Distraction — MEG Decoding of Speech Representations Under Reduced Attention

Question: Are speech representations automatic, or do they depend on focused attention?  ·  Study: Bonnaire2026 (Le Petit Prince, French MEG)  ·  Design: Listen (NoTask) vs Distraction (math / memory / list).

15 subjects · MEG 306 ch cache v1.3 192 configs · 128 cached · 64 ongoing 21 figures generated 2026-06-24 16:58

Headline findings

Early lexical decoding is automaticWordFrequency / WordLength / surprisal decode above chance and Listen ≈ Distraction — early lexical & predictive processing survives distraction.
Alpha band carries the attention effectThe clearest Listen > Distraction effect is in the alpha band (8–13 Hz); broadband decoding masks it.
Attention extends maintenance, not peakListen TGMs spread further off-diagonal — representations are held longer under attention, peak decodability is similar.
Manipulation decodable but tonicCondition decodes above chance (~0.29 vs 0.25) but flat across the epoch and pre-onset — a block/session effect, not word-locked.

Across the linguistic hierarchy a consistent story emerges (§8, §16): words and phonemes are represented under distraction, and early lexical and predictive processing is essentially unchanged — strong evidence for automaticity. Attention leaves its mark only at higher, slower levels: a band-specific alpha Listen>Distraction effect, a faint attention-dependence in deep contextual embeddings, and longer representational maintenance (broader temporal-generalisation). The manipulation check confirms the conditions differ in the brain, but as a tonic, block-level state rather than word-locked processing. Each experiment below carries a coloured “What we learn” takeaway.

Reading guide

■ Listen (NoTask)■ Distraction (math/memory/list)■ Listen − Distraction■ sig. cluster (p<.05)

Decoding curves show mean across subjects ± 95% CI; dotted vertical line = word onset. Method: continuous features (frequency, surprisal, embeddings) use ridge regression scored by correlation r; categorical targets (phoneme, POS, condition) use one-hot ridge / a classifier — condition decoding alone is balanced accuracy. Each section's exact config is under "Show config".

Configuration coverage

SectionConfigsCache status
acoustic4cached (15/15)
acoustic_phoneme4cached (15/15)
phoneme_id2cached (15/15)
lexical12cached (15/15)
pos_decoding4cached (15/15)
condition_decoding2cached (15/15)
surprisal4cached (15/15)
entropy4cached (15/15)
llm8cached (15/15)
residual_surprisal4cached (15/15)
timegen12cached (15/15)
oscillations60cached (15/15)
encoding16ongoing (0)
oov12ongoing (0)
oov_temporal12ongoing (0)
llm_context_size32partial (8/32)
Total192128 cached · 64 ongoing

Level 1Acoustic — word-locked above chance

Hypothesis

The speech envelope and mel spectrum are low-level acoustic features that cortex should track automatically at word onset, forming the sensory bedrock on which all lexical processing is built. Attention is not expected to gate input this early (§4.1).

What the figure shows

Sliding-window ridge decoding of SpeechEnvelope and MelSpectrum time-locked to word onset, Listen vs Distraction. Acoustics do not depend on syntactic category, so the data are split by condition only (not content / non-content) to preserve power.

Interpretation

Acoustic information is recoverable around word onset in both conditions: the auditory signal reaches and is represented in cortex comparably whether or not the listener is distracted.
acoustic__MelSpectrum.png
MelSpectrum
acoustic__SpeechEnvelope.png
SpeechEnvelope

What we learn

The acoustic input is faithfully tracked in cortex even under distraction — the sensory floor is intact. Any condition difference found later must therefore arise downstream, not from degraded input.
▸ Show config (4 configs)
{
  "_section": "acoustic",
  "n_configs": 4,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "ToStatic",
    "extractor": {
      "name": "SpeechEnvelope",
      "frequency": 100.0,
      "aggregation": "single",
      "allow_missing": true
    },
    "event_types": "Word"
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask'",
    "train: index>=0  ||  test: condition=='NoTask'"
  ]
}

Level 1Acoustic — phoneme-locked above chance

Hypothesis

Realigning the same acoustic features to phoneme onset indexes fine-grained, phoneme-rate sensory tracking — a stricter test of automatic low-level encoding than the word-locked version.

What the figure shows

SpeechEnvelope / MelSpectrum decoded time-locked to phoneme onset (≈3–4× the word rate), Listen vs Distraction.

Interpretation

Phoneme-rate acoustic tracking is present in both conditions, providing the sub-lexical sensory baseline against which the phoneme-identity analysis is read.
acoustic_phoneme__MelSpectrum.png
MelSpectrum
acoustic_phoneme__SpeechEnvelope.png
SpeechEnvelope

What we learn

Sensory tracking persists down to the phoneme timescale regardless of attention — distraction does not blunt the fast, low-level acoustic encoding of speech.
▸ Show config (4 configs)
{
  "_section": "acoustic_phoneme",
  "n_configs": 4,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Phoneme",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "ToStatic",
    "extractor": {
      "name": "SpeechEnvelope",
      "frequency": 100.0,
      "aggregation": "single",
      "allow_missing": true
    },
    "event_types": "Phoneme"
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask'",
    "train: index>=0  ||  test: condition=='NoTask'"
  ]
}

Level 2Phoneme identity above chance

Hypothesis

Phoneme identity is a categorical sub-lexical code. If sub-lexical processing is automatic, phoneme identity should be decodable from MEG and remain so under distraction.

What the figure shows

LabelEncoder decoding of phoneme identity (one-hot target scored by correlation r), phoneme-locked, Listen vs Distraction.

Interpretation

Phoneme identity is recoverable in both conditions — the brain distinguishes which phoneme was heard whether attending or distracted, consistent with the project's claim that words and phonemes are represented without focused attention (§16).
phoneme_id__LabelEncoder_text.png
LabelEncoder text

What we learn

Distraction does not erase phoneme-level representations: sub-lexical linguistic units are encoded automatically, one level above raw acoustics.
▸ Show config (2 configs)
{
  "_section": "phoneme_id",
  "n_configs": 2,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Phoneme",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "LabelEncoder",
    "event_types": "Phoneme",
    "event_field": "text",
    "aggregation": "trigger",
    "return_one_hot": true
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask'",
    "train: index>=0  ||  test: condition=='NoTask'"
  ]
}

Level 3Lexical — frequency / length / embedding null (L ≈ D)

Hypothesis

Static word features (frequency, length) are tagged rapidly and automatically by cortex (~100–250 ms), so they should decode equally well in Listen and Distraction (Hypothesis 1, §4.1).

What the figure shows

Decoding of WordFrequency, WordLength and the 300-d spaCy static embedding, split by content vs non-content words, with a Listen−Distraction difference panel and a cluster-permutation test.

Interpretation

All three features decode clearly above chance, peaking ~100–200 ms (WordFrequency r≈0.12; WordLength r≈0.08–0.10; spaCy r≈0.02–0.03). The Listen and Distraction curves overlap almost entirely and the difference panel shows no significant cluster (§7.1).
lexical__SpacyEmbedding.png
SpacyEmbedding
lexical__WordFrequency.png
WordFrequency
lexical__WordLength.png
WordLength

What we learn

This is the project's central result: low-level lexical features are encoded automatically and rapidly, independent of attentional state. (A null is suggestive of automaticity, not proof of equivalence — formal TOST / Bayes-factor tests are still pending.)
▸ Show config (12 configs)
{
  "_section": "lexical",
  "n_configs": 12,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "WordFrequency",
    "aggregation": "trigger",
    "language": "fr"
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition!='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']"
  ]
}

Level 3Part-of-speech above chance

Hypothesis

If the cortical signal carries syntactic-category information, part-of-speech should be decodable from word-locked MEG; whether attention modulates it is an open question.

What the figure shows

LabelEncoder decoding of POS category (French tags NC / V / ADJ / ADV…) from word-locked MEG, Listen vs Distraction across content and non-content words.

Interpretation

Probes whether neural activity discriminates grammatical category and whether that discrimination survives distraction — a bridge between lexical form (Level 3) and the contextual / compositional level (Level 4).
pos_decoding__LabelEncoder_pos.png
LabelEncoder pos

What we learn

Tests whether grammatical category is represented automatically, linking word-form decoding to the higher, meaning-level analyses that follow.
▸ Show config (4 configs)
{
  "_section": "pos_decoding",
  "n_configs": 4,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "LabelEncoder",
    "event_types": "Word",
    "event_field": "pos",
    "aggregation": "trigger",
    "return_one_hot": true
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition!='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']"
  ]
}

Level 3Condition decoding (manipulation check) manipulation check

Hypothesis

If the distraction manipulation is effective, the task condition (NoTask / math / memory / list) should itself be decodable from MEG above chance.

What the figure shows

A 4-way RidgeClassifierCV scored by balanced accuracy, collapsed over time into a per-subject bar / strip: the decoder against a label-shuffling DummyClassifier and the 1/4 theoretical-chance line (per-subject points with mean ± 95% CI).

Interpretation

Condition decodes above chance (decoder ≈0.29 vs 0.25; vs 1/4 t-test p≈3.8e-5; vs shuffled Wilcoxon p≈3e-4). Crucially the accuracy is flat across the whole epoch and already present before word onset — a tonic state/session difference, not word-locked processing. Because condition is blocked per run and CV groups by timeline, this partly reflects run/session structure rather than linguistic signal, which is why it is shown time-collapsed rather than as a time-course.
condition_decoding__balanced_accuracy.png
balanced accuracy

What we learn

The manipulation leaves a clear, sustained brain-state signature — but it is tonic and block-confounded. It confirms the conditions differ at the session level without proving a word-locked attentional effect.
▸ Show config (2 configs)
{
  "_section": "condition_decoding",
  "n_configs": 2,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "LabelEncoder",
    "event_types": "Word",
    "event_field": "condition",
    "aggregation": "trigger",
    "return_one_hot": false
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeClassifierCV"
    }
  },
  "scoring": {
    "name": "bac"
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: index>=0"
  ]
}

Level 4Surprisal (−log p(word|context)) null (L ≈ D)

Hypothesis

Predictive processing — tracking a word's surprisal given its context — is automatic and should survive distraction (Hypothesis 5, §4.5).

What the figure shows

Decoding of per-word Mistral-7B surprisal, content vs non-content words, Listen vs Distraction with a difference panel.

Interpretation

Surprisal decodes strongly in both conditions, peaking r≈0.06–0.08 around 0–200 ms, with content words giving a cleaner, higher peak; no Listen−Distraction difference emerges (§7.3).
surprisal__ConditionalLogProba.png
ConditionalLogProba

What we learn

The brain robustly tracks word predictability whether or not it is attending — online predictive processing of incoming words appears automatic, like the lexical features.
▸ Show config (4 configs)
{
  "_section": "surprisal",
  "n_configs": 4,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "ConditionalLogProba",
    "model_name": "mistralai/Mistral-7B-v0.1",
    "context_size": 128,
    "dim_cast_surprise": 1,
    "pretrained": true,
    "aggregation": "trigger",
    "contextualized": true,
    "infra": {
      "folder": "/data/jbonnaire/cache/Bonnaire2026"
    }
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition!='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']"
  ]
}

Level 4Entropy (H(word|context)) above chance

Hypothesis

Forward predictive uncertainty — entropy over the next word — may be attention-dependent: a Listen>Distraction effect would mean predictions are actively generated only under attention (§8).

What the figure shows

Decoding of per-word Mistral-7B entropy, content vs non-content, Listen vs Distraction with a difference panel.

Interpretation

Tests whether forward uncertainty is tracked and, via the difference panel, whether that tracking is gated by attention — the natural complement to the (automatic) surprisal result.
entropy__LexicalEntropy.png
LexicalEntropy

What we learn

Separates reacting to a word's surprisal (automatic) from predicting ahead (possibly attention-gated). The difference panel is where a Listen>Distraction prediction effect would surface if predictions require attention.
▸ Show config (4 configs)
{
  "_section": "entropy",
  "n_configs": 4,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "LexicalEntropy",
    "model_name": "mistralai/Mistral-7B-v0.1",
    "context_size": 128,
    "dim_cast_surprise": 1,
    "pretrained": true,
    "aggregation": "trigger",
    "contextualized": true,
    "infra": {
      "folder": "/data/jbonnaire/cache/Bonnaire2026"
    }
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition!='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']"
  ]
}

Level 4LLM contextual embeddings (Mistral-7B) Listen > Distraction

Hypothesis

Deeper, contextual representations should be more attention-sensitive than surface ones: attention gates compositional / contextual processing but not surface word form (Hypothesis 2, §4.2).

What the figure shows

Decoding of Mistral-7B embeddings at L1 (first layer — surface form) and L16 (middle layer — contextual / compositional), content vs non-content, Listen vs Distraction.

Interpretation

L1 (surface) is essentially flat / barely above zero in both conditions. L16 (contextual) is marginally above chance with a small Listen>Distraction trend around 0–500 ms, though confidence intervals overlap (§7.2).
llm__HuggingFaceText_L0.0.png
HuggingFaceText L0.0
llm__HuggingFaceText_L0.5.png
HuggingFaceText L0.5

What we learn

The little contextual representation that is decodable shows a hint of attention-dependence at the compositional layer but not the surface layer — the clearest sign that attention preferentially gates higher-level meaning, even if the effect is weak.
▸ Show config (8 configs)
{
  "_section": "llm",
  "n_configs": 8,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "HuggingFaceText",
    "model_name": "mistralai/Mistral-7B-v0.1",
    "layers": 0.0,
    "token_aggregation": "last",
    "contextualized": true,
    "aggregation": "trigger",
    "infra": {
      "folder": "/data/jbonnaire/cache/Bonnaire2026"
    }
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition!='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']"
  ]
}

Level 4Residualized surprisal (⊥ word frequency) above chance

Hypothesis

The genuinely contextual part of surprisal — with the word-frequency confound regressed out — is what truly reflects prediction rather than lexical familiarity.

What the figure shows

Decoding of surprisal after OLS-removing word-frequency variance (precomputed residual column), Listen vs Distraction.

Interpretation

Read against the raw-surprisal section, this shows how much of the surprisal decoding survives once its overlap with simple word frequency is removed.
residual_surprisal__EventField_residual_surprisal.png
EventField residual surprisal

What we learn

Disentangles contextual prediction from the lexical-frequency confound — it tells us whether surprisal decoding reflects genuine prediction or merely how common the word is.
▸ Show config (4 configs)
{
  "_section": "residual_surprisal",
  "n_configs": 4,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "EventField",
    "event_types": "Word",
    "event_field": "residual_surprisal",
    "aggregation": "trigger"
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition!='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']"
  ]
}

Cross-cuttingFrequency-band decoding Listen > Distraction

Hypothesis

If an attention effect exists at all, it should be clearest in the alpha band (8–13 Hz), classically tied to attentional engagement and suppression (Hypothesis 3, §4.3).

What the figure shows

WordFrequency / WordLength / spaCy decoded after bandpass filtering MEG into delta / theta / alpha / beta / gamma; for each band a Listen-vs-Distraction curve plus a Listen−Distraction difference with a cluster test (content words).

Interpretation

Alpha is the strongest finding in the project: clear above-chance decoding with Listen consistently above Distraction (a significant cluster). Beta is also above chance with a weaker L>D trend. The broadband null hides this because the effect lives in a specific band (§7.4).
oscillations__SpacyEmbedding.png
SpacyEmbedding
oscillations__WordFrequency.png
WordFrequency
oscillations__WordLength.png
WordLength

What we learn

Attention does modulate lexical decoding — but selectively, in the alpha band. The broadband null masks a real, band-specific Listen>Distraction effect: alpha oscillations carry more lexical information under attentive listening.
▸ Show config (60 configs)
{
  "_section": "oscillations",
  "n_configs": 60,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "WordFrequency",
    "aggregation": "trigger",
    "language": "fr"
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition!='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']"
  ]
}

Cross-cuttingTemporal generalisation (TGM) Listen > Distraction

Hypothesis

Attention may not change peak decodability but may extend how long a representation is maintained, seen as off-diagonal spread in the temporal-generalisation matrix (Hypothesis 4, §4.4).

What the figure shows

King & Dehaene temporal-generalisation matrices — a decoder trained at each train-time and tested at every test-time — shown as Listen | Distraction | their difference (content words).

Interpretation

Above-chance decoding concentrates on the main diagonal (a transient code). Listen TGMs are visually broader off-diagonal, especially for WordFrequency (§7.5).
timegen__SpacyEmbedding.png
SpacyEmbedding
timegen__WordFrequency.png
WordFrequency
timegen__WordLength.png
WordLength

What we learn

Attention appears to extend the maintenance of word representations rather than boost their peak — the same information is held for longer when listening attentively.
▸ Show config (12 configs)
{
  "_section": "timegen",
  "n_configs": 12,
  "runner": "TimeGen",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "WordFrequency",
    "aggregation": "trigger",
    "language": "fr"
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition!='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']"
  ]
}

Overall picture — what it means

What the null means (and doesn't)

Broadband lexical decoding (frequency, length, surprisal) is nearly identical in Listen and Distraction — strong evidence that early lexical processing (~100–300 ms) is automatic. But non-significance is not equivalence: TOST + Bayes-factor tests are still needed before claiming the conditions are truly equal.

Where attention does matter

Four convergent places: alpha-band decoding (Listen>Distraction, the most robust effect); temporal maintenance (Listen TGMs spread further off-diagonal); deep contextual embeddings (small L16 trend); and — from the companion evoked/source analyses — a larger late sustained response (>300 ms) under attentive listening.

Take-home

Speech representations are largely automatic: words, phonemes and surprisal survive distraction. What attention adds is not the initial representation but its maintenance and contextual elaboration — visible in alpha oscillations, off-diagonal TGM spread and deep LLM layers, not in broadband peak decodability.

Data note: all figures use cache v1.3, which fixed a ~10 s word-onset misalignment for sub-01 / sub-04 (weak MISC005 trigger). Earlier v1.2 figures for those subjects are invalid.

Ongoing — not yet computed

These cross-cutting analyses are scripted in exps_all.py but not yet run (64 configs). They will populate this report once computed.

Encoding models (contextual features) ongoing

Reverse direction: predict MEG from surprisal / entropy / 2 LLM layers (RidgeCV, alpha_per_target). Isolates each feature's neural contribution. 16 configs — none cached yet.

Status: 16 configs · none cached yet — run run_exp exps_all encoding

▸ Show planned config
{
  "_section": "encoding",
  "n_configs": 16,
  "runner": "SlidingWindow",
  "mode": "encod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "ConditionalLogProba",
    "model_name": "mistralai/Mistral-7B-v0.1",
    "context_size": 128,
    "dim_cast_surprise": 1,
    "pretrained": true,
    "aggregation": "trigger",
    "contextualized": true,
    "infra": {
      "folder": "/data/jbonnaire/cache/Bonnaire2026"
    }
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": true,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition!='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']"
  ]
}

Out-of-vocabulary subscoring (group CV by word) ongoing

Standard CV but word identity is the CV group, so no word form is seen in train and test — tests whether decoding generalises to unseen words. 12 configs — none cached yet.

Status: 12 configs · none cached yet — run run_exp exps_all oov

▸ Show planned config
{
  "_section": "oov",
  "n_configs": 12,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "WordFrequency",
    "aggregation": "trigger",
    "language": "fr"
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "text",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition!='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']"
  ]
}

Temporal OOV (first→second half transfer) ongoing

Train on the first half of each timeline, test on the second, grouped by word form — a stricter cross-time transfer test. 12 configs — none cached yet.

Status: 12 configs · none cached yet — run run_exp exps_all oov_temporal

▸ Show planned config
{
  "_section": "oov_temporal",
  "n_configs": 12,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "WordFrequency",
    "aggregation": "trigger",
    "language": "fr"
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "text",
  "query_pairs": [
    "train: chunk=='first_half'  ||  test: chunk=='second_half' and condition!='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: chunk=='first_half'  ||  test: chunk=='second_half' and condition!='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: chunk=='first_half'  ||  test: chunk=='second_half' and condition=='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: chunk=='first_half'  ||  test: chunk=='second_half' and condition=='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']"
  ]
}

LLM context-size ablation ongoing

Does neural tracking of LLM embeddings build up with longer context (4/16/64/128 words) under Listen but not Distraction? Tests 'context does not build up' (§4.6).

Status: 8/32 configs cached · 24 still to run — run run_exp exps_all llm_context_size

▸ Show planned config
{
  "_section": "llm_context_size",
  "n_configs": 32,
  "runner": "SlidingWindow",
  "mode": "decod",
  "study_chain": "Chain",
  "event_type": "Word",
  "epoch": "-2.0 \u2026 2.0 s",
  "extractor": {
    "name": "HuggingFaceText",
    "model_name": "mistralai/Mistral-7B-v0.1",
    "layers": 0.0,
    "token_aggregation": "last",
    "contextualized": true,
    "aggregation": "trigger",
    "infra": {
      "folder": "/data/jbonnaire/cache/Bonnaire2026"
    }
  },
  "model": {
    "sklearn_model": {
      "name": "RidgeCV",
      "alpha_per_target": false,
      "y_scaler": "StandardScaler"
    }
  },
  "cv.group": "timeline",
  "query_pairs": [
    "train: index>=0  ||  test: condition!='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition!='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and not pos in ['NC', 'V', 'ADJ', 'ADV']",
    "train: index>=0  ||  test: condition=='NoTask' and pos in ['NC', 'V', 'ADJ', 'ADV']"
  ]
}