of the 100-s window'. That split fails the 0.5-s sustain check, the search moves past it, and the first switch is scored undetected (SDT = Inf, score 0) even when every predicted label is correct. Under the default synthetic schedule (gaps drawn from 30-90 s) this happens to 414 of 3113 switches (13.3%) for a perfect decoder.

*Why it matters:* L is a median over switches, so this loss is invisible in the headline L of the perfect decoder (1.00) but not for real decoders. Synthetic switch spacing must match the real test protocol before integrators are tuned, and results on closely spaced switches are pessimistic by construction.

**Finding 2. There is no speaker switch anywhere in the training data; the KUL-AV-GC 'switch' is a change of side.**

In all 99 KUL-AV-GC trials the attended talker stays the same and moves from one side to the other at exactly t = 300 s. A decoder trained on one half and tested on the other scores 84.4% at 60 s if the talker is constant and 20.3% if it had switched. The trial-level label is wrong for 50% of direction samples.

*Why it matters:* Speaker-switch detection can only be trained and evaluated with SYNTHETIC switches. A stimulus-reconstruction decoder cannot see these genuine switches at all, and direction labels must be taken per sample, never from the trial label.

**Finding 3. Stimulus names are unreliable, and two datasets share the same audio.**

Clustering all 3,958 envelopes by content shows that in KUL-AV-GC 29 of 30 distinct audios carry more than one streamWav name across subjects, so the same name does not mean the same audio. All 240 KUL-UHD envelopes also occur in KUL2016, and KUL2016 itself has only 8 distinct audios.

*Why it matters:* Folds that separate stories by name leak stimuli between training and test data. Group folds by envelope identity, and keep KUL2016 and KUL-UHD on the same side of any split.

**Finding 4. Spatial decoders learn which trial a window came from, not where the listener attends.**

With 5-s windows the same features predict the trial ID at 63%-89% (chance 2-25%). Accuracy drops by a median 22.4 pp (up to 49 pp) when whole trials are held out, and across datasets it is at chance (49.0%-54.2%). In USTC and KUL-UHD it even falls clearly BELOW chance, because labels alternate trial by trial, so the neighbouring trials that look most similar carry the opposite label.

*Why it matters:* Any spatial or direction decoder evaluated on random windows is inflated. On the 99 genuine KUL-AV-GC switches the best spatial decoder reaches only 51.7% sample accuracy.

**Finding 5. Shortcuts are built into the design of three datasets.**

The trial label equals trial-number parity in 100% of KUL-UHD and USTC trials. In KUL2016 the attended envelope is simply louder in 65% of trials, and a classifier that sees only envelope statistics picks the attended stream at 82.5% for unseen subjects, without looking at the EEG.

*Why it matters:* A model can score well on these datasets without decoding attention. Envelope inputs should be scale-symmetric, and no feature may carry trial order or slow drift.

**Finding 6. Cross-dataset decoding works, but at about half the within-subject signal-to-noise ratio.**

Every dataset shows neural tracking, with the attended talker tracked more strongly than the ignored one (peak 93.8-188 ms). The 1-s evidence SNR is 1.4-2.3x higher within subject than across datasets. Accuracy follows Phi(k * SNR1 * sqrt(window)) with k = 0.74-0.98, and USTC (Mandarin, NeuroScan) is the hardest held-out dataset (52.8% at 1 s).

*Why it matters:* Improving SNR1 (better decoders, better alignment) is worth more than tuning the integrator, which changes H by only 0.019-0.068 over plain argmax.

**Finding 7. One suspected problem turned out not to exist.**

The example-trial EDA suggested the USTC stream columns were swapped. At full scale, column 1 wins only 25.3% of USTC trials and no subject shows a significant column-1 majority. USTC is simply the weakest dataset, not a mislabelled one.

*Why it matters:* USTC labels can be used as they are. Its difficulty (envelopes with a different extraction, no high-pass, edge transients in every trial) is a front-end problem, not a labelling problem.

## 4. Corpus inventory

*Table 1. Corpus inventory (SuperHuge raw stage, 128 Hz; mapping to the CND sources verified sample by sample: EEG max abs. difference 0, envelope r = 1.000 for every trial).*

| dataset | subjects | trials | hours | trial length s (median, range) | stored channels | distinct audio (envelope clusters) | EOG channels | conditions |
|---|---|---|---|---|---|---|---|---|
| KUL-AV-GC | 13 | 99 | 16.5 | 600 (600-600) | 64 | 30 | yes | FixedVideo, MovingTargetNoise, MovingVideo, NoVisuals |
| DTU | 18 | 1080 | 15.0 | 50 (50-50) | 64 | 124 | yes | Anechoic, HighReverberation, MildReverberation |
| KUL2016 | 16 | 320 | 20.5 | 124 (124-399) | 64 | 8 | no | dry, hrtf |
| KUL-UHD | 30 | 120 | 13.0 | 391 (389-395) | 253 | 4 | no | - |
| USTC | 18 | 360 | 12.0 | 120 (120-120) | 64 | 80 | yes | - |

> **How to read this:** One row per dataset. 'Distinct audio' counts different speech recordings found by comparing envelopes, not file names - a small number means the same stories are reused across trials and subjects. 'EOG channels' says whether eye-movement electrodes were stored.

Montages (matched by upper-cased channel name): common intersection of all five datasets = 57 channels (BioSemi-64 minus AF7, AF8, P9, P10, Iz, AFz, FCz); BioSemi-64 ∩ USTC = 58 (missing AF7, AF8, P9, P10, Iz, AFz; identical to the official baseline's list); BioSemi-64 mapped onto KUL-UHD = 62 (KUL-UHD lacks AFz and FCz). The BioSemi datasets spell AFz as 'Afz', USTC uses upper case throughout (FP1, FPZ, ...), so index-based pooling would be wrong.

## 5. Phase A: signal quality and recording 'dialects'

*Table 2. Phase A, per dataset (subject-level medians; 149,336 channel-trials from 1,979 trials). Bad channel: robust SD > 5x the dataset median; bad trial: >= 10% bad or flat channels; edge transient: first-difference jump > 20 robust SD within 0.5 s of either trial edge in any channel.*

| dataset | median ch. SD [95% CI] | RMS/SD | |DC| | 1/f slope | alpha peak Hz | 50 Hz ratio | sub-1 Hz ratio | edge-transient trials | bad channels | bad trials |
|---|---|---|---|---|---|---|---|---|---|---|
| KUL-AV-GC | 7.96 [6.829, 8.808] | 1 | 0.000861 | 1.21 | 9.88 | 0.0695 | 0.0244 | 3.0% | 0.13% | 0.0% |
| DTU | 39.1 [36.71, 39.8] | 211 | 8.49e+03 | 1.4 | 9.56 | 1.44 | 6.05 | 2.8% | 2.33% | 3.0% |
| KUL2016 | 9.64 [8.546, 11.38] | 1 | 0.00828 | 1.04 | 10.1 | 1.06 | 1.31 | 4.7% | 0.00% | 0.0% |
| KUL-UHD | 131 [118.9, 148.8] | 51.4 | 7.87e+03 | 1.12 | 9.94 | 3.8 | 6.35 | 100.0% | 5.14% | 16.7% |
| USTC | 63.6 [49.32, 71.49] | 97.6 | 5.2e+03 | 1.14 | 10.8 | 2.25 | 11.4 | 100.0% | 1.43% | 4.2% |

> **How to read this:** One row per dataset; values are medians over subjects. Compare a column down the rows: SD and |DC| show the amplitude scale and offset (large = raw, unfiltered data); RMS/SD = 1 means no offset; 50 Hz ratio > 1 means mains noise, < 1 a notch filter; sub-1 Hz ratio < 1 means the data were high-passed; the last three columns are the share of problematic trials or channels.

![Figure A1. Trial-level QC metrics per dataset (dot = trial, bar = median; log10 where marked). n trials: KUL-AV-GC 99, DTU 1,080, KUL2016 320, KUL-UHD 120, USTC 360.](figures/a1_qc_metric_distributions.png)

*Figure A1. Trial-level QC metrics per dataset (dot = trial, bar = median; log10 where marked). n trials: KUL-AV-GC 99, DTU 1,080, KUL2016 320, KUL-UHD 120, USTC 360.*

> **How to read this:** Each small panel is one quality metric; its title says which, and 'log10' means the axis is logarithmic (each unit = 10x). Each column of dots is a dataset (one dot per trial); the shaded shape (violin) shows where trials cluster and the black bar is the median. Look for datasets whose cloud sits far from the others: e.g. median_rms_sd near 0 (log10 of 1) for KUL-AV-GC and KUL2016 means no DC offset, while DTU, KUL-UHD and USTC sit at 50-200x; median_abs_dc tells the same story.

![Figure A2. Median-channel Welch PSD (4-s Hann, 50% overlap), median over trials with IQR band.](figures/a2_psd_median_iqr.png)

*Figure A2. Median-channel Welch PSD (4-s Hann, 50% overlap), median over trials with IQR band.*

> **How to read this:** Power spectrum of the raw EEG: x = frequency (Hz), y = power, both on log scales. The line is the median trial of a dataset and the shaded band covers the middle 50% of trials. What to see: the red line (KUL-AV-GC) collapses below 1 Hz and above ~40 Hz (it was filtered before release); the bump near 10 Hz is the alpha rhythm (strongest in KUL2016); the sharp spike at 50 Hz is mains noise (KUL-UHD, USTC); the other datasets keep large power below 1 Hz because they were never high-passed.

*Table 3. Inferred pre-filtering history (from RMS/SD, sub-1 Hz power ratio, aperiodic-fit roll-off and the 50 Hz line ratio; the reference check uses the SD of the across-channel mean relative to the channel SD).*

| ds | inferred_highpass | inferred_lowpass | inferred_notch | inferred_reference |
|---|---|---|---|---|
| KUL-AV-GC | HP ~1 Hz or higher | LP, roll-off ~40.8 Hz | notch at 50 Hz | not average-referenced |
| DTU | no HP (DC offsets present) | LP, roll-off ~64.0 Hz | no notable 50 Hz line | not average-referenced |
| KUL2016 | HP < 1 Hz (or none, DC removed) | LP, roll-off ~64.0 Hz | no notable 50 Hz line | not average-referenced |
| KUL-UHD | no HP (DC offsets present) | LP, roll-off ~62.8 Hz | strong 50 Hz line | not average-referenced |
| USTC | no HP (DC offsets present) | LP, roll-off ~64.0 Hz | no notable 50 Hz line | not average-referenced |

> **How to read this:** The processing each dataset most likely received before release, inferred from the spectra. Read across a row to get one dataset's history; datasets with different rows need harmonised preprocessing before pooling.

Interpretation. The stored units differ by an order of magnitude and three datasets carry DC offsets of thousands of units, so a causal high-pass with state initialisation and a causal running z-score are required before any pooling. KUL-AV-GC has almost no power below 1 Hz (sub-1 Hz ratio 0.0244), which makes slow EOG/drift information unavailable in that dataset only. None of the datasets is stored average-referenced. The 50 Hz line in KUL-UHD and USTC is real but 2.8-4.1x weaker (median over subjects) than the example-trial values suggested.

![Figure A3. Share of trials in which each channel is flagged bad (KUL-UHD shown on its 253-channel layout).](figures/a3_bad_channel_topography.png)

*Figure A3. Share of trials in which each channel is flagged bad (KUL-UHD shown on its 253-channel layout).*

> **How to read this:** Each disc is a head seen from above (nose at the top, left ear on the left); dots are electrodes. The colour shows in what share of trials that electrode was flagged bad (white = never, dark red = often). Each panel has its own colour scale (see its colour bar), so compare patterns rather than colours across panels. KUL-UHD is drawn as dots on its 253-channel layout. Clusters at the edge (e.g. frontal USTC, temporal DTU) point to poor electrode contact.

![Figure A4. Share of trials with an edge transient, split by trial start and end.](figures/a4_edge_transients.png)

*Figure A4. Share of trials with an edge transient, split by trial start and end.*

> **How to read this:** Bars give the share of trials with a large jump (> 20 robust SD) within 0.5 s of the trial start (solid) or end (faded). A value of 1.0 means every trial. KUL-UHD and USTC reach 100%: every one of their trials needs the edges trimmed or the filters warmed up before use.

## 6. Phase B: labels, streams and design confounds

*Table 4. Ground-truth audit from the CND per-sample attendedStream / attendedAzimuth fields (per-trial detail in tables/label_audit.csv). SuperHuge column 0 equals the majority-attended stream in every trial (r = 1.000).*

| dataset | trials | within-trial speaker switches | within-trial direction switches | direction samples mislabelled by trial label | column 0 = CND stream 1 |
|---|---|---|---|---|---|
| KUL-AV-GC | 99 | 0 | 99 | 50% | 100% |
| DTU | 1080 | 0 | 0 | 0% | 49% |
| KUL2016 | 320 | 0 | 0 | 0% | 80% |
| KUL-UHD | 120 | 0 | 0 | 0% | 50% |
| USTC | 360 | 0 | 0 | 0% | 50% |

> **How to read this:** Counts from the per-sample ground-truth labels. 'Within-trial speaker switches' is 0 everywhere; only KUL-AV-GC has direction switches. 'Direction samples mislabelled' is the share of samples where the single trial-level label is wrong (50% in KUL-AV-GC because the side changes half way).

*Table 5. H6 cross-half test on all 99 KUL-AV-GC trials (13 subjects): backward decoder (0-500 ms, ridge) trained on one half and tested on the other; median subject accuracy % [95% CI] under each hypothesis.*

| window s | A: speaker switches at 5 min | B: same speaker all trial |
|---|---|---|
| 1 | 45.4 [44.3, 46.0] | 56.6 [55.8, 57.8] |
| 10 | 36.2 [31.9, 39.0] | 67.9 [65.9, 71.8] |
| 60 | 20.3 [12.5, 28.1] | 84.4 [79.7, 89.1] |

> **How to read this:** Two competing explanations of the KUL-AV-GC 5-min event, tested by decoding. Each cell is the accuracy expected if that explanation were true. The 'same speaker all trial' column rises with window length (84% at 60 s) while the other falls far below 50%, so the speaker did not change.

The KUL-AV-GC 5-min event is a change of attended direction with the same attended talker (the two talkers swap sides). For a stimulus-reconstruction decoder there is no switch at all; for a spatial decoder the correct per-sample label flips at 300 s and the trial-level label is wrong for the second half of every trial.

*Table 6. H7 envelope asymmetry and the 'loud column' shortcut (a logistic classifier seeing only envelope statistics of the two columns in random order).*

| dataset | trials | col0 SD > col1 SD | binomial p | subject Wilcoxon p | distinct SD-ratio levels | loud-column classifier LOSO % |
|---|---|---|---|---|---|---|
| KUL-AV-GC | 99 | 58.6% | 0.11 | 0.048 | 88 | 50.0 [50.0, 66.7] |
| DTU | 1080 | 49.8% | 0.93 | 0.97 | 111 | 45.8 [38.3, 47.5] |
| KUL2016 | 320 | 65.0% | 8.8e-08 | 3.1e-05 | 12 | 82.5 [80.0, 90.0] |
| KUL-UHD | 120 | 50.0% | 1 | 0.00015 | 7 | 50.0 [50.0, 50.0] |
| USTC | 360 | 48.9% | 0.71 | 0.016 | 334 | 50.0 [45.0, 55.1] |

> **How to read this:** 'col0 SD > col1 SD' is the share of trials where the attended envelope is louder (50% = no bias). The p-values test whether that share differs from 50%. The last column shows how well a classifier picks the attended talker from loudness statistics alone - anything clearly above 50% is a shortcut (KUL2016: 82.5%).

![Figure B1. log2 SD ratio of column 0 (attended) to column 1 per trial.](figures/b4_env_sd_ratio.png)

*Figure B1. log2 SD ratio of column 0 (attended) to column 1 per trial.*

> **How to read this:** Each dot is one trial: the loudness (SD) of the attended envelope divided by that of the ignored one, on a log2 scale. 0 = equally loud, +1 = attended twice as loud, -1 = half as loud. The black bar is the median. A cloud shifted above 0 (KUL2016) means the attended talker is systematically louder - a shortcut. Dots lined up in horizontal rows are the same stimulus pair reused across subjects (KUL2016 has only 12 distinct rows).

*Table 7. H8 design confounds. The drift-only classifier uses per-channel DC, drift slope and sub-1 Hz power (no trial index); within-subject evaluation is balanced leave-pair-out because plain leave-one-trial-out is biased below 50% with balanced labels.*

| dataset | label vs trial parity agreement | subjects fully parity-locked | drift-only classifier, within subject | drift-only classifier, LOSO |
|---|---|---|---|---|
| KUL-AV-GC | 55.6% | 2/13 | 56.2% | 50.0% |
| DTU | 51.3% | 0/18 | 54.2% | 49.2% |
| KUL2016 | 50.0% | 16/16 | 44.2% | 60.0% |
| KUL-UHD | 100.0% | 30/30 | 37.5% | 50.0% |
| USTC | 100.0% | 18/18 | 47.9% | 50.0% |

> **How to read this:** 'Label vs trial parity agreement' of 100% means the attended side can be read off whether the trial number is odd or even. The two classifier columns show how well slow drift alone predicts the label (50% = no information).

*Table 8. Stimulus reuse from envelope identity (zero-lag correlation of the 8 Hz envelopes, first 40 s). All 240 KUL-UHD envelopes also occur in KUL2016 (4 shared audio clusters).*

| dataset | envelopes | streamWav names | distinct audio (r > 0.99) | audio clusters with several names | max subjects per audio |
|---|---|---|---|---|---|
| KUL-AV-GC | 198 | 30 | 30 | 29 | 10 |
| DTU | 2160 | 124 | 124 | 0 | 18 |
| KUL2016 | 640 | 16 | 8 | 8 | 16 |
| KUL-UHD | 240 | 4 | 4 | 0 | 30 |
| USTC | 720 | 80 | 80 | 0 | 10 |

> **How to read this:** How many different recordings of speech exist per dataset, found from the envelopes themselves. 'Audio clusters with several names' > 0 means one recording appears under different file names - so file names cannot be trusted to build story-disjoint folds.

Envelope character: no dataset stores per-trial z-scored envelopes; all are signed/mean-removed (about half the samples are negative), and inter-stream correlation is near zero (median r KUL-AV-GC 0.0016, DTU -0.0082, KUL2016 0.01, KUL-UHD 0.032, USTC 0.012). KUL-AV-GC envelopes are sharply high-passed at about 1 Hz; USTC envelopes show a 4-5 Hz (syllable-rate) bump and a platykurtic distribution (kurtosis -1.4), i.e. a different envelope extraction. Match the envelope front end across datasets before pooling.

![Figure B2. Envelope modulation spectrum (column 0, median over trials).](figures/b7_env_modspec.png)

*Figure B2. Envelope modulation spectrum (column 0, median over trials).*

> **How to read this:** How fast the speech loudness fluctuates: x = modulation frequency (Hz), y = relative power, log scales. Curves on top of each other mean the envelopes were extracted the same way. KUL-AV-GC (red) drops sharply below 1 Hz (its envelopes were high-passed); USTC (orange) has a bump at 4-5 Hz, the syllable rate of Mandarin, and a different shape - a different envelope extraction.

*Table 9. H9 stream-order audit (whole-trial correlation of a held-out decoder with each column). Subjects with a column-1 majority at FDR q < 0.05: 0 of 95.*

| dataset | col1 wins, LOSO | col1 wins, LODO | median r col0 / col1 (LOSO) |
|---|---|---|---|
| KUL-AV-GC | 8.1% | 5.1% | 0.058 / 0.017 |
| DTU | 22.9% | 20.3% | 0.074 / 0.018 |
| KUL2016 | 10.9% | 13.4% | 0.081 / 0.031 |
| KUL-UHD | 5.8% | 8.3% | 0.047 / 0.0054 |
| USTC | 25.3% | 32.8% | 0.035 / 0.0057 |

> **How to read this:** For each dataset, how often a decoder that never saw the trial prefers column 1 (the stream marked 'ignored'). Well below 50% means the columns are labelled correctly; a swap would push it above 50%.

## 7. Phase C: neural tracking, decoders and the SNR law

Causal pipeline used for every decoding result unless stated: 1-s warm-up and trim at both edges; causal Butterworth high-pass 0.5 Hz (sosfilt, sosfilt_zi initialised); common-average reference over the montage; causal band-pass 1-9 Hz; causal anti-alias 12 Hz and decimation to 32 Hz; causal running robust z-score (tau 30 s, clip +-5). The same filters are applied to the envelopes. Backward ridge decoders use EEG lags 0-250 or 0-500 ms after the envelope sample; lambda is chosen by inner cross-validation on training groups only.

*Table 10. Envelope-EEG cross-correlation (RMS over 57 channels, subject means of trials) at the attended peak within 0-500 ms. A per-trial circular-shift null (100 shifts, 40 trials per dataset) is in tables/xcorr_null_trials.csv.*

| dataset | subjects | peak lag ms | peak lag 95% CI | RMS r attended | RMS r ignored | att - ign | att - ign 95% CI | Wilcoxon p |
|---|---|---|---|---|---|---|---|---|
| KUL-AV-GC | 13 | 93.75 | [62.5, 156.2] | 0.0152 | 0.00982 | 0.00719 | [0.003924, 0.00969] | 0.01 |
| DTU | 18 | 156.25 | [140.6, 234.4] | 0.0186 | 0.00766 | 0.0114 | [0.007976, 0.01479] | 7.6e-06 |
| KUL2016 | 16 | 140.625 | [93.75, 187.5] | 0.0204 | 0.0107 | 0.00859 | [0.006106, 0.01177] | 3.1e-05 |
| KUL-UHD | 30 | 187.5 | [156.2, 203.1] | 0.0151 | 0.00854 | 0.00649 | [0.004993, 0.008899] | 8e-08 |
| USTC | 18 | 187.5 | [125, 296.9] | 0.0148 | 0.00816 | 0.0076 | [0.003161, 0.008991] | 0.0034 |

> **How to read this:** Size and timing of neural tracking per dataset: the delay of the peak (ms), the correlation strength for attended and ignored speech, their difference with CI, and a p-value testing that attended > ignored across subjects. All p-values are small, so the attention effect is present everywhere.

![Figure C1. Tracking curves (median over subjects, IQR band); solid = attended, dashed = ignored.](figures/c1_tracking_curves.png)

*Figure C1. Tracking curves (median over subjects, IQR band); solid = attended, dashed = ignored.*

> **How to read this:** How strongly the EEG follows the speech envelope at each delay. x = delay in ms (positive = EEG after the sound), y = correlation strength averaged over the 57 channels (RMS). Solid = attended talker, dashed = ignored talker; the band covers the middle 50% of subjects. A peak at 100-200 ms is the brain's response to the speech; the gap between solid and dashed at the peak is the attention effect every decoder relies on. It is visible in all five datasets but small (correlations around 0.01-0.02).

![Figure C2. Topographies at the dataset median peak lag (attended, ignored, difference).](figures/c1_topographies.png)

*Figure C2. Topographies at the dataset median peak lag (attended, ignored, difference).*

> **How to read this:** Head maps (nose up) of the envelope-EEG correlation at each dataset's peak delay. Rows: attended talker, ignored talker, and their difference. Red = positive, blue = negative correlation; within a column the three maps share one colour scale. Where the attended map is stronger than the ignored one (strong colour in the bottom row) is where attention information lives - mostly central/frontal and temporal electrodes.

![Figure C3. Forward TRF global field power (own ridge implementation; MNE is not installed).](figures/c1_forward_trf_gfp.png)

*Figure C3. Forward TRF global field power (own ridge implementation; MNE is not installed).*

> **How to read this:** The forward model (temporal response function): how the EEG responds over time to the speech envelope. x = delay (ms), y = response strength summed over channels (global field power). Solid = attended, dashed = ignored. Peaks between roughly 100 and 250 ms are the auditory response; a higher solid curve means attended speech drives the brain more strongly.

*Table 11. Accuracy (%) of the backward decoder by window length, median over subjects [95% CI] (lags 0-500 ms, 57 channels, non-overlapping windows; DTU trials are 48 s after trimming, so no 60-s windows exist). All values: tables/decoder_accuracy_vs_window.csv.*

| scheme | window s | KUL-AV-GC | DTU | KUL2016 | KUL-UHD | USTC |
|---|---|---|---|---|---|---|
| within-subject LOTO | 1 | 57.3 [56.3, 58.2] | 57.6 [56.7, 59.8] | 58.0 [56.9, 58.9] | 55.9 [54.4, 57.0] | 56.1 [55.0, 58.3] |
| within-subject LOTO | 10 | 70.0 [68.9, 71.0] | 71.9 [67.9, 74.4] | 72.9 [71.3, 75.6] | 65.9 [61.8, 69.7] | 66.8 [63.4, 72.3] |
| within-subject LOTO | 60 | 88.9 [84.7, 91.7] | n/a | 91.0 [86.1, 94.4] | 81.2 [72.9, 87.5] | 77.5 [75.0, 85.0] |
| LOSO | 1 | 53.8 [53.0, 54.9] | 55.3 [53.6, 56.9] | 54.7 [54.3, 56.8] | 54.6 [53.6, 55.2] | 54.1 [52.1, 55.2] |
| LOSO | 10 | 61.9 [60.8, 64.8] | 66.0 [60.4, 68.3] | 66.1 [61.3, 68.9] | 63.5 [60.5, 64.9] | 60.2 [55.2, 64.8] |
| LOSO | 60 | 72.2 [69.4, 81.5] | n/a | 84.0 [73.6, 87.5] | 77.1 [72.9, 83.3] | 67.5 [57.5, 80.0] |
| LODO | 1 | 54.4 [54.0, 55.6] | 55.6 [54.6, 56.8] | 54.5 [52.9, 55.3] | 53.9 [53.0, 54.8] | 52.8 [50.4, 54.1] |
| LODO | 10 | 61.9 [59.3, 65.7] | 66.5 [62.9, 69.2] | 64.1 [57.8, 66.9] | 62.8 [56.9, 64.5] | 56.4 [52.7, 61.1] |
| LODO | 60 | 77.8 [75.0, 79.2] | n/a | 79.2 [66.7, 83.3] | 75.0 [70.8, 79.2] | 62.5 [50.0, 75.0] |

> **How to read this:** Accuracy (%) for the three validation schemes (rows) at 1, 10 and 60-s windows, per dataset (columns). Compare down a column to see the cost of generalising: within subject > new subject > new dataset. Brackets are 95% CIs over subjects.

![Figure C4. Accuracy vs window length per scheme (n = 13/18/16/30/18 subjects).](figures/c4_accuracy_vs_window.png)

*Figure C4. Accuracy vs window length per scheme (n = 13/18/16/30/18 subjects).*

> **How to read this:** Decoding accuracy (y) against the length of the decision window in seconds (x, log scale), for three validation schemes: within subject (left), new subject from the same dataset (middle) and new dataset (right). Points are the median over subjects, bars the 95% CI; the grey line is chance (0.5). All curves rise with window length (more data per decision = fewer errors). The right panel is the realistic one for the challenge: about 53-56% at 1 s and 56-67% at 10 s. DTU has no 60-s point because its trials are only 48 s long after trimming.

*Table 12. 1-s evidence SNR1 = mean(d)/SD(d), d = r(att) - r(ign), per subject (lags 0-500 ms). LODO variants: pool = dataset-equal, subject-equal pooled covariance; raw = sample-weighted pooled covariance; avgdec = average of per-subject decoders. The official baseline's 53.94% at 1 s implies SNR1 = 0.0989.*

| scheme | KUL-AV-GC | DTU | KUL2016 | KUL-UHD | USTC |
|---|---|---|---|---|---|
| loto | 0.185 [0.162, 0.205] | 0.177 [0.166, 0.23] | 0.197 [0.18, 0.231] | 0.142 [0.11, 0.157] | 0.157 [0.126, 0.213] |
| loso | 0.0871 [0.0783, 0.11] | 0.142 [0.0909, 0.166] | 0.133 [0.105, 0.173] | 0.12 [0.0897, 0.13] | 0.102 [0.0474, 0.13] |
| lodo_pool | 0.107 [0.0939, 0.132] | 0.131 [0.119, 0.166] | 0.12 [0.0673, 0.129] | 0.0971 [0.0698, 0.122] | 0.0691 [0.0143, 0.108] |
| lodo_raw | 0.106 [0.0914, 0.127] | 0.134 [0.123, 0.167] | 0.118 [0.068, 0.128] | 0.0984 [0.072, 0.125] | 0.0694 [0.0213, 0.112] |
| lodo_avgdec | 0.0921 [0.0814, 0.12] | 0.13 [0.113, 0.168] | 0.122 [0.0741, 0.133] | 0.111 [0.0903, 0.132] | 0.0644 [0.0526, 0.106] |

> **How to read this:** SNR1 per scheme (rows) and dataset (columns). Higher = stronger evidence per second. The three 'lodo' rows are different ways of pooling training data; they are almost identical. Compare with 0.099, the value implied by the official baseline.

SNR law: fitting accuracy(w) = Phi(k * SNR1 * sqrt(w)) over subjects and windows gives k = 0.74-0.98 (median 0.91). k < 1 means consecutive 1-s windows are positively correlated, so averaging buys a little less than sqrt(w). Within-subject SNR1 is 1.4-2.3x the LODO SNR1 (per dataset), which is the main cost of cross-dataset transfer. The three LODO aggregations (dataset-equal pooled covariance, sample-weighted pooled covariance, averaged per-subject decoders) differ by <= 0.3 pp at 1 s.

![Figure C5. Measured vs predicted accuracy under the SNR law (dot = subject x window length).](figures/c4_snr_law.png)

*Figure C5. Measured vs predicted accuracy under the SNR law (dot = subject x window length).*

> **How to read this:** Test of the SNR law. Each dot is one subject at one window length; x = the accuracy predicted from that subject's 1-s SNR alone, y = the accuracy actually measured. Dots on the dashed diagonal mean the prediction is exact. The legend gives the fitted k per dataset (1 = consecutive seconds independent; below 1 = they are correlated, so longer windows help slightly less than ideal).

*Table 13. Pipeline comparison under LODO, all 95 subjects: median accuracy % (paired median difference vs the reference in pp; Wilcoxon p across subjects).*

| pipeline | 0-250 ms, 1 s | 0-250 ms, 10 s | 0-500 ms, 1 s | 0-500 ms, 10 s |
|---|---|---|---|---|
| causal 1-9 Hz, CAR, 57 ch (reference) | 54.2 | 62.9 | 54.3 | 62.3 |
| zero-phase 1-9 Hz | 53.8 (-0.32; p=0.0022) | 61.8 (-1.33; p=5.7e-05) | 54.1 (-0.37; p=0.011) | 61.8 (-0.28; p=0.24) |
| supplied 32 Hz preprocessed data | 54.3 (-0.61; p=0.12) | 62.1 (-0.91; p=0.2) | 54.1 (-0.34; p=0.53) | 63.2 (+0.22; p=0.27) |
| causal 0.5-30 Hz (64 Hz) | 54.1 (-0.10; p=0.78) | 62.5 (+0.21; p=0.5) | 54.6 (+0.00; p=0.086) | 62.9 (+1.33; p=0.041) |
| no re-reference | 53.8 (-0.45; p=0.002) | 60.8 (-1.36; p=0.00013) | 53.9 (-0.26; p=0.032) | 61.2 (-0.44; p=0.2) |
| 58-ch montage | 54.3 (-0.04; p=0.075) | 62.3 (-0.44; p=0.044) | 54.2 (+0.00; p=0.92) | 61.4 (+0.00; p=0.55) |
| 62-ch montage (USTC zero-filled) | 54.0 (+0.04; p=1) | 62.5 (+0.00; p=0.42) | 54.2 (+0.00; p=0.78) | 61.8 (+0.42; p=0.73) |

> **How to read this:** Each row is one change to the preprocessing, compared with the first row (reference). The number is the median accuracy; in brackets the median change vs the reference in percentage points and its p-value. Negative changes mean the variant is worse. Nothing beats the causal reference with a significant margin except, marginally, the 0.5-30 Hz variant at 10 s.

H5 verdict: causal filtering costs nothing. Zero-phase filtering is not better; where the paired test is significant, it is slightly worse. The supplied 32 Hz preprocessed data are not better either, and are worse for KUL-AV-GC at 1 s (-0.92 pp, p = 0.00024). Dropping CAR lowers accuracy and SNR1. The causal 0.5-30 Hz variant (64 Hz) is the only one ahead of the reference: +1.33 pp at 10 s (p = 0.041, uncorrected for the 7 comparisons) and +0.00 pp at 1 s (p = 0.086); worth a follow-up, not yet a conclusion.

## 8. Phase D: domain shift and spatial shortcuts

*Table 14. H10 fingerprints: 5-s epochs (n = 30,936 from 1,979 trials), 1-9 Hz and 8-13 Hz covariances (57 -> 56-dim zero-sum projection, log-Euclidean tangent space), PCA(100) + logistic regression, 5-fold CV grouped by trial. Dataset-ID chance: 31.4% (majority), 20% (uniform).*

| features / alignment | dataset ID acc | balanced acc | subject ID KUL-AV-GC | subject ID DTU | subject ID KUL2016 | subject ID KUL-UHD | subject ID USTC |
|---|---|---|---|---|---|---|---|
| none (raw scale) | 99.7% | 99.4% | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% |
| none, trace-normalised | 99.6% | 99.3% | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% |
| EA per recording | 84.3% | 79.1% | 72.5% | 89.0% | 91.8% | 97.6% | 100.0% |
| causal running EA | 87.1% | 82.7% | 89.8% | 96.0% | 92.6% | 97.3% | 99.7% |
| amplitude raw | 98.9% | 98.0% | - | - | - | - | - |
| amplitude after causal running z | 86.0% | 75.5% | - | - | - | - | - |

> **How to read this:** How well a classifier can tell which dataset (or which subject) a 5-s EEG segment came from. Rows: without and with alignment. Chance for dataset ID is 31% (always guessing the largest dataset). Values near 100% mean a strong fingerprint; alignment lowers it only to 84-87%.

![Figure D1. Tangent-space PCA of 5-s covariances before and after alignment (8,000 random epochs).](figures/d1_embedding.png)

*Figure D1. Tangent-space PCA of 5-s covariances before and after alignment (8,000 random epochs).*

> **How to read this:** Each dot is one 5-s EEG segment, placed by the two main directions of variation of its spatial covariance (axes have no units). Colour = dataset. Four panels: no alignment, scale-normalised, Euclidean alignment, causal running alignment; the title gives how accurately a classifier can tell the dataset. Separate coloured islands mean each dataset has its own 'fingerprint'; the more the colours mix after alignment, the more the alignment removed it. They mix only partly (84-87% still identifiable).

*Table 15. H11 attended-side decoding (filter bank 1-4/4-8/8-13/13-30 Hz; FBCSP-LDA and Riemannian tangent-space logistic regression), median subject accuracy. *KUL-AV-GC: leave-one-trial-out (each trial holds both sides); other datasets: balanced leave-pair-out. Trial-ID chance = 1/trials per subject (12.5%, 1.7%, 5%, 25%, 5%). Windows capped at 24 (5 s) / 15 (1 s) per trial.*

| window | model | dataset | within-trial split | LOTO* | LOSO | LODO | trial-ID |
|---|---|---|---|---|---|---|---|
| 5 s | fbcsp | KUL-AV-GC | 57.3% | 44.8% | 50.0% | 53.1% | - |
| 5 s | fbcsp | DTU | 73.9% | 49.4% | 50.1% | 51.3% | - |
| 5 s | fbcsp | KUL2016 | 80.7% | 60.3% | 64.6% | 51.7% | - |
| 5 s | fbcsp | KUL-UHD | 92.7% | 44.3% | 50.5% | 54.2% | - |
| 5 s | fbcsp | USTC | 67.6% | 29.2% | 50.9% | 49.0% | - |
| 5 s | riemann | KUL-AV-GC | 46.9% | 43.2% | 51.0% | 50.7% | 70.8% |
| 5 s | riemann | DTU | 66.1% | 51.9% | 51.9% | 51.5% | 72.4% |
| 5 s | riemann | KUL2016 | 78.2% | 58.7% | 66.0% | 50.8% | 62.9% |
| 5 s | riemann | KUL-UHD | 92.2% | 43.2% | 51.0% | 51.0% | 89.1% |
| 5 s | riemann | USTC | 73.2% | 37.2% | 50.1% | 50.0% | 81.3% |
| 1 s | fbcsp | KUL-AV-GC | 55.0% | 51.7% | 50.0% | 49.2% | - |
| 1 s | fbcsp | DTU | 67.7% | 50.8% | 48.8% | 51.3% | - |
| 1 s | fbcsp | KUL2016 | 66.8% | 54.7% | 61.3% | 53.0% | - |
| 1 s | fbcsp | KUL-UHD | 73.3% | 44.2% | 50.0% | 52.5% | - |
| 1 s | fbcsp | USTC | 59.5% | 33.0% | 51.2% | 49.5% | - |
| 1 s | riemann | KUL-AV-GC | 50.0% | 47.5% | 52.5% | 47.5% | 37.5% |
| 1 s | riemann | DTU | 61.1% | 51.7% | 51.0% | 50.2% | 40.5% |
| 1 s | riemann | KUL2016 | 62.0% | 55.2% | 56.5% | 52.3% | 28.0% |
| 1 s | riemann | KUL-UHD | 65.8% | 41.2% | 50.0% | 51.7% | 56.7% |
| 1 s | riemann | USTC | 55.3% | 38.0% | 49.8% | 50.2% | 39.2% |

> **How to read this:** Direction (left/right) decoding with spatial filters. Read along a row from left to right: the leaky within-trial split is highest, then accuracy falls as training and test data get further apart (held-out trial, new subject, new dataset). The last column shows the same features identifying the trial itself - far above chance - which is what inflates the first column.

Spatial decoders look good only when windows of the same trial appear in training and test data: the median within-trial-split minus held-out-trial gap is 22.4 pp at 5 s (up to 49 pp, KUL-UHD FBCSP 92.7% -> 44.3%). The same features identify the trial. In USTC and KUL-UHD held-out-trial accuracy is significantly BELOW 50%: labels alternate by trial parity, so the trials nearest in time (most similar drift state) carry the opposite label. Across datasets (LODO) nothing transfers.

*Table 16. H12 physiology vs attended side: per-subject Cohen's d (right minus left, 10-s bins), across-subject Wilcoxon, FDR over all tests. ALI = (R - L)/(R + L) parieto-occipital 8-13 Hz power; heog = mean(AF7,F7,FT7) - mean(AF8,F8,FT8) below 1 Hz; eog0 = first recorded external channel (USTC: HEO).*

| ds | feature | n_subjects | d [95% CI] | share_positive | wilcoxon_p | fdr_q |
|---|---|---|---|---|---|---|
| KUL-AV-GC | ali | 13 | -0.053 [-0.15, 0.0043] | 30.8% | 0.068 | 0.32 |
| KUL-AV-GC | heog | 13 | 0.022 [-0.016, 0.042] | 61.5% | 0.34 | 0.57 |
| KUL-AV-GC | eog0 | 13 | 0.012 [-0.0083, 0.042] | 61.5% | 0.41 | 0.6 |
| DTU | ali | 18 | 0.18 [-0.017, 0.33] | 72.2% | 0.074 | 0.32 |
| DTU | heog | 18 | -0.085 [-0.27, 0.069] | 38.9% | 0.12 | 0.43 |
| DTU | eog0 | 18 | -0.0068 [-0.089, 0.094] | 44.4% | 0.8 | 0.83 |
| KUL2016 | ali | 16 | -0.18 [-0.37, 0.12] | 37.5% | 0.14 | 0.43 |
| KUL2016 | heog | 16 | -0.021 [-0.028, -0.00071] | 25.0% | 0.0063 | 0.082 |
| KUL-UHD | ali | 30 | -0.056 [-0.34, 0.29] | 46.7% | 0.73 | 0.83 |
| KUL-UHD | heog | 30 | -0.11 [-0.66, 0.13] | 46.7% | 0.16 | 0.43 |
| USTC | ali | 18 | -0.057 [-0.31, 0.038] | 33.3% | 0.074 | 0.32 |
| USTC | heog | 18 | -0.17 [-0.28, 0.052] | 38.9% | 0.021 | 0.18 |
| USTC | eog0 | 18 | 0.43 [0.27, 1] | 83.3% | 0.00053 | 0.014 |

> **How to read this:** Does brain or eye activity differ between attending right and left? d is the effect size per subject (median, 95% CI); 'share positive' is the share of subjects with d > 0; q is the p-value corrected for all tests in the table. Only USTC's eye channel (eog0) shows a clear effect.

*Table 17. KUL-AV-GC switch-locked averages (+-60 s around each of the 99 genuine direction switches, sign-aligned).*

| signal | subjects | switches | post - pre (median [95% CI]) | Wilcoxon p |
|---|---|---|---|---|
| ALI (1-s) | 13 | 99 | -0.017 [-0.0245, 0.00628] | 0.19 |
| hEOG proxy (4 Hz) | 13 | 99 | 0.000437 [-0.000633, 0.0032] | 0.24 |
| EOG ch0 | 13 | 99 | -0.00275 [-0.004, 0.00164] | 0.22 |
| EOG ch1 | 13 | 99 | -0.000211 [-0.00453, 0.00367] | 0.79 |
| EOG ch2 | 13 | 99 | 0.000781 [-0.00247, 0.00443] | 0.38 |
| EOG ch3 | 13 | 99 | -0.00323 [-0.00871, 0.00189] | 0.08 |

> **How to read this:** The change from 60 s before to 60 s after each real KUL-AV-GC switch (median over subjects, 95% CI) and its p-value. Values near 0 with large p mean nothing changes at the switch.

![Figure D2. KUL-AV-GC switch-locked ALI, hEOG proxy and EOG channels (median over subjects, 95% bootstrap band).](figures/d4_switch_locked.png)

*Figure D2. KUL-AV-GC switch-locked ALI, hEOG proxy and EOG channels (median over subjects, 95% bootstrap band).*

> **How to read this:** Signals averaged around the 99 real KUL-AV-GC side switches: x = time from the switch (s; 0 = switch), y = signal level, flipped so that a left-to-right switch would push the curve upward. Left: alpha lateralisation; middle: eye-movement proxy from frontal electrodes; right: the recorded EOG channels. A physiological marker of the switch would appear as a step at 0. The curves stay flat (the band is the 95% CI), so no marker was found.

There is no alpha-lateralisation or EOG step at the KUL-AV-GC switch. The KUL-AV-GC recordings are high-passed near 1 Hz, so a sustained gaze-related EOG level cannot survive in them; the proxy is uninformative by construction there and in the mean-removed KUL2016 data. The one robust side marker is USTC's recorded HEO channel. Condition-wise ALI effects in KUL-AV-GC (FixedVideo d = -0.16, MovingVideo d = 0.092) have opposite signs and are not FDR-corrected across the 8 condition tests; treat them as exploratory.

## 9. Phase E: switch detection and the accuracy-latency trade-off

Scorer: MATLAB/Octave are not installed, so computeSwitchLatency.m and scorePredictions.m were ported line by line (scripts/scorer.py). Equivalence: 34/34 hand-derived cases pass, including 3,000 random sequences with zero mismatches against an independent brute-force implementation of the documented split search, plus 300 random subjects for the compiled version. No MATLAB-scored outputs are shipped (scoring_result.zip is empty), so no direct comparison with MATLAB was possible.

- Two switches within 100 s: the search window of the first switch contains the second. If the label reverts, the 'old label until the end of the window' split can outscore the true split; it then fails the 0.5-s sustain check, the search resumes past it, and the first switch is scored undetected (SDT = Inf). A perfect decoder is detected only if the next reverting switch is >= 50 s later (cases O in tables/scorer_equivalence_tests.csv). Under the default synthetic schedule (gaps U(30, 90) s) this hits 414/3113 switches for an oracle decoder; its L still reads 1.00 because L is the median over switches.
- The sustain check is not clipped to the trial: a switch in the last 0.5 s of a trial can be confirmed by the next trial's labels.

*Table 18. Lookahead budget of the evaluated configurations (decoder lags 0-500 ms). With 0-250 ms lags the stamping advance is 750 ms. Causal filter group delay is lag, not lookahead.*

| component | lookahead ms |
|---|---|
| causal filters (HP, BP, anti-alias, running z) | 0.0 |
| backward decoder lags 0-500 ms (EEG after the envelope sample) | 500.0 |
| 1-s correlation window (stamped at its end) | 0.0 |
| decision stamping advance (labels samples from t_end + L - 1 s) | 500.0 |
| integrator (causal forward filter / moving average / ladder) | 0.0 |
| TOTAL | 1000.0 |

> **How to read this:** How the allowed 1-s lookahead is spent: the decoder looks 500 ms into the future of each sample, and decisions are used 500 ms earlier than the end of their window. The total must not exceed 1,000 ms.

Lookahead unit test (tables/lookahead_test.csv): large perturbations of all inputs (EEG and both envelopes) after t + 1 s leave every label up to t unchanged, for 3 integrators x 7 values of t (21/21 pass). Two negative controls are flagged as violations: a 2-s budget, and zero-phase filtering. A first 1.25-s control was not flagged (the running z-score clips small leaks), so the test has limited sensitivity to leaks shorter than about 0.5 s.

*Table 19. SYNTHETIC switches (generateSyntheticSwitches.m schedule; label and evidence columns swapped together), LODO evidence (1-s windows, 0.25-s hop, lags 0-500 ms), best integrator on the median held-out dataset = ma_10s. Null = sign-randomised evidence with the same integrator.*

| held-out dataset | SYNTHETIC switches | Acc % | SDT s | L | H | miss rate | null Acc % | null L | null H | oracle misses |
|---|---|---|---|---|---|---|---|---|---|---|
| KUL-AV-GC | 890 | 61.2 | 19.7 | 0.721 | 0.662 | 12.8% | 50.6 | 0.48 | 0.493 | 142/890 |
| DTU | 220 | 64.1 | 6.39 | 0.904 | 0.75 | 13.6% | 50.5 | 0.73 | 0.597 | 0/220 |
| KUL2016 | 934 | 61.3 | 15.5 | 0.73 | 0.666 | 13.5% | 49.8 | 0.549 | 0.522 | 138/934 |
| KUL-UHD | 661 | 59.9 | 19.6 | 0.696 | 0.644 | 11.5% | 50.7 | 0.528 | 0.517 | 113/661 |
| USTC | 408 | 55.3 | 19.5 | 0.746 | 0.635 | 12.3% | 50.1 | 0.649 | 0.566 | 21/408 |

> **How to read this:** Switch-detection scores per held-out dataset for the best integrator. Acc, SDT (seconds to detect), L and H are the official metrics; 'miss rate' = share of switches never detected. The 'null' columns are the same pipeline with random evidence, i.e. the floor. 'Oracle misses' counts switches that even a perfect decoder loses because of the scorer's 50-s property.

*Table 20. Best H per integrator family and held-out dataset (SYNTHETIC; 0-500 ms lags). Families: argmax; moving average 1-40 s; 2-state HMM (p_stay grid, emissions fitted on the training datasets' out-of-fold evidence); HMM on running-SD-normalised evidence with temperature; N-state ladder (MESD-style, N = 3-9).*

| integrator family (best member) | KUL-AV-GC | DTU | KUL2016 | KUL-UHD | USTC |
|---|---|---|---|---|---|
| argmax | 0.602 | 0.694 | 0.623 | 0.599 | 0.622 |
| hmm | 0.655 | 0.76 | 0.672 | 0.649 | 0.639 |
| hmmz | 0.659 | 0.762 | 0.673 | 0.65 | 0.641 |
| ladder | 0.619 | 0.719 | 0.639 | 0.622 | 0.629 |
| ma | 0.662 | 0.75 | 0.666 | 0.647 | 0.635 |

> **How to read this:** The best H each integrator family reaches on each held-out dataset. Compare down a column: smoothing integrators (moving average, HMM) beat plain argmax by 0.02-0.07 H.

![Figure E1. Acc-L trade-off per held-out dataset (SYNTHETIC switches; colour = real evidence, grey = sign-randomised null; lines = iso-H).](figures/e1_tradeoff_c19-common57-car_lodo_pool_0-500ms.png)

*Figure E1. Acc-L trade-off per held-out dataset (SYNTHETIC switches; colour = real evidence, grey = sign-randomised null; lines = iso-H).*

> **How to read this:** The accuracy-versus-reaction trade-off, one panel per held-out dataset. x = Acc (% of samples correct), y = L (switch-detection score, 1 = instant). Each coloured marker is one integrator setting (marker shape = family: circle argmax, square moving average, triangles HMM, diamond ladder); grey markers are the same settings fed with random evidence (the null). The curved grey lines are iso-H lines: every point on a line has the same headline score H, labelled on the line. Better = up and to the right. The coloured cloud sits clearly above the grey null, but far below the dotted oracle line - there is room to improve.

With 0-250 ms lags (750 ms stamping advance) the same integrator gives H = KUL-AV-GC 0.663, DTU 0.748, KUL2016 0.677, KUL-UHD 0.63, USTC 0.636, essentially the same trade-off.

Genuine switches. The 99 KUL-AV-GC direction switches are invisible to a speaker decoder. Scored with direction evidence from spatial decoders trained on the other four datasets, the best configuration (riemann, ma_2s) reaches Acc = 51.7%, SDT = 47.4 s, L = 0.604 and H = 0.557, against a null L of 0.451. Sample accuracy is at chance and L is only modestly above the null, so direction is essentially not decodable across datasets with these features. Mapping the speaker decoder to a direction through the known stream azimuths gives an apparently high H (0.908, hmmz_p0.999_T0.25), but that number is INVALID. Column 0 is the attended speaker for the whole KUL-AV-GC trial by construction, so any sticky integrator that locks onto column 0 turns into a perfect direction track. It also assumes stream positions are known at test time.

## 10. Hypothesis verdicts (H1-H12)

What the hypotheses are. An earlier exploratory analysis looked at a single example trial from each dataset and proposed twelve claims (H1-H12) about the data. A single trial cannot show whether an effect is general, so each claim is stated in full below with the example-trial value that motivated it, why it matters for the challenge, and how it was tested here on the full corpus.

*Table 21. The twelve hypotheses: what each claims, why it matters and how it was tested.*

| ID | short name | the hypothesis in full (from the example-trial EDA) | why it matters | how it was tested on the full corpus |
|---|---|---|---|---|
| H1 | Amplitude scale and DC offsets | The five datasets store EEG on very different amplitude scales (median channel SD about 16x apart), and DTU, KUL-UHD and USTC carry large constant (DC) offsets because they were never high-pass filtered. | Pooling data with different scales and offsets lets a model tell datasets apart by amplitude alone and breaks fixed normalisation constants (like the official HMM's). | Per channel and per trial: mean, SD, robust SD, RMS/SD and drift for all 149,336 channel-trials; subject medians with bootstrap CIs. |
| H2 | Pre-filtering history | The datasets were filtered differently before release: KUL-AV-GC low-passed near 40 Hz and high-passed below 1 Hz; KUL-UHD and USTC contain strong 50 Hz mains noise. | Different band limits change what frequency content a model can rely on; a model trained on one filtering may not transfer. | Welch spectra of every trial: 1/f fit, spectral roll-off, power below 1 Hz, 50 Hz line ratio; inference table per dataset. |
| H3 | Edge transients and bad channels | Trials in USTC and KUL-UHD begin and end with large electrical jumps (thousands of units at t = 0), and some channels are bad (e.g. KUL-UHD P7 with SD 2,381 vs 71 at Cz). | Filters ring on such jumps and one bad channel can dominate a common-average reference, corrupting the first seconds of every trial. | Largest first-difference jump near each edge (> 20 robust SD), bad/flat channel flags against dataset medians, share per dataset. |
| H4 | The SNR law | Decoding accuracy for a window of w seconds is predicted by the 1-s evidence signal-to-noise ratio: accuracy = Phi(SNR1 * sqrt(w)). Within subject, KUL-AV-GC SNR1 is about 0.16; the official baseline (53.94% at 1 s) implies about 0.10. | If true, one number (SNR1) predicts accuracy at every window length and tells how much a better decoder is worth. | SNR1 = mean/SD of r(attended) - r(ignored) over 1-s windows per subject; fit of k in Phi(k * SNR1 * sqrt(w)) for w = 1-60 s under within-subject, cross-subject and cross-dataset validation. |
| H5 | Causal vs zero-phase filtering | Real-time (causal) filtering, which the challenge requires, performs no worse than offline zero-phase filtering. | The 1-s lookahead rule forbids zero-phase filters; if they were clearly better, the rule would cost accuracy. | The same cross-dataset decoder run on causal and zero-phase versions of the data (and on the supplied 32 Hz data), compared per subject with a paired Wilcoxon test. |
| H6 | What the KUL-AV-GC 5-minute switch is | In KUL-AV-GC the change at 5 minutes switches the attended direction (the talker moves to the other side), not the attended talker. | It decides whether the only genuine within-trial switches in the training data can be used to train a speaker-switch detector. | Per-sample CND labels (attendedStream, attendedAzimuth) for all 99 trials, plus a cross-half decoding test: train on one half, test on the other under 'switch' and 'no switch' assumptions. |
| H7 | The attended stream is louder | In KUL2016 and KUL-AV-GC the attended audio envelope has a larger amplitude than the ignored one (65% and 59% of trials). | A model could pick the louder stream without using the EEG at all, which would not work on test data. | SD ratio of the two envelopes in every trial, binomial and Wilcoxon tests, and a classifier that sees only envelope statistics. |
| H8 | The label follows trial order | In KUL-UHD and USTC the attended side (left/right) alternates strictly with trial number: odd trials one side, even trials the other. | Anything correlated with trial order (drift, fatigue, electrode settling) then predicts the label, a shortcut that fails on unseen subjects. | Cross-tabulation of label vs trial parity and order, and a 'drift-only' classifier using slow features (DC, drift, sub-1 Hz power). |
| H9 | USTC stream columns may be swapped | In one USTC example, a decoder that never saw the trial correlated more with the stream marked 'ignored' (r = 0.083 vs 0.026), suggesting the attended/ignored columns are swapped. | Swapped labels would make USTC poison both training and evaluation. | Decoders trained without the subject (LOSO) or without the dataset (LODO); whole-trial correlation with each column in all 360 USTC trials; per-subject binomial tests with FDR correction. |
| H10 | Covariance fingerprints the dataset | The spatial covariance of the EEG identifies which dataset a recording came from almost perfectly (99.7%), and Euclidean alignment (EA) removes only part of this (to 76%). | A strong dataset fingerprint means models can learn recording setup instead of attention, and cross-dataset transfer suffers. | Classifiers predicting dataset and subject from 30,936 5-s covariance epochs, with no alignment, EA, and causal running EA, cross-validated so no trial is split. |
| H11 | Spatial decoders exploit trial fingerprints | Spatial (CSP / Riemannian) decoders of the attended side look good only when windows of the same trial are in training and test data; they collapse when whole trials are held out (literature: trial ID ~100%, held-out trials 41-45% on KUL-AV-GC). | Direction decoding is what the eye-gaze-controlled test sets require; if it only memorises trials it will fail there. | FB-CSP and Riemannian decoders under random within-trial splits, held-out trials, held-out subjects and held-out datasets, plus predicting the trial ID from the same features. |
| H12 | No physiological marker at the switch | At the KUL-AV-GC side switch there is no change in alpha-band lateralisation or horizontal eye movements (example: ALI 0.040 -> 0.045, p = 0.77). | Alpha lateralisation and gaze are the classic markers of spatial attention; if absent, direction must be inferred another way. | Parieto-occipital alpha lateralisation index and EOG signals vs the attended side in all datasets, and time-locked to all 99 genuine switches (+-60 s). |

> **How to read this:** One row per hypothesis: what it claims, why it matters for the challenge and how it was tested here.

How to read the verdicts:

- **confirmed**: The full-corpus data support the claim: the effect is present, in the stated direction, across subjects (95% CI excluding 'no effect' where a test applies), not just in the example trial.
- **partly**: Part of the claim holds and part does not (for example it holds for one dataset but not another, or the effect is real but much smaller than the example trial suggested).
- **refuted**: The full-corpus data contradict the claim: the effect is absent or points the other way.

Result: 10 of 12 hypotheses confirmed, 1 partly confirmed, 1 refuted. H7 is counted as confirmed because it holds fully for KUL2016, but for KUL-AV-GC it only partly holds.

*Table 22. Verdicts in plain language: what was claimed, the verdict and what the full corpus shows.*

| ID | claim tested | verdict | what the full corpus shows |
|---|---|---|---|
| H1 | Datasets are stored on very different amplitude scales, and three carry large DC offsets. | confirmed | Yes. Typical channel SD differs 16.5x between datasets, and DTU, KUL-UHD and USTC sit on offsets of thousands of units. Every model needs per-channel normalisation (a causal running z-score) before datasets can be pooled. |
| H2 | Datasets were filtered differently before release. | partly | Partly. KUL-AV-GC is clearly band-limited (about 1-41 Hz) and notched at 50 Hz. KUL-UHD and USTC do have a 50 Hz line, but it is 2.8-4.1x weaker than the example trial suggested. The other datasets are not high-passed. |
| H3 | Trials start and end with electrical jumps in USTC and KUL-UHD, and some channels are bad. | confirmed | Yes. Every USTC and KUL-UHD trial has a large jump near an edge (100%), against 3-5% elsewhere. 5.1% of KUL-UHD channels are bad. Edges must be trimmed or the filters warmed up. |
| H4 | Decoding accuracy grows with window length in a predictable way, set by the 1-s signal-to-noise ratio. | confirmed | Yes. Accuracy follows Phi(k * SNR1 * sqrt(w)) with k = 0.74-0.98 in every dataset and scheme. For KUL-AV-GC, SNR1 is 0.185 within subject and 0.107 across datasets, close to the 0.099 implied by the official baseline. |
| H5 | Real-time (causal) filtering loses nothing compared with offline zero-phase filtering. | confirmed | Yes. Across all 95 subjects, causal filtering is never worse; zero-phase filtering is slightly worse where the difference is significant. The challenge's causality rule therefore costs nothing at the filtering stage. |
| H6 | The KUL-AV-GC 5-minute event changes where the listener attends, not whom they listen to. | confirmed | Yes. The CND labels show no speaker switch and 99 side switches, and a decoder trained on one half of the trials still tracks the same talker in the other half (84.4% at 60 s vs 20.3% if the talker had changed). |
| H7 | The attended audio stream is systematically louder in KUL2016 and KUL-AV-GC. | confirmed (KUL2016); partly (KUL-AV-GC) | Confirmed for KUL2016 (65% of trials; an envelope-only classifier gets 82.5%). Only partly for KUL-AV-GC (59%, not significant per trial, and useless to a classifier). This is a shortcut a model could exploit. |
| H8 | In KUL-UHD and USTC the attended side is fixed by whether the trial number is odd or even. | confirmed | Yes, in 100% of trials and every subject. Slow-drift features alone predict the label only weakly (at most 60.0%), but any model that sees trial order can learn the label without using the EEG. |
| H9 | The USTC audio columns may be swapped (the ignored stream stored as the attended one). | refuted | No. A held-out decoder prefers column 0 in 74.7% of USTC trials, and no subject shows the reverse. USTC is just harder to decode. |
| H10 | EEG covariance reveals which dataset a recording came from, and alignment removes only part of that. | confirmed | Yes. Dataset identity is predicted at 99.7% from covariance, still 84.3% after Euclidean alignment and 87.1% after the causal version (chance 31.4%). Recording setup is a strong confound for pooled training. |
| H11 | Spatial (direction) decoders look good only because they memorise trials. | confirmed | Yes. The same features identify the trial at up to 89%, accuracy drops by a median 22.4 pp when whole trials are held out, and it is at chance across datasets. For KUL-AV-GC: 57.3% within trial, 44.8% on held-out trials, 53.1% across datasets. |
| H12 | Brain or eye signals of spatial attention do not change at the KUL-AV-GC side switch. | confirmed | Confirmed: neither alpha lateralisation nor eye movements change at the 99 switches (all p >= 0.08). But the KUL-AV-GC data are high-passed near 1 Hz, which removes slow eye-movement signals, so this is weak evidence of absence. USTC's recorded eye channel does separate left from right trials. |

> **How to read this:** One row per hypothesis: the claim, the verdict (colour-coded in the key above) and a plain-language summary of the evidence.

*Table 23. Verdicts with the numbers behind them (also tables/hypotheses_verdicts.csv, which lists the evidence files).*

| ID | previous estimate (n = 1 trial/dataset) | full-corpus estimate (with CIs) | n | verdict |
|---|---|---|---|---|
| H1 | SD 8.3/39.2/9.6/132.1/61.5; RMS/SD 1/210/1/58/94 | median channel SD 7.96/39.1/9.64/131/63.6 (max/min 16.5x); RMS/SD 1/211/1/51.4/97.6; subject-level 95% CIs in qc_dataset_summary.csv | 1979 trials / 95 subjects (149,336 channel-trials) | confirmed |
| H2 | line ratio 10.8 (UHD), 9.3 (USTC) | KUL-AV-GC roll-off 40.8 Hz in 100% of trials, sub-1 Hz ratio 0.0244 (HP ~1 Hz), 50 Hz notched (line ratio 0.0695); 50 Hz line ratio UHD 3.8 [2.624, 8.669], USTC 2.25 [1.533, 3.306], DTU 1.44, KUL2016 1.06; DTU/UHD/USTC not high-passed (DC offsets); KUL2016 mean-removed | 1979 trials / 95 subjects | partly |
| H3 | jumps 9,249 / 13,011 units at t=0; UHD P7 SD 2,381 vs Cz 71 | trials with a >20 robust-SD jump within 0.5 s of an edge: KUL-AV-GC 3.03%, DTU 2.78%, KUL2016 4.69%, KUL-UHD 100%, USTC 100%; bad-channel share KUL-AV-GC 0.126%, DTU 2.33%, KUL2016 0%, KUL-UHD 5.14%, USTC 1.43%; bad trials (>=10% bad/flat ch.) UHD 16.7% | 1979 trials | confirmed |
| H4 | 54.5% @1 s, 66.1% @10 s, 88.9% @60 s (n=9 windows) | LOTO SNR1 0.185 [0.162, 0.205]; acc 57.3% @1 s, 70% @10 s, 88.9% @60 s; fitted k 0.886 (all datasets/schemes 0.744-0.977); LODO SNR1 0.107 [0.0939, 0.132] vs baseline-implied 0.0989 | 13 subjects / 99 trials (KUL-AV-GC); 95 subjects overall | confirmed |
| H5 | within-trial 50-58% @1 s either way | LODO, paired per subject: 0-250ms w=1s: delta acc (zero-phase - causal) -0.323 pp [-0.513, -0.128], Wilcoxon p=0.00223; 0-250ms w=10s: delta acc (zero-phase - causal) -1.33 pp [-1.97, -0.658], Wilcoxon p=5.66e-05; 0-500ms w=1s: delta acc (zero-phase - causal) -0.372 pp [-0.47, -0.0709], Wilcoxon p=0.0115; 0-500ms w=10s: delta acc (zero-phase - causal) -0.282 pp [-0.834, 0.282], Wilcoxon p=0.24 | 95 subjects (LODO) | confirmed: causal is never worse; zero-phase is slightly worse where significant |
| H6 | cross-half: switch hyp. 0.0% @60 s; no-switch hyp. 88.9% | CND: 0 speaker switches, 99 direction switches (all at t=300 s) in 99 trials; cross-half @60 s: no-switch 84.4% [79.7, 89.1] vs switch 20.3% [12.5, 28.1]; trial-level label wrong for 50% of direction samples | 13 subjects / 99 trials | confirmed |
| H7 | 65% of 320 (p=2.3e-9); 59% of 99 (p=0.0095) | KUL2016 65% of 320 (binomial p=8.82e-08; 12 distinct ratio levels = 12 stimulus pairs); KUL-AV-GC 58.6% of 99 (binomial p=0.107, Wilcoxon log-ratio p=0.0096); others ~50%. Envelope-only 'loud column' classifier LOSO: KUL2016 82.5% [80, 90], KUL-AV-GC 50%, DTU 45.8%, UHD 50%, USTC 50% | 1979 trials / 95 subjects | confirmed (KUL2016); partly (KUL-AV-GC) |
| H8 | 60/60 and 180/180 | KUL-UHD agreement 100% of 120 trials (30/30 subjects fully locked); USTC 100% of 360 (18/18); others KUL-AV-GC 55.6%, DTU 51.3%, KUL2016 50%. Drift-only classifier: LOSO KUL2016 60% (p=0.0102), within-subject DTU 54.2% (p=0.00769); others n.s. | 1979 trials | confirmed |
| H9 | r = 0.083 vs 0.026 | USTC column 1 wins 25.3% of 360 trials (LOSO) and 32.8% (LODO); median r col0 0.0355 vs col1 0.0057; subjects with a column-1 majority at FDR q<0.05: 0 | 18 USTC subjects / 360 trials; all 95 subjects | refuted |
| H10 | 99.7% -> 76.1% (chance 39%) | dataset-ID accuracy: none 99.7%, trace-normalised 99.6%, EA 84.3%, causal running EA 87.1% (majority chance 31.4%); subject-ID (within dataset) after EA 72.5-100%, after running EA 89.8-99.7%; amplitude-only features 98.9% raw, 86% after running z | 30,936 5-s epochs / 1979 trials | confirmed |
| H11 | not tested (literature: CSP trial-ID ~100%, LOTO 41-45% on KUL-AV-GC) | 5-s windows: within-trial minus LOTO gap median 22.4 pp (range 3.65 to 49; e.g. KUL-UHD FBCSP 92.7% -> 44.3%); LODO 49-54.2%; trial-ID accuracy KUL-AV-GC 70.8% (chance 12.5%), DTU 72.4% (chance 1.67%), KUL2016 62.9% (chance 5%), KUL-UHD 89.1% (chance 25%), USTC 81.3% (chance 5%); KUL-AV-GC FBCSP within-trial 57.3%, LOTO 44.8%, LODO 53.1% // 1-s windows: within-trial minus LOTO gap median 14.5 pp (range 2.5 to 29.2; e.g. KUL-UHD FBCSP 73.3% -> 44.2%); LODO 47.5-53%; trial-ID accuracy KUL-AV-GC 37.5% (chance 12.5%), DTU 40.5% (chance 1.67%), KUL2016 28% (chance 5%), KUL-UHD 56.7% (chance 25%), USTC 39.2% (chance 5%); KUL-AV-GC FBCSP within-trial 55%, LOTO 51.7%, LODO 49.2%. LOTO is BELOW chance in USTC/KUL-UHD: labels alternate by trial parity, so temporally neighbouring trials (similar drift state) carry the opposite label | 95 subjects; 30,936 5-s and 29,685 1-s windows (<= 24 / 15 per trial) | confirmed |
| H12 | ALI 0.040 -> 0.045 (p=0.77); hEOG p=0.98 | switch-locked post-minus-pre (99 switches, 13 subjects): ALI p=0.191, hEOG proxy p=0.244, EOG channels p>=0.0803. Elsewhere: USTC HEO separates left/right trials (d=0.432, q=0.0139) | 99 switches / 13 subjects | confirmed |

> **How to read this:** The same verdicts with the full numbers: the previous (single-trial) estimate next to the full-corpus estimate, and the sample size n behind it.

## 11. What this means for modelling

- Front end: use the causal chain above (state-initialised high-pass, CAR over the 57 common channels matched by name, causal band-pass, running robust z-score with clipping). It is never worse than zero-phase filtering or the supplied 32 Hz data (Table 13), it is causal, and it removes the DC offsets and scale differences of Table 2. Add explicit edge handling, because KUL-UHD and USTC have transients in every trial.
- Labels: train speaker decoders on the per-sample CND attendedStream. For direction targets use the per-sample azimuth, never the trial label (wrong for 50% of KUL-AV-GC samples). Do not interpret SuperHuge column 0 as 'attended' at test time.
- Validation: report LODO with KUL-AV-GC as the proxy fold, aggregate per subject, and group folds by envelope identity (KUL2016 and KUL-UHD share all of their audio; KUL-AV-GC names are unreliable). Never score windows from training trials: spatial decoders gain up to 49 pp from trial fingerprints (Table 15).
- Shortcuts to neutralise: parity-locked labels (KUL-UHD, USTC) make any slow-drift or order feature a label proxy. The KUL2016 'loud column' lets an envelope-only model reach 82.5%; use envelope-symmetric model inputs or equalise the envelope scale per trial. The dataset fingerprint survives Euclidean alignment (84.3% with EA, 87.1% with causal running EA).
- Integrator: a 10-20 s moving average or a variance-normalised HMM with p_stay 0.98-0.999 per 0.25-s step is near the frontier (Table 20). Integrator choice is worth 0.019-0.068 H over argmax; raising SNR1 matters more (accuracy ~ Phi(0.9 * SNR1 * sqrt(w))). Choose on the median held-out dataset, and check the switch spacing against the scorer's 50-s property.
- Direction decoding: spatial/EOG cues do not transfer across datasets here. Unless the blind sets provide stream positions, direction labels must come from speaker decoding plus position metadata; confirm with the organisers what is available at test time.

## 12. Limitations

- MATLAB/Octave unavailable: the scorer is a verified Python port, not the official binary run; the synthetic-switch generator is also a port (same schedule, independent random stream), so SYNTHETIC numbers are not seed-identical to the MATLAB baseline.
- MNE, statsmodels and UMAP are not installed: TRFs and topographies use own implementations; embeddings use PCA; subject-level Wilcoxon tests replace mixed models.
- The published baseline's hold-out split and its synthetic-switch settings are not reproduced exactly, so Table 19 is not directly comparable with 'H 0.7002'.
- Spatial decoding used capped window counts (<= 24 5-s and <= 15 1-s windows per trial) to fit in memory; continuous windows were computed only for the KUL-AV-GC genuine-switch evaluation. The optional CCA decoder and a fixed-lag smoother were not run (the lookahead budget is fully used by the decision-stamping advance).
- Leave-pair-out replaces leave-one-trial-out for single-label datasets in classification analyses. With 4 trials per KUL-UHD subject it trains on 2 trials, so those numbers are noisy.
- The dry-run gate was applied to extraction, QC, caching and the spatial cache (one subject per dataset). Decoders were checked on the first subjects' outputs during the full run rather than in a separate dry run.
- Several scripts were edited by another session during the run (checkpointing and window capping in d3_spatial, the FFT null in c1_xcorr, figure titles). All affected steps were rerun, and the Phase E results reproduced exactly.
- The lookahead test has limited sensitivity to leaks shorter than about 0.5 s (see Phase E).

## 13. Reproducibility

Environment: .venv (Python 3.12.13); exact package versions in reports/eda_full/requirements.txt; every parameter in reports/eda_full/config.yaml (seed 20270101). Command: cd reports/eda_full && bash scripts/run_all.sh (every step is resumable from cache/). Hardware: 8 logical CPUs and 23 GB RAM, shared with other users (load average 5-15); at most 4 workers x 2 BLAS threads. Wall-clock time of this run, including writing and debugging the scripts: about 4 h (03:14-07:15 UTC, 2026-09-24); per-step timings are in logs/. Outputs: tables/ (all CSVs), figures/ (PNG at 150 dpi, each with its underlying CSV), logs/, cache/ (36 GB of intermediate files; data/ was never modified).
