From 763aa9f97640a09ca4fb85aef7b5c9ef7ff452f6 Mon Sep 17 00:00:00 2001 From: Seyed Yahya Shirazi Date: Wed, 20 May 2026 14:39:52 -0700 Subject: [PATCH] narrative review phase 4: paper-review + humanizer --- manuscript/narrative-review/manuscript.md | 38 +++-- .../narrative-review/reviews/humanizer-log.md | 68 +++++++++ .../reviews/internal-review.md | 141 ++++++++++++++++++ 3 files changed, 227 insertions(+), 20 deletions(-) create mode 100644 manuscript/narrative-review/reviews/humanizer-log.md create mode 100644 manuscript/narrative-review/reviews/internal-review.md diff --git a/manuscript/narrative-review/manuscript.md b/manuscript/narrative-review/manuscript.md index 61febe4..805c54c 100644 --- a/manuscript/narrative-review/manuscript.md +++ b/manuscript/narrative-review/manuscript.md @@ -32,7 +32,7 @@ word_budget: ## Abstract -Naturalistic-stimulus neuroscience has moved in two waves from whole-clip inter-subject correlation (ISC) metrics to event-locked methods that interrogate individual shots. Most empirical evidence to date comes from adult functional magnetic resonance imaging (fMRI), adult intracranial electroencephalography (iEEG), or adult scalp electroencephalography (EEG) ISC. Per-shot event-related spectral perturbation (ERSP) in a developmental cohort viewing silent character animation has no published precedent. We review the four-strand corpus that constrains this design space, argue that psychophysics, action, language, and emotion make divergent and partly-falsifiable predictions about the 0 to 500 ms post-shot-onset window, and lay out the topographic-and-band rejection region a pre-registered group analysis can adopt before opening the data. The Healthy Brain Network EEG Release 3 cohort viewing *The Present* (Pixar 2014) sits at this empty intersection. +Naturalistic-stimulus neuroscience has moved from whole-clip inter-subject correlation (ISC) to event-locked methods that interrogate individual shots. Most empirical evidence is adult functional magnetic resonance imaging (fMRI), adult intracranial electroencephalography (iEEG), or adult scalp electroencephalography (EEG) ISC. Per-shot event-related spectral perturbation (ERSP) in a developmental cohort viewing silent character animation has no published precedent. We review the corpus that constrains this design space, argue that psychophysics, action, language, and emotion make divergent and partly-falsifiable predictions about the 0 to 500 ms post-shot-onset window, and lay out the topographic-and-band rejection region a pre-registered group analysis can adopt before opening the data. The Healthy Brain Network EEG Release 3 cohort viewing *The Present* (Pixar 2014) sits at this empty intersection, with the local 100 Hz working set capping beta-band claims until a 500 Hz validation pass. ## 1. Introduction: the per-shot turn @@ -40,15 +40,17 @@ Naturalistic-stimulus neuroscience moved from controlled gratings to feature fil A separate developmental tradition has used Pixar shorts in fMRI to map theory of mind (ToM) and pain networks in children as young as three [Richardson2018DevelopmentOT] and silent abstract animation to improve magnetic resonance imaging (MRI) compliance and reveal reliable network-level activity [Vanderwal2015InscapesAM]. Cross-sectional EEG-ISC across ages 6 to 44 is the closest electrophysiological developmental anchor; ISC is highest in children and declines into adulthood [Petroni2018TheVO]. None of these traditions has reported per-shot ERSP at the 0 to 500 ms post-onset window in a child cohort viewing animation. -This Review argues that four research perspectives, psychophysics, action, language, and emotion, make divergent and partly-falsifiable predictions about this empty cell. Sections 2 to 6 develop the perspectives in order. Section 7 synthesises them into a topographic-and-band rejection region that a pre-registered group analysis can adopt before opening the data. Box 1 anchors the argument to the Healthy Brain Network EEG (HBN-EEG) Release 3 cohort viewing *The Present* (Pixar 2014), the empty-cell stimulus that motivates the Review. +This review argues that four research perspectives, psychophysics, action, language, and emotion, make divergent and partly-falsifiable predictions about this empty cell. Sections 2 to 6 develop the perspectives in order. Section 7 synthesises them into a topographic-and-band rejection region that a pre-registered group analysis can adopt before opening the data. Box 1 anchors the argument to the Healthy Brain Network EEG (HBN-EEG) Release 3 cohort viewing *The Present* (Pixar 2014), the empty-cell stimulus that motivates the review. ## 2. The four-perspective scaffold The four-perspective scaffold is structural rather than decorative. Each perspective makes a different *kind* of prediction. Psychophysics names a regressor of no interest that must be partialled before any social claim can be defended. Action names a band-and-topography prediction (mu-band event-related desynchronisation [ERD] over central rolandic cortex) with adult precedent. Language names a method that structurally cannot transfer (language-model surprisal aligned to spoken transcripts) plus a positive sub-thread of silent-narrative findings that does transfer. Emotion names two distinct predictions at incompatible latencies (early occipital alpha desynchronisation and later frontal-asymmetric alpha). Together the four make a hierarchy of prior evidence depth that the data can rerank. -The perspectives cross 15 corpus themes catalogued in our Phase 2 science map. Theme 1 (ISC as reliability metric) is the most heavily represented analytic theme, originating in fMRI [Hasson2004IntersubjectSO] and migrating to EEG [dmochowski2012correlated], MEG [Lankinen2014IntersubjectCO], peripheral physiology [Madsen2022CognitivePO], and audience prediction [Dmochowski2014AudiencePA]. Theme 2 (event boundaries and segmentation) is anchored in event-segmentation theory and HMM event-state recovery [zacks2007event; baldassano2017event; speer2007narrative; Ben-Yakov2018TheHF]. Theme 3 (naturalness gradient) defines the stimulus continuum from controlled gratings to Heider-Simmel triangles to character animation to live-action film (Figure 2). Theme 4 (low-level feature regressors) is the psychophysics backbone [Adelson1985SpatiotemporalEM; Carandini2011NormalizationAA; Nishimoto2011ReconstructingVE]. Theme 6 (mu rhythm and action observation) anchors the action perspective [hari1998action; pineda2005mu]. Theme 9 (language models as regressors) is the language perspective's structural comparator [Goldstein2022SharedCP; Caucheteux2022BrainsAA]. Themes 7, 12, and 13 (affective dynamics; pet, animal, and baby-schema affective response; developmental neuroimaging in cinematic paradigms) collectively carry the emotion perspective. Themes 5, 8, 10, 11, 14, and 15 cut across multiple perspectives. Figure 1 maps each perspective to the themes it contributes to substantially. +The perspectives cross 15 corpus themes catalogued in our Phase 2 science map (Figure 1). Two themes anchor the analytic backbone independent of perspective: ISC as a reliability metric (Theme 1), originating in fMRI [Hasson2004IntersubjectSO] and migrating to EEG [dmochowski2012correlated], MEG [Lankinen2014IntersubjectCO], peripheral physiology [Madsen2022CognitivePO], and audience prediction [Dmochowski2014AudiencePA]; and event segmentation (Theme 2), anchored in event-segmentation theory and hidden-Markov-model event-state recovery [zacks2007event; baldassano2017event; speer2007narrative; Ben-Yakov2018TheHF]. Theme 3 (naturalness gradient; Figure 2) places the stimulus on a continuum from controlled gratings to live-action film, with character animation as the intermediate point that motivates the empty-cell framing. -Theme overlap is intentional. The perspectives interact at the per-shot ERSP level rather than partitioning the variance cleanly. Sections 3 to 6 develop the four perspectives in order, naming the band-by-topography signature each makes and the falsification region attached to each (Figure 4). Section 7 closes by combining the four rejection regions into a single pre-registerable test before group analysis. +The four perspectives then sit in specific corners of this theme space. Psychophysics owns Themes 4 (low-level feature regressors) [Adelson1985SpatiotemporalEM; Carandini2011NormalizationAA; Nishimoto2011ReconstructingVE], 5 (time-resolved EEG and MEG), and 11 (free-viewing EEG with eye coregistration). Action owns Themes 6 (mu rhythm and action observation) [hari1998action; pineda2005mu] and 8 (social cognition through biological motion) and contributes to Themes 2 and 14 (distributed multivariate signatures). Language owns Theme 9 (LMs as regressors) [Goldstein2022SharedCP; Caucheteux2022BrainsAA] as a structural comparator and Theme 10 (audiovisual integration), but its silent-narrative sub-thread cuts across Themes 8 (social cognition; default-mode network as narrative integrator) and 13 (developmental neuroimaging in cinematic paradigms). Emotion owns Themes 7 (affective dynamics), 12 (pet, animal, and baby-schema affective response), and 13. Theme 15 (predictive processing) is a cross-perspective unifier: it ties mu-band ERD to mirror-system prediction error, LM surprisal to next-word prediction, and event boundaries to prediction-error transients. + +Perspective overlap is intentional rather than residual; the perspectives interact at the per-shot ERSP level rather than partitioning variance cleanly. Sections 3 to 6 develop them in order, naming the band-by-topography signature each makes and the falsification region attached to each (Figure 4). Section 7 closes by combining the four rejection regions into a single pre-registerable test before group analysis. ## 3. Psychophysics: the bottom-up floor @@ -56,7 +58,7 @@ Psychophysics anchors the bottom-up floor that every per-shot analysis must clea The closest electrophysiological analogue to per-shot ERSP during naturalistic film is the intracranial study of Nentwich and colleagues, who showed that motion outranks luminance for occipitoparietal cortex when triple-regressed against optical-flow magnitude, saccade onsets, and film-cut onsets [Nentwich2023SemanticNM]. That result establishes a quantitative ranking among low-level regressors: per-shot log luminance ratio (LLR) is one of several low-level features that needs accounting. EEG ISC at the whole-clip scale tracks low-level features at occipital electrodes more strongly than higher-order content [dmochowski2012correlated; Madsen2022CognitivePO; Cohen2016MemorableAN], although attention strongly modulates this baseline [Ki2016AttentionSM]. An envelope-only auditory control isolating low-level acoustic structure from higher-level musical structure [Kaneshiro2021InterSubjectEC] is the methodological template the LLR-as-covariate plan inherits. -A second class of bottom-up drivers operates through the eye. Free-viewing EEG depends on eye-movement coregistration to separate stimulus-onset responses from saccade-locked and fixation-related potentials [Dimigen2011CoregistrationOE; Plöchl2012CombiningEA], and regression deconvolution of overlapping events is the methodological state of the art [Dimigen2021RegressionbasedAO]. Gaze coherence varies with stimulus class, highest on Hollywood trailers and lowest on stop-motion or stills [Dorr2010VariabilityOE]; a Pixar short sits between these extremes. The HBN-EEG cohort carries no synchronous eye tracker, which means a per-shot analysis cannot deconvolve overlapping saccade-locked transients from shot-onset responses. Independent component analysis (ICA)-based artifact rejection through adaptive mixture ICA (AMICA) and IC classification (ICLabel) is the operating compromise [Bell1997TheC]. The implication for per-shot ERSP is asymmetric: per-shot LLR is the minimum partialling for any social-content claim. Motion energy computed offline from the stimulus video is the named first follow-up regressor [Nishimoto2011ReconstructingVE; Nentwich2023SemanticNM]. The multivariate temporal response function (mTRF) toolbox supplies the production regression framework [Crosse2016TheMT]. Figure 2 places the empty cell on the naturalness gradient. +A second class of bottom-up drivers operates through the eye. Free-viewing EEG depends on eye-movement coregistration to separate stimulus-onset responses from saccade-locked and fixation-related potentials [Dimigen2011CoregistrationOE; Plöchl2012CombiningEA], and regression deconvolution of overlapping events is the methodological state of the art [Dimigen2021RegressionbasedAO]. Gaze coherence varies with stimulus class, highest on Hollywood trailers and lowest on natural movie clips and static images [Dorr2010VariabilityOE]; a Pixar short sits between these extremes. The HBN-EEG cohort carries no synchronous eye tracker, which means a per-shot analysis cannot deconvolve overlapping saccade-locked transients from shot-onset responses. Independent component analysis (ICA)-based artifact rejection through adaptive mixture ICA (AMICA) and IC classification (ICLabel) is the operating compromise [Bell1997TheC]. The implication for per-shot ERSP is asymmetric: per-shot LLR is the minimum partialling for any social-content claim. Motion energy computed offline from the stimulus video is the named first follow-up regressor [Nishimoto2011ReconstructingVE; Nentwich2023SemanticNM]. The multivariate temporal response function (mTRF) toolbox supplies the production regression framework [Crosse2016TheMT]. Figure 2 places the empty cell on the naturalness gradient. ## 4. Action: mu-band ERD and event segmentation @@ -66,29 +68,29 @@ Even with that tempering, the prediction is specific. Shots dominated by charact The second action beat is event segmentation. Speer and colleagues found posterior cingulate, middle-temporal, and posterior STS boundary-locked transients in fMRI during narrative listening [speer2007narrative]. Baldassano and colleagues recovered a hierarchy of event boundaries from Sherlock-movie fMRI using HMM, with hippocampal boundary signals predicting subsequent free recall [baldassano2017event]. Lerner and colleagues mapped temporal receptive windows from sensory cortex (milliseconds) to default-mode regions (tens of seconds) [lerner2011temporal]. Chen and colleagues showed event-specific patterns in the default-mode network are shared across viewers and reactivated at recall [chen2017shared]. Ben-Yakov and Henson distinguished within-event camera cuts, which produce minimal hippocampal responses, from across-event narrative boundaries, which produce robust ones [Ben-Yakov2018TheHF]. Magliano and Zacks supplied the behavioural foundation that viewers segment edited films along cuts independent of dialogue [Magliano2011TheIO]. -A third action beat concerns single-agent versus two-agent shots. Sliwa and Freiwald documented a dedicated cortical network in macaque for processing two-agent social interaction, separable from single-agent action perception [sliwa2017macaque]. This motivates excluding two-agent shots from a clean single-agent contrast, since the social-interaction network may dominate two-agent variance. Klin and colleagues showed that toddlers with autism orient to audiovisual contingency rather than upright biological motion [klin2009biological], and adolescents with autism fixate eyes 50 percent as often during emotionally evocative film viewing [klin2002visual]. The HBN cohort includes a substantial autism-spectrum subsample, making autism status a candidate moderator rather than a noise source. +A third action beat concerns single-agent versus two-agent shots. Sliwa and Freiwald documented a dedicated cortical network in macaque for processing two-agent social interaction, separable from single-agent action perception [sliwa2017macaque]. This motivates excluding two-agent shots from a clean single-agent contrast, since the social-interaction network may dominate two-agent variance. ## 5. Language: comparator of non-transfer plus silent-narrative sub-thread ### 5a. Language-model regressors are structurally non-transferable -The contemporary methodological mainstream in naturalistic neuroimaging is built around transformer-based language-model (LM) regressors aligned to spoken or read transcripts. Goldstein and colleagues showed pre-onset prediction, post-onset surprise, and contextual-embedding signatures shared between word-by-word intracranial cortical recording (ECoG) and autoregressive LMs [Goldstein2022SharedCP]. Each signature depends on speech-onset alignment. Heilbron and colleagues separated lexical, syntactic, and semantic surprisal regressors during MEG audiobook listening, all derived from LMs with word-onset alignment [Heilbron2020AHO]. Caucheteux and colleagues mapped transformer intermediate layers to fMRI and MEG responses to natural narrative [Caucheteux2022BrainsAA] and a cortical hierarchy of prediction timescales [Caucheteux2023EvidenceOA]. Antonello and colleagues documented log-linear scaling of brain prediction with LM parameter count up to 30B [Antonello2023ScalingLF]. Schrimpf and colleagues showed that next-word-prediction quality drives brain score on fMRI, ECoG, and reading-time benchmarks [schrimpf2021the]. Toneva and Wehbe used BERT to predict reading fMRI and MEG, with attention-head ablations linking brain prediction to natural-language processing performance [toneva2019interpreting]. Huth and colleagues built the canonical voxelwise word-embedding encoding atlas tiling cortex with semantic clusters; this method requires spoken transcripts [Huth2016NaturalSR]. Nelson and colleagues tracked open-node count during syntactic merge using intracranial high-gamma dynamics, explicitly reading-based [Nelson2017NeurophysiologicalDO]. The N400 family bridges to picture-context paradigms at the cost of dynamic stimulus [Kutas2011ThirtyYA; DeLong2005ProbabilisticWP]. +The contemporary methodological mainstream in naturalistic neuroimaging is built around transformer-based language-model (LM) regressors aligned to spoken or read transcripts. Goldstein and colleagues showed pre-onset prediction, post-onset surprise, and contextual-embedding signatures shared between word-by-word electrocorticography (ECoG) and autoregressive LMs [Goldstein2022SharedCP]. Each signature depends on speech-onset alignment. Heilbron and colleagues separated lexical, syntactic, and semantic surprisal regressors during MEG audiobook listening, all derived from LMs with word-onset alignment [Heilbron2020AHO]. Caucheteux and colleagues mapped transformer intermediate layers to fMRI and MEG responses to natural narrative [Caucheteux2022BrainsAA] and a cortical hierarchy of prediction timescales [Caucheteux2023EvidenceOA]. Antonello and colleagues documented log-linear scaling of brain prediction with LM parameter count up to 30B [Antonello2023ScalingLF]. Schrimpf and colleagues showed that next-word-prediction quality drives brain score on fMRI, ECoG, and reading-time benchmarks [schrimpf2021the]. Toneva and Wehbe used BERT to predict reading fMRI and MEG, with attention-head ablations linking brain prediction to natural-language processing performance [toneva2019interpreting]. Huth and colleagues built the canonical voxelwise word-embedding encoding atlas tiling cortex with semantic clusters; this method requires spoken transcripts [Huth2016NaturalSR]. Nelson and colleagues tracked open-node count during syntactic merge using intracranial high-gamma dynamics, explicitly reading-based [Nelson2017NeurophysiologicalDO]. The N400 family bridges to picture-context paradigms at the cost of dynamic stimulus [Kutas2011ThirtyYA; DeLong2005ProbabilisticWP]. Each method depends on word-level alignment to spoken or read stimuli. *The Present* is wordless. All seven Category G cards in our language ontology (and 12 cards corpus-wide) carry `transfer-to-silent: no`. A vision-side analogue, multimodal vision-language model embeddings or scene-difference deep-network features as continuous regressors, does not yet exist in the corpus for scalp-EEG ERSP. The Lipkin frontotemporal language network atlas [Lipkin2022ProbabilisticAF] is included as the negative-control region of interest in the falsification region of Section 7. ### 5b. Silent-narrative neural correlates that do transfer -A second strand of language-strand cards documents what silent narrative engages independent of speech. Castelli and colleagues showed that silent geometric-shape animations engage medial prefrontal cortex, the temporo-parietal junction, and STS when motion implies social interaction, with no speech required [Castelli2000MovementAM; castelli2000heider]; the same paradigm in autism shows reduced engagement [Castelli2002AutismAS]. Vanderwal and colleagues built Inscapes, a purpose-built silent abstract animation that improves MRI compliance and produces reliable network-level activity, used by the HBN cohort itself [Vanderwal2015InscapesAM]. Naci and colleagues used a Hitchcock excerpt as a covert assessment, showing that high-order cortex can be probed from a near-silent narrative [Naci2014ACN]. Lankinen and colleagues report source-space MEG reliable across viewers in occipital and temporal cortex during silent-visual and audiovisual movie conditions, the closest electrophysiological analogue with a deliberate silent-visual condition [Lankinen2014IntersubjectCO]. The Studyforrest infrastructure provides an audio-only foundation that has been extended to silent-cohort contrasts [Hanke2014AH7]. Schroeder and colleagues described modality-general delta- and theta-band phase alignment to attended event onsets, providing the mechanistic frame for shot-onset ERSP independent of speech [Schroeder2009LowfrequencyNO]. Senkowski and colleagues described transient gamma synchronisation and low-frequency phase coupling for cross-modal binding [Senkowski2008CrossmodalBT]. Buckner, Simony, Yeshurun, Mar, and Tamir developed the default-mode network as narrative integrator, with framing context driving within-stimulus divergence [Buckner2008TheBD; Simony2016DynamicRO; Yeshurun2017SameSD; Mar2011TheNB; Tamir2016ReadingFA]. +Silent-narrative neural correlates do transfer to scalp-EEG ERSP analysis even when language-model regressors cannot. Castelli and colleagues showed that silent geometric-shape animations engage medial prefrontal cortex, the temporo-parietal junction, and STS when motion implies social interaction, with no speech required [Castelli2000MovementAM; castelli2000heider]; the same paradigm in autism shows reduced engagement [Castelli2002AutismAS]. Vanderwal and colleagues built Inscapes, a purpose-built silent abstract animation that improves MRI compliance and produces reliable network-level activity, used by the HBN cohort itself [Vanderwal2015InscapesAM]. Naci and colleagues used a Hitchcock excerpt as a covert assessment, showing that high-order cortex can be probed from a near-silent narrative [Naci2014ACN]. Lankinen and colleagues report source-space MEG reliable across viewers in occipital and temporal cortex during silent-visual and audiovisual movie conditions, the closest electrophysiological analogue with a deliberate silent-visual condition [Lankinen2014IntersubjectCO]. The Studyforrest infrastructure provides an audio-only foundation that has been extended to silent-cohort contrasts [Hanke2014AH7]. Schroeder and colleagues described modality-general delta- and theta-band phase alignment to attended event onsets, providing the mechanistic frame for shot-onset ERSP independent of speech [Schroeder2009LowfrequencyNO]. Senkowski and colleagues described transient gamma synchronisation and low-frequency phase coupling for cross-modal binding [Senkowski2008CrossmodalBT]. Buckner, Simony, Yeshurun, Mar, and Tamir developed the default-mode network as narrative integrator, with framing context driving within-stimulus divergence [Buckner2008TheBD; Simony2016DynamicRO; Yeshurun2017SameSD; Mar2011TheNB; Tamir2016ReadingFA]. -The language perspective therefore enters our Review twice. Subsection 5a defines the methodological isolation of silent-stimulus designs: the dominant LM-as-regressor framework cannot transfer. Subsection 5b supplies the cortical substrates that silent narrative is expected to engage: medial prefrontal cortex, the temporo-parietal junction, the STS, and the default-mode network. Their independent-component-cluster analogues in EEG are the search regions the per-shot ERSP analysis targets. Figure 3 makes the gap structure explicit. +The language perspective therefore plays two roles. The 5a sub-thread isolates the silent-stimulus design from the dominant LM-as-regressor framework. The 5b sub-thread supplies the cortical substrates that silent narrative engages: medial prefrontal cortex, the temporo-parietal junction, the STS, and the default-mode network. Their independent-component-cluster analogues in EEG are the search regions for the per-shot ERSP analysis. Figure 3 makes the gap structure explicit. ## 6. Emotion: two predictions at different latencies -The emotion perspective makes two predictions with different latencies and different implicated structures. The first is an early visual-cortex emotion-schema response. Kragel and colleagues built EmoNet, a deep-learning model showing that emotion schemas are encoded in early visual cortex, predicting that emotion-tuned visual representations should appear in early-latency occipital ERSP [Kragel2018EmotionSA]. Saarimaki and colleagues decoded six basic emotions during emotional movie viewing using fMRI multi-voxel pattern analysis [Saarimäki2016DiscreteNS]; Cowen and Keltner extended the taxonomy to 27 distinguishable categories from short videos [Cowen2017SelfreportC2]. Distributed-network meta-analysis argues for distributed signatures over strict regional localisation [Lindquist2012TheBB], with the neurologic pain signature as a methodological exemplar of multivariate signatures of affect [Wager2013AnFN]. The closest EEG correlate at the 0 to 500 ms scale is alpha-band desynchronisation during emotional pictures. Codispoti and colleagues (2023) review the EEG alpha-band literature on emotional picture perception and conclude that alpha desynchronisation is a robust correlate of attentional engagement by emotional stimuli, with parametric arousal modulation [Codispoti2023AlphabandOA]. Whether this transfers to dynamic naturalistic stimuli at sub-second timescales in a child cohort is untested. +The emotion perspective makes two predictions with different latencies and different implicated structures. The first is an early visual-cortex emotion-schema response. Kragel and colleagues built EmoNet, a deep-learning model showing that emotion schemas are encoded in early visual cortex, predicting that emotion-tuned visual representations should appear in early-latency occipital ERSP [Kragel2018EmotionSA]. Saarimaki and colleagues decoded six basic emotions during emotional movie viewing using fMRI multi-voxel pattern analysis [Saarimäki2016DiscreteNS]; Cowen and Keltner extended the taxonomy to 27 distinguishable categories from short videos [Cowen2017SelfreportC2]. Distributed-network meta-analysis argues for distributed signatures over strict regional localisation [Lindquist2012TheBB], with the neurologic pain signature as a methodological exemplar of multivariate signatures of affect [Wager2013AnFN]. The closest EEG correlate at the 0 to 500 ms scale is early occipital alpha desynchronisation (80 to 300 ms post-shot-onset, extrapolated from static-picture latencies). Codispoti and colleagues (2023) review the EEG alpha-band literature on emotional picture perception and conclude that alpha desynchronisation is a robust correlate of attentional engagement by emotional stimuli, with parametric arousal modulation [Codispoti2023AlphabandOA]. Whether this transfers to dynamic naturalistic stimuli at sub-second timescales in a child cohort is untested. The second prediction is a longer-latency cuteness or affiliative response. Stoeckel and colleagues reported common activation across child and dog spanning emotion, reward, affiliation, visual processing, and social cognition regions in adult mothers viewing photographs of own child versus own dog [Stoeckel2014PatternsOB]. Glocker and colleagues showed that baby schema parametrically modulates nucleus accumbens reward in adults [Glocker2009BabySM]. Borgi and colleagues demonstrated that children aged 3 to 6 already show parametric cuteness ratings and gaze bias for human infant, puppy, and kitten faces [Borgi2014BabySI]; this is the behavioural anchor that the cuteness response is established well before adolescence. The interpretation implication is that Stoeckel measures identity-level pair-bonding and Borgi measures generic baby schema. HBN viewers have no identity-level bond with an animated puppy, so the relevant inference is from generic baby schema rather than pair-bonding circuitry. -Two EEG routes connect these predictions to observables. The first is alpha-band desynchronisation as an arousal-modulated correlate of attentional engagement [Codispoti2023AlphabandOA]. The second is frontal alpha asymmetry as an approach-withdrawal index [Davidson2000AffectiveSP; Coan2004FrontalEA]. An updated meta-analytic critique documents smaller effect sizes and substantial reliability concerns [Reznik2018FrontalAA]. The corpus contains no card applying asymmetry analysis to per-event sub-second windows during a continuous naturalistic stimulus, and none in a developmental cohort viewing film. Frontal asymmetry at shot-onset latency is therefore exploratory rather than confirmatory. The third emotion beat is social cognition. Richardson and colleagues documented ToM and pain networks present from age three and refining with age, using Pixar shorts in 122 children [Richardson2018DevelopmentOT]; this is the load-bearing developmental anchor. Mar synthesised narrative comprehension as a social-cognitive activity [Mar2011TheNB]; Singer and colleagues documented affective pain-region engagement during observed pain [Singer2004EmpathyFP]; Zaki and Ochsner formalised the tripartite empathy model bridging experience sharing and mental-state attribution [Zaki2012TheNO]. Nummenmaa and colleagues showed emotion intensity modulates ISC in midline cortex during film viewing [Nummenmaa2012EmotionsPS]; Schmaelzle and Grall theorised ISC as audience captivation [Schmälzle2020TheCB]. Two predictions sit at incompatible latencies and topographies; an LLR-partialled per-shot generalised linear model (GLM) adjudicates between them. +Two EEG routes connect these predictions to observables. The first is early occipital alpha-band desynchronisation (80 to 300 ms) as an arousal-modulated correlate of attentional engagement [Codispoti2023AlphabandOA]. The second is later frontal alpha asymmetry (200 to 500 ms; extrapolated downward from the seconds-to-minutes Davidson tradition) as an approach-withdrawal index [Davidson2000AffectiveSP; Coan2004FrontalEA]. An updated meta-analytic critique documents smaller effect sizes and substantial reliability concerns [Reznik2018FrontalAA]. The corpus contains no card applying asymmetry analysis to per-event sub-second windows during a continuous naturalistic stimulus, and none in a developmental cohort viewing film. Frontal asymmetry at shot-onset latency is therefore exploratory rather than confirmatory. The third emotion beat is social cognition. Richardson and colleagues documented ToM and pain networks present from age three and refining with age, using Pixar shorts in 122 children [Richardson2018DevelopmentOT]; this is the load-bearing developmental anchor. Mar synthesised narrative comprehension as a social-cognitive activity [Mar2011TheNB]; Singer and colleagues documented affective pain-region engagement during observed pain [Singer2004EmpathyFP]; Zaki and Ochsner formalised the tripartite empathy model bridging experience sharing and mental-state attribution [Zaki2012TheNO]. Nummenmaa and colleagues showed emotion intensity modulates ISC in midline cortex during film viewing [Nummenmaa2012EmotionsPS]; Schmaelzle and Grall theorised ISC as audience captivation [Schmälzle2020TheCB]. Two predictions sit at incompatible latencies and topographies; an LLR-partialled per-shot generalised linear model (GLM) adjudicates between them. ## 7. Synthesis: integration, falsifiability, and open questions @@ -104,13 +106,9 @@ External precedent: Petroni and colleagues recorded 64-channel EEG at 500 Hz fro A topographic-and-band rejection region for the four-perspective ranking can be pre-registered before group analysis. A surviving central-rolandic mu-band cluster (electrodes C3, Cz, and C4; 8 to 13 Hz) confirms the action prediction. A surviving frontal-asymmetric alpha cluster (electrodes F3 and F4; 8 to 13 Hz) confirms the emotion prediction. A surviving cluster in left frontotemporal IC space, overlapping the Lipkin language-network atlas [Lipkin2022ProbabilisticAF] used as a negative-control mask, falsifies the four-perspective ranking by relocating the surviving signal into a perspective the thesis says should not transfer. A null result on the LLR-partialled GLM at a pre-registered cluster-level alpha (p < 0.05 corrected by mass-univariate cluster-based permutation, with the mTRF toolbox precedent [Crosse2016TheMT]) also falsifies the four-perspective ranking, by localising per-shot ERSP variance entirely to bottom-up features in this cohort. Pinning the rejection region before data analysis is the publication discipline that constrains analyst degrees of freedom. -### 7.4 Narrative-position objection - -Boy-only and puppy-only shots in *The Present* differ on three-act narrative position; boy-only clusters in the early-act setup, puppy-only in the late-act resolution. Any boy-vs-puppy ERSP difference may therefore be confounded with prediction-error or arousal trajectories driven by narrative position. The corpus-grounded response is to add shot-index-in-narrative as a continuous covariate in the group GLM and to fit a within-act stratified analysis as a named follow-up [Magliano2011TheIO; zacks2007event; baldassano2017event; chen2017shared]. Both are tractable from the existing shot-event annotation without new behavioural coding. - -### 7.5 Open questions and limitations +### 7.4 Open questions and limitations -The corpus is honest about what it cannot say. The Hickok-style mu-system critique is not represented in our cards, which weakens the action-perspective prediction. The emotion literature is predominantly adult, and the three pet-evoked affective cards are fMRI or behavioural rather than EEG. Frontal asymmetry at sub-second timescales is unprecedented and reliability-limited. The single-stimulus design forbids generalisation beyond *The Present*. The 100 Hz local working set caps any beta-band and gamma-band claims until the 500 Hz validation pass. The HBN cohort is psychiatrically heterogeneous; stratified analyses (autism-spectrum, attention, social skill) are exploratory follow-ups, not primary tests. The Outstanding Questions Box collects the forward-looking adjudication targets. +Narrative position is a within-stimulus confound. Boy-only and puppy-only shots in *The Present* differ on three-act position: boy-only clusters in the early-act setup, puppy-only in the late-act resolution. Any boy-vs-puppy ERSP difference may therefore be confounded with prediction-error or arousal trajectories. The response is to add shot-index-in-narrative as a continuous covariate in the group GLM and to fit a within-act stratified analysis as a named follow-up [Magliano2011TheIO; baldassano2017event; chen2017shared]; both are tractable from the existing shot-event annotation. Beyond narrative position, several gaps in the corpus limit what this Review can claim. The Hickok-style mu-system critique is not represented in our cards, which weakens the action prediction. Klin and colleagues showed that toddlers with autism orient to audiovisual contingency rather than upright biological motion [klin2009biological] and that adolescents with autism fixate eyes 50 percent as often during emotionally evocative viewing [klin2002visual]; the HBN cohort includes a substantial autism-spectrum subsample, so autism status is a candidate moderator, but stratified analyses (autism-spectrum, attention, social skill) are exploratory follow-ups rather than primary tests. The emotion literature is predominantly adult; the three pet-evoked affective cards are fMRI or behavioural, not EEG. Frontal asymmetry at sub-second timescales is unprecedented and reliability-limited. The single-stimulus design forbids generalisation beyond *The Present*. The 100 Hz local working set caps beta-band and gamma-band claims until the 500 Hz validation pass. The Outstanding Questions Box collects the forward-looking adjudication targets. ## Box 1: HBN-EEG Release 3 as the anchor cohort @@ -133,7 +131,7 @@ Recent advances make the per-shot framing newly tractable. 2. Is mu-band ERD over central rolandic clusters elicited by animated-character action observation, as it is by hand-action observation in adults? 3. Does cuteness-driven affective response in children produce a sub-second EEG signature distinguishable from generic arousal in the alpha band, and is the signature compatible with frontal asymmetry at sub-second timescales given the meta-analytic reliability concerns? 4. Can a topographic-and-band rejection region for the four-perspective ranking be pre-registered before group analysis, and is the central-rolandic-versus-frontal-asymmetric-versus-language-network discrimination operationalisable from EEG IC clusters? -5. Can a vision-side multimodal embedding regressor substitute for the language-model-surprisal regressor framework on silent stimuli, and what method bridges the syntactic and semantic granularity of the language-model framework into vision? +5. Can a multimodal vision-language embedding regressor substitute for language-model surprisal on silent stimuli? 6. Does within-stimulus narrative position (three-act trajectory) explain condition-level effects that survive low-level partialling in single-stimulus designs, and how should shot-index-in-narrative be operationalised as a regressor? 7. What is the residual saccade-locked variance contamination in shot-onset EEG ERSP without a synchronous eye tracker, and at what cohort size does ICA-only artifact rejection become sufficient? @@ -163,7 +161,7 @@ Recent advances make the per-shot framing newly tractable. **Baby schema.** A set of infantile physical features (large head, large eyes, round cheeks) that elicit attentional, affective, and caregiving responses. -**Naturalistic stimulus.** A continuous, ecologically valid stimulus (typically a film, audiobook, or video game) presented without trial-by-trial structuring. +**Naturalistic stimulus.** A continuous, ecologically valid stimulus (typically a film, audiobook, or video game) presented without trial-by-trial structuring. Naturalness is a continuum from controlled gratings to live-action film, with character animation and abstract Heider-Simmel triangles as intermediate points (Figure 2). **Event segmentation.** The cognitive process of parsing continuous experience into discrete events at moments of high prediction error, organised hierarchically. @@ -175,7 +173,7 @@ Recent advances make the per-shot framing newly tractable. **Figure 2. Naturalness gradient and developmental cohort coverage.** Stimulus naturalness on the x-axis (controlled gratings, static photographs, Heider-Simmel triangles, abstract animation, character animation, live-action film) versus participant cohort on the y-axis (adult, adolescent, child). Markers are sized by number of corpus cards and shaped and coloured by modality (fMRI as circle, EEG as square, MEG as triangle, intracranial EEG as diamond; behavioural-only entries as the letter b). The dashed yellow rectangle at (child, character animation) marks the target cell for per-shot EEG ERSP at the 0 to 500 ms window: existing coverage is whole-clip ISC, not per-shot ERSP. -**Figure 3. Gap matrix.** Eight named gaps from the four-strand corpus (rows) versus four prior-effort axes (cinematic fMRI, naturalistic scalp EEG, intracranial and MEG, behavioural and eye-tracking; columns). Filled cells list a representative card slug; cells marked "no coverage" with a vermillion dashed border indicate uncovered combinations. Twelve cells across the eight rows show no coverage in at least one column, defining the design space for per-shot developmental EEG ERSP. +**Figure 3. Gap matrix.** Eight named gaps from the four-strand corpus (rows) versus four prior-effort axes (cinematic fMRI, naturalistic scalp EEG, intracranial and MEG, behavioural and eye-tracking; columns). Filled cells list a representative card slug; cells marked "no coverage" with a vermillion dashed border indicate uncovered combinations. Thirteen cells across the eight rows carry no coverage, defining the design space for per-shot developmental EEG ERSP. **Figure 4. Predictions and falsification regions, per perspective.** Each perspective (row) is named with its predicted topography (with a head schematic showing the topographic focus), band, latency, and pre-registered falsification region. Psychophysics is the covariate, not the prediction. Action predicts central-rolandic mu-band (8 to 13 Hz) ERD over electrodes C3, Cz, and C4, with possible beta rebound (15 to 25 Hz). Language predicts no signal locally; a surviving cluster in left-frontotemporal IC space (Lipkin atlas negative-control mask) falsifies the four-perspective ranking. Emotion predicts early occipital alpha desynchronisation (80 to 300 ms) and later frontal F3/F4 asymmetry (200 to 500 ms), at incompatible latencies and topographies. The cluster-level alpha for falsification is p < 0.05 corrected by mass-univariate permutation. diff --git a/manuscript/narrative-review/reviews/humanizer-log.md b/manuscript/narrative-review/reviews/humanizer-log.md new file mode 100644 index 0000000..9b3a87d --- /dev/null +++ b/manuscript/narrative-review/reviews/humanizer-log.md @@ -0,0 +1,68 @@ +# Humanizer pass log + +**Skill:** `manuscript:humanizer` (research-skills marketplace, v0.5.0) +**Date:** 2026-05-20 +**Phase:** 4 (paper-review + humanizer) + +## Detection summary + +Scanned the full manuscript (sections 1-7, Box 1, Trends Box, Outstanding Questions, Glossary, figure legends) for the 29 humanizer patterns. The manuscript already showed strong baseline humanizer discipline because the project style guide enforces several patterns upfront: + +| Pattern | Status | Notes | +|---|---|---| +| 14: em-dash overuse | clean | Project rule "no em-dashes"; grep for `—` and `--` returned no body matches | +| 18: emojis | clean | Project rule "no emojis"; none present | +| 17: title case headings | clean | Sentence-case throughout (matches Cell Press house style for body headings) | +| 19: curly quotes | clean | Straight quotes only | +| 7: AI vocabulary (delve, intricate, tapestry, underscore, pivotal, vital, crucial filler, showcase, boast) | clean | grep returned 0 matches in body; "robust" appears once in a scientific context (alpha desynchronisation as "robust correlate") and was kept | +| 8: copula avoidance | clean | grep for "stands as / serves as / functions as / represents a" returned 0 body matches | +| 23: filler phrases | clean | No "it is important to note", "in order to", "due to the fact that" | +| 24: excessive hedging | clean | Single hedges in Discussion (legitimate); no stacked "could potentially possibly" | +| 25: generic positive conclusions | clean | No "opens exciting new avenues" or "the future is bright" | +| 20: collaborative artifacts | clean | No "I hope this helps" | +| 22: sycophantic tone | clean | No "great question" patterns | + +Two minor patterns were detected and rewritten. + +## Patterns detected and fixes applied + +### Pattern 3 (superficial -ing endings) — minor instance in Section 5 + +**Detected:** Section 5b closing paragraph had "Their independent-component-cluster analogues in EEG are the search regions the per-shot ERSP analysis targets" with the implicit -ing in "targets" pulling a verb-as-rhetorical-flourish. + +**Fix:** Tightened to "...are the search regions for the per-shot ERSP analysis", removing the verb-as-action framing. + +### Pattern 25 (anthropomorphic / soft assertions) — Section 7.4 + +**Detected:** "The corpus is honest about what it cannot say." This is an anthropomorphic framing (corpora are not honest); softer than the rest of the section. + +**Fix:** Rewrote to "Several gaps in the corpus limit what this Review can claim." Direct, claim-bounded, and not anthropomorphic. + +### Pattern 11 light touch (synonym cycling) — Section 5 + +**Detected:** "The language perspective therefore enters our review twice" uses "enters" as a quasi-metaphor that could read as filler. + +**Fix:** Rewrote to "The language perspective therefore plays two roles" plus "The 5a sub-thread isolates..." and "The 5b sub-thread supplies..." giving each sub-thread an explicit subject and verb. + +## Patterns explicitly preserved per research-writing calibration + +- Passive voice in methods-adjacent claims (Box 1 pipeline description): preserved. +- Single hedges in Section 7.5 ("may", "suggests", "is consistent with"): kept. +- Established compound modifiers ("event-related spectral perturbation", "log luminance ratio", "default-mode network"): hyphenation preserved per scientific style. +- Selective bold in Trends Box bullet leads ("**Whole-brain shot-cut response in adult intracranial EEG.**"): preserved per pattern 15 calibration (selective bold for callout headers is fine). +- Cite-card slug citations in `[KeyYYYY...]` form: preserved per task instructions; Cell Press numbered conversion happens in Phase 5. + +## Net effect on word count + +Before humanizer pass (after M1-M4 paper-review fixes): +- Main text sections 1-7 total: 3255 words + +After humanizer pass: +- Main text sections 1-7 total: 3249 words (six words trimmed via pattern 3 and pattern 25 fixes) + +## Carry-forward to Phase 5 + +- Cell Press numbered reference conversion (Phase 5). +- Final TiCS-specific formatting (Phase 5). +- Font-size + DPI fixes on figures (Phase 5). +- Stimulus thumbnails (Fig 2) + brain icons (Fig 4) via figures:transparent-icons (Phase 5b). diff --git a/manuscript/narrative-review/reviews/internal-review.md b/manuscript/narrative-review/reviews/internal-review.md new file mode 100644 index 0000000..23f0ad3 --- /dev/null +++ b/manuscript/narrative-review/reviews/internal-review.md @@ -0,0 +1,141 @@ +# Internal peer review of `manuscript.md` + +**Reviewer:** internal pass via `manuscript:paper-review` (Phase 4 of epic #46). +**Target:** *Trends in Cognitive Sciences* Forum Review. +**Date:** 2026-05-20. + +## Synopsis + +This Forum Review argues that per-shot EEG event-related spectral perturbation (ERSP) in a developmental cohort viewing silent character animation is an empty cell at the intersection of four research traditions (psychophysics, action, language, emotion). The four-perspective scaffold makes the argument operationally tractable, naming a regressor of no interest (psychophysics), a band-and-topography prediction with adult precedent (action), a method that structurally cannot transfer plus a positive sub-thread (language), and two latency-distinct predictions (emotion). Section 7 names a topographic-and-band rejection region that a pre-registered group analysis can adopt before opening the data, anchored externally on Petroni-Cohen 2018 and internally on a partly-validated EEGLAB-style pipeline (Box 1). The corpus underlying the synthesis is 94 cite-card-backed references from a four-strand collection. Headline assessment: this is a publishable Forum Review draft. The four-perspective scaffold is coherent; F1-F5 carry-forwards from the prior self-review are addressed in prose; the falsifiability section is concretely operationalisable. The most consequential remaining issues are (1) a quantitative caption error in Figure 3, (2) Section 2 still reads more as a theme catalogue than an integrated argument, (3) Section 7 runs ~175 words over its budget while Sections 1-2 sit ~390 words under, and (4) the abstract omits the 100 Hz sampling-rate constraint which is a load-bearing methodological qualifier in the body. + +## Critical Issues + +None. There are no methodological flaws, invalid statistics, or unsupported claims that would block submission of this Review. Notes below are major or minor concerns. + +## Major Concerns + +### M1. Figure 3 caption count is wrong: 13 no-coverage cells, not 12 + +**Where**: Figure 3 caption (line 178); referenced from Section 5b body and from the gap-matrix figure source SVG. + +**Issue**: The caption states "twelve cells across the eight rows show no coverage in at least one column". Counting the no-coverage cells in `fig3_gap-matrix.svg`: Gap 1 (1), Gap 2 (1), Gap 3 (2), Gap 4 (1), Gap 5 (1), Gap 6 (2), Gap 7 (2), Gap 8 (3) = **13**. A peer reviewer who counts the dashed rectangles will find the mismatch immediately. This is the same class of factual error that the prior self-review flagged as F1 (Category G cardinality 12 vs 7), so addressing it preserves the rigor-checklist discipline. + +**Fix**: Change "twelve" to "thirteen" in the Figure 3 caption. Verify against the SVG one more time before Phase 5. + +### M2. Section 2 still reads as a theme catalogue rather than an argument + +**Where**: Section 2 ("The four-perspective scaffold"), middle paragraph (line 49). + +**Issue**: The prior self-review flagged I4 (Section 2 breaks the narrative arc). The current draft adds the bridge sentence at the open and forward-reference at the close, both of which are improvements. The middle paragraph still enumerates 15 themes back-to-back with a perspective tag per theme, which reads as an inventory rather than an argument. The 334-word section also sits under its 600-word budget; there is headroom to develop the perspective interactions more substantively. Compared to Section 4 (action), where multiple "beats" structure the argument, Section 2 has no structural rhythm beyond the enumeration. + +**Fix**: Restructure the middle paragraph around the four perspectives, not the 15 themes. One paragraph per perspective summarising which themes carry the perspective's prediction, briefly. Or compress the theme enumeration to two sentences and use the freed budget to elaborate the four-perspective interaction (when do action and emotion compete, when do they reinforce, what is the relationship between language 5b and emotion social-cognition). Either path raises Section 2 from list to argument. + +### M3. Section 7 over-runs its budget by ~175 words while Sections 1-2 sit ~390 words under + +**Where**: Sections 1 (326 vs 450 target), 2 (334 vs 600 target), 7 (624 vs 450 target). + +**Issue**: The body-text rebalancing in the PR description ("Sections 1-2 are under target by ~390 words combined; Section 7 is over by ~175") is true to the budgets but reflects a deeper structural skew. Section 7 is doing five sub-jobs (integration, anchor case, falsifiability, narrative-position objection, open questions and limitations) inside one section, while Sections 1 and 2 are doing under what their budgets would support. A peer reviewer will notice that the synthesis is heavier than the introduction, which inverts the usual Forum Review weighting. + +**Fix**: Move Section 7.5 (open questions and limitations) content partly to Section 1 (limitations preview) and partly to the Outstanding Questions Box (where it already lives, redundantly). Move Section 7.4 (narrative-position objection) tightened to ~60 words inside 7.3 falsifiability, since it is in fact a falsification-region concern. Use the freed budget in Section 2 to do the perspective-interaction work flagged in M2. + +### M4. Abstract omits the 100 Hz constraint, which the body names as load-bearing + +**Where**: Abstract (lines 33-35). + +**Issue**: The abstract claims that per-shot ERSP in a developmental cohort has no published precedent and that HBN-EEG R3 sits at the empty intersection. The body (Section 7.5; Box 1) names the 100 Hz local working set as a constraint that caps beta-band and gamma-band claims until the 500 Hz validation pass. A peer reviewer reading only the abstract would not learn that the empirical follow-on test is sampling-rate-limited at the dev tier, which a TiCS abstract should disclose at the level of one clause. + +**Fix**: Add one clause to the abstract acknowledging the sampling-rate constraint: "...sits at this empty intersection, with the local 100 Hz working set capping beta-band claims until a 500 Hz validation pass." This makes the abstract honest at the abstract level. Word budget: the abstract is currently 125 words; this adds ~15 words, putting it ~10 words over its 110-120 target. Offset by tightening the sentence on "Most empirical evidence to date comes from...". + +## Minor Concerns + +### m1. Abstract is 125 words; TiCS target is ~80-120 + +**Where**: Abstract (line 35). + +**Issue**: 125 words is 5 over the upper bound of the 110-120 target the scaffold set, and TiCS-published abstracts cluster at 80-120. Combined with M4 (add sampling-rate clause), aim for 110-120 net. + +**Fix**: Trim "in two waves" filler and "Most empirical evidence to date" to one phrase each. Drop "We review the four-strand corpus that constrains this design space" in favour of "We review the corpus". + +### m2. "stop-motion or stills" misnames Dorr 2010 + +**Where**: Section 3, paragraph 3 (line 59). + +**Issue**: The sentence "Gaze coherence varies with stimulus class, highest on Hollywood trailers and lowest on stop-motion or stills" misrepresents Dorr and colleagues 2010, which compared Hollywood trailers, natural movie clips, and static images. Stop-motion is not a category in that study; stills is. + +**Fix**: Replace "stop-motion or stills" with "natural movie clips and static images". + +### m3. Section 4 paragraph 1 drops the Hickok critique in a single sentence + +**Where**: Section 4, paragraph 1 (line 63), last sentence. + +**Issue**: The Hickok-style mu-system critique is mentioned in a single half-sentence ("...not represented as cards in our corpus and which temper the weight that the action-perspective prediction can carry"). The prior self-review (I3) flagged this as a Phase 1 grounding gap: the corpus does not contain the strongest steelman objection to the deepest specific prediction. Treating the critique as a one-sentence hedge under-weights it. + +**Fix**: Either add a sentence elaborating the substance of the Hickok objection (mu suppression also reflects general attention to motion, not a one-to-one mirror-system signature) so the reader knows what the critique actually says, or commit to a follow-up paragraph in Section 7.5 (limitations) that explicitly names the corpus gap. The current treatment is honest but thin. + +### m4. "This Review" capitalisation is inconsistent with Cell Press house style + +**Where**: Section 1, last paragraph (line 43): "This Review argues that..."; Section 5b last paragraph (line 83): "The language perspective therefore enters our Review twice." + +**Issue**: Cell Press style typically uses lowercase "review" inside the article body and capitalises "Review" only when referring to the article type as a noun in metadata. The current capitalisation is inconsistent across sections. + +**Fix**: Lowercase throughout the body. Reserve "Review" capitalisation for the YAML frontmatter / article-type-designation context. + +### m5. Section 5b lacks a clear thesis sentence to mirror 5a + +**Where**: Section 5b (lines 81-83). + +**Issue**: Subsection 5a opens with "The contemporary methodological mainstream...", a clear thesis. Subsection 5b opens with "A second strand of language-strand cards documents what silent narrative engages independent of speech", which is descriptive rather than thesis-asserting. The 5a-vs-5b split is structurally valuable (per I5 carry-forward), but the asymmetry in opening rhetoric makes 5b read like a continuation rather than a counterpart. + +**Fix**: Open 5b with "Silent-narrative neural correlates do transfer to scalp-EEG ERSP analysis even when language-model regressors cannot." Then proceed with Castelli, Vanderwal, Naci, Lankinen, Schroeder, default-mode-network synthesis. + +### m6. Glossary entry "Naturalistic stimulus" risks being too narrow + +**Where**: Glossary (line 166). + +**Issue**: The definition reads "A continuous, ecologically valid stimulus (typically a film, audiobook, or video game) presented without trial-by-trial structuring." A TiCS reader from a non-cinematic-fMRI background may not understand why audiobooks and video games are listed alongside film. The naturalness-gradient framing of Section 3 and Figure 2 implies that the stimulus class is broader than examples given. + +**Fix**: Add one sentence to the glossary entry: "Naturalness is a continuum from controlled gratings to live-action film, with character animation and abstract Heider-Simmel triangles as intermediate points (Figure 2)." + +### m7. Outstanding Questions Box question 5 is dense + +**Where**: Outstanding Questions Box, question 5 (line 136). + +**Issue**: Question 5 ("Can a vision-side multimodal embedding regressor substitute for the language-model-surprisal regressor framework on silent stimuli, and what method bridges the syntactic and semantic granularity of the language-model framework into vision?") packs two distinct questions into one and asks the reader to track "vision-side multimodal embedding regressor" and "language-model-surprisal regressor framework" in the same sentence. The compound structure dilutes the question's force. + +**Fix**: Split into two questions or shorten to "Can a multimodal vision-language embedding regressor substitute for language-model surprisal on silent stimuli?" Drop the second clause; the bridging question is a follow-up rather than a forward-looking adjudication target. + +### m8. Section 4 paragraph 4 abruptly introduces the autism subsample + +**Where**: Section 4, last paragraph (line 69), starting "A third action beat concerns single-agent versus two-agent shots". + +**Issue**: This paragraph compresses three different topics: macaque two-agent network (Sliwa), single-agent contrast design rationale, and the autism-spectrum subsample of HBN. The autism topic appears here in two sentences without prior setup. It belongs either with Section 7.5 (limitations as moderator) or as its own short paragraph. + +**Fix**: Move the autism sentences to Section 7.5 limitations. Section 4 paragraph 4 then reads as a tight macaque-to-design-rationale argument. + +### m9. Figure 4 caption mentions "occipital alpha desynchronisation" without naming the latency window in the text + +**Where**: Figure 4 caption (line 180); Section 6 body (line 87). + +**Issue**: Figure 4 caption gives "80 to 300 ms (occipital)" and "200 to 500 ms (frontal)". Section 6 body does not give explicit latency windows. A reader inspecting the figure first and then the body will notice the latency-window precision in the figure that is absent in the body. + +**Fix**: Add the latency windows to Section 6 body: "early occipital alpha desynchronisation (80 to 300 ms post-shot-onset)" and "later frontal-asymmetric alpha (200 to 500 ms)". Citation precedent for the windows: the Codispoti review for occipital, and the Davidson tradition (seconds-to-minutes) extrapolated downward for frontal. The body should be explicit about the extrapolation since the frontal latency is not directly precedented. + +### m10. "ECoG" is introduced inside a parenthetical without being defined + +**Where**: Section 5a (line 75): "shared between word-by-word intracranial cortical recording (ECoG) and autoregressive LMs". + +**Issue**: ECoG (electrocorticography) is introduced as an apposition to "intracranial cortical recording" but the acronym is not actually defined here; the parenthetical reads as defining "intracranial cortical recording" with the acronym ECoG, but ECoG specifically means electrocorticography. A non-EEG reader (TiCS audience is cognitive-science-broad) may be confused. + +**Fix**: Either expand to "electrocorticography (ECoG)" or remove the acronym since it is used only once more (Section 5a, in the Schrimpf citation). Defining-on-first-use convention from the project style guide. + +## Editor Note + +Recommendation: minor revision (TiCS terminology). The four-perspective scaffold and the falsifiability operationalisation are publication-ready in argumentative substance. The remaining work is rebalancing (Sections 1, 2, 7 word counts), one factual caption error (M1), and polish (minor concerns). All flagged items are addressable in a single Phase 4 revision pass without restructuring the argument. The 5a/5b split (per prior self-review I5) is well-executed and the falsifiability rejection region (per prior self-review F5) is concrete enough to pre-register before group analysis. + +## Carry-forward to Phase 5 + +- M1 (caption count): correct before Phase 5 assembly. +- M2 (Section 2 restructure): apply in Phase 4 humanizer / copy-edit pass. +- M3 (word-count rebalance): apply in Phase 4 copy-edit pass. +- M4 (abstract sampling-rate clause): apply in Phase 4 copy-edit pass. +- m1-m10: apply opportunistically in Phase 4 humanizer pass; verify in Phase 5 final assembly.