diff --git a/manuscript/narrative-review/reviews/humanizer-log.md b/manuscript/narrative-review/reviews/humanizer-log.md
new file mode 100644
index 0000000..7654f18
--- /dev/null
+++ b/manuscript/narrative-review/reviews/humanizer-log.md
@@ -0,0 +1,81 @@
+# Humanizer pass log
+
+**Skill:** `manuscript:humanizer` (research-skills marketplace, v0.5.0)
+**Date:** 2026-05-21 (v2 Phase 4)
+**Scope:** all per-section files under `manuscript/narrative-review/`
+
+## Method
+
+The skill loaded its 29-pattern catalogue plus research-writing calibrations. I scanned every per-section file via grep for the high-frequency patterns:
+
+- Pattern 1 (significance inflation): pivotal, evolving landscape, groundbreaking, deeply rooted, indelible mark, marking/shaping the, focal point
+- Pattern 3 (superficial -ing endings): showcasing, highlighting, underscoring, emphasising, emphasizing, fostering, encompassing, cultivating
+- Pattern 4 (promotional): vibrant, rich, nestled, in the heart of
+- Pattern 7 (AI vocabulary): delve, tapestry, intricate, showcase, boast, additionally, moreover, furthermore, notably
+- Pattern 8 (copula avoidance): stands as, serves as, functions as, boasts, represents a [novel/key/...]
+- Pattern 10 (rule of three): scanned visually
+- Pattern 14 (em-dashes): grep for U+2014
+- Pattern 18 (emojis): grep for emoji range
+- Pattern 19 (curly quotes): grep for U+201C/D
+- Pattern 23 (filler): in order to, due to the fact, it is worth noting, it is important to note
+- Pattern 24 (excess hedging): could potentially, may possibly, might potentially
+- Pattern 25 (generic positive conclusions): opens new avenues, paves the way, sets the stage, paradigm shift
+- Pattern 27 (persuasive tropes): at its core, the real question, fundamentally, the heart of the matter
+- Pattern 28 (signposting): let us dive, let us explore, here is what you need to know
+- Pattern 20 (chatbot artifacts): I hope this, let me know, Of course, Certainly, Great question
+
+## Findings
+
+| File | Patterns matched | Fixes applied | Notes |
+|---|---|---|---|
+| `highlights.md` | 0 | 0 | 5 bullets all <=80 chars; no AI patterns |
+| `abstract.md` | 0 | 0 | Clean |
+| `sections/01_introduction.md` | 0 (pattern 28 candidate flagged for Phase 5) | 0 | Last sentence "Section 2 begins with..." is meta-tutorial signposting per pattern 28; paper-review m5 already scheduled for Phase 5 final assembly, so deferred (per task instructions: do not silently rewrite Phase 5 items) |
+| `sections/02_scaffold.md` | 0 | 0 | Clean |
+| `sections/03_psychophysics.md` | 0 | 0 | Methods-adjacent passage in paragraph 3 uses passive voice ("The HBN-EEG Release 3 cohort carries no synchronous eye tracker, which means a per-shot analysis cannot deconvolve..."). Per skill calibration, passive in methods-adjacent technical descriptions is acceptable; not flagged |
+| `sections/04_action.md` | 1 (scientific use of "robust", not significance inflation) | 0 | "...which produce robust ones [Ben-Yakov2018TheHF]" describes hippocampal-response amplitude; citation-anchored scientific use, not pattern 1 |
+| `sections/05_language.md` | 0 | 0 | 5a + 5b structure clean |
+| `sections/06_emotion.md` | 1 (scientific use of "robust") | 0 | "alpha desynchronisation is a robust correlate of attentional engagement" describes the Codispoti review's conclusion; citation-anchored scientific use, not pattern 1 |
+| `sections/07_synthesis.md` | 0 | 0 | Clean. Single hedges in 7.4 ("may therefore be confounded", "is the response") are legitimate research-writing usage per skill calibration |
+| `boxes/box1_anchor.md` | 0 | 0 | Methods-style pipeline description; passive voice acceptable per skill calibration |
+| `boxes/trends.md` | 0 | 0 | Pattern 16 (inline-header vertical list with bold leads) used here, but each bullet is a substantive recent development with citation, not generic boilerplate. Per skill: selective bold for callout headers is acceptable |
+| `boxes/outstanding-questions.md` | 0 | 0 | 7 questions, each forward-looking and specific |
+| `glossary.md` | 0 | 0 | Definitions are technical and grounded; no AI inflation |
+| `figures.md` | 0 | 0 | Captions are descriptive, not promotional |
+
+## Per-section word-count delta
+
+Before pass: 3397 main-text words + 117 abstract + 60 highlights + 161 + 191 + 174 (boxes) + 417 glossary + 425 figure legends = 4942 words
+After pass: identical (no rewrites applied)
+
+## Why the pass was a no-op
+
+The manuscript was drafted in Phase 2 with the v1 self-review's F1-F5 and I1-I5 carry-forwards already enforcing discipline against the worst AI-writing tells (signposted Section 1 transition, falsifiability discipline in 7.3, narrative-position objection in 7.4, perspective-driven Section 2 over theme-catalogue, parallel thesis sentences in 5a/5b). Additionally, the project style guide ([CLAUDE.md](../../../CLAUDE.md)) bans em-dashes, emojis, and AI attribution upfront. The manuscript-writing skill's IMRAD Writing Principles (Clarity, Precision, Flow, Citations) further constrain word choice. Together, these prevent the patterns from entering the prose in the first place.
+
+The humanizer pass as a finishing step is therefore most useful in projects without an upstream style-guide layer; in this project it functions as an audit confirmation.
+
+## Pattern-28 deferral note
+
+The closing sentence of Section 1 paragraph 3 ("Section 2 begins with the four-perspective scaffold the rest of the review builds on") matches pattern 28 (signposting / meta-narration). The paper-review pass independently flagged this as concern **m5** for Phase 5 attention. Per the task instructions to preserve concerns scoped for Phase 5, this is not rewritten here.
+
+## Final anti-AI audit
+
+> "What makes the below so obviously AI generated?"
+
+Re-reading the assembled manuscript with that prompt in mind:
+
+- The prose is dense with citations (often 3-5 per sentence). Density is genre-appropriate for a TiCS Forum Review and not an AI tell.
+- Sentence rhythm varies (short topic sentences alternating with longer evidence sentences); not the uniform pacing typical of AI text.
+- No "key" / "crucial" / "vital" inflation.
+- Section transitions are explicit but not tutorial-script ("Section 2 begins with..." flagged above).
+- Single-paragraph perspective definitions read as an actual argumentative scaffold, not a "let me cover all the bases" enumeration.
+
+> "Now make it not obviously AI generated."
+
+No rewrites required. The manuscript reads as authored.
+
+## Carry-forward to Phase 5
+
+- Pattern 28 fix on Section 1 closing sentence (covered by paper-review m5).
+- All other m1-m6 paper-review items.
+- No humanizer-specific changes required.
diff --git a/manuscript/narrative-review/reviews/internal-review.md b/manuscript/narrative-review/reviews/internal-review.md
new file mode 100644
index 0000000..73a7bd4
--- /dev/null
+++ b/manuscript/narrative-review/reviews/internal-review.md
@@ -0,0 +1,84 @@
+# Internal peer review of v2 manuscript
+
+**Skill:** `manuscript:paper-review` (research-skills marketplace, v0.5.0)
+**Reviewer pass:** internal critique against the *Trends in Cognitive Sciences* (TiCS) Forum Review submission standard
+**Date:** 2026-05-21 (v2 Phase 4)
+**Source:** assembled from `manuscript/narrative-review/{highlights,abstract,sections/*,boxes/*,glossary,figures}.md` (3397 main-text words; 5619 words total)
+
+## 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 perspectives (psychophysics, action, language, emotion). The four-perspective scaffold is structurally coherent: psychophysics is the bottom-up covariate that must be partialled; action makes the most specific positive oscillatory prediction; language is structurally non-transferable for LM regressors but supplies positive silent-narrative correlates via its 5b sub-thread; emotion makes two latency-distinct predictions. Section 7 articulates a pre-registerable topographic-and-band rejection region (central-rolandic mu confirms action; frontal-asymmetric alpha confirms emotion; left-frontotemporal cluster falsifies; null at cluster-level alpha also falsifies via bottom-up exhaustion). The five F1-F5 critical findings from the prior v1 self-review are all addressed in prose. Headline assessment: **publishable Forum Review draft.** Zero critical issues. Six minor concerns are listed below for the Phase 5 copy-edit pass.
+
+## Critical Issues
+
+None. The manuscript has no methodological flaws, unsupported claims, or structural failures that would block submission.
+
+## Major Concerns
+
+None. The argument is internally consistent and the four-perspective scaffold cleanly maps to Sections 3-6 and to Figure 4's predictions table.
+
+## Minor Concerns
+
+### m1. Highlight #2 does not name the four perspectives
+
+**Where:** `highlights.md`, bullet 2.
+**Issue:** "Four perspectives diverge on per-shot EEG predictions of naturalistic film" does not name which four. A reader scanning Highlights without continuing to the abstract cannot identify the perspectives. Cell Press / TiCS Highlights typically name the key concepts so scanning readers can decide whether to read further.
+**Suggestion:** Either expand to "Psychophysics, action, language, and emotion diverge on per-shot EEG predictions" (66 chars; fits the <=85 limit) or drop "Four perspectives" and lead with "Psychophysics, action, language, emotion diverge on per-shot EEG predictions" (78 chars).
+
+### m2. Abstract pre-registration disclosure is opaque without body context
+
+**Where:** `abstract.md`, final clause.
+**Issue:** "...the 100 Hz local working set caps beta-band claims until a 500 Hz validation pass" assumes the reader already knows what the 100 Hz vs 500 Hz distinction means. In an abstract this disclosure is honest but cryptic.
+**Suggestion:** Either tighten to "we acknowledge a sampling-rate constraint that caps beta-band claims" (no numbers; gain in clarity, loss in specificity) or expand by one clause to make the band-limit consequence explicit: "...HBN-EEG R3 sits at this empty intersection, although the local 100 Hz cohort caps beta-band and gamma-band claims pending a 500 Hz validation."
+
+### m3. Section 4 Hickok-system critique is referenced but not characterised
+
+**Where:** Section 4 paragraph 1 (last sentence).
+**Issue:** "Hickok-style objections to one-to-one mirror-interpretations of mu suppression" names the critique tradition but does not say what the critique actually argues (i.e., that mu suppression also reflects general attention to motion, not a one-to-one mirror-system signature). A reader unfamiliar with the literature learns the critique exists but not its substance. The v1 self-review's I3 carry-forward asked for either elaboration or grounding via a corpus card; the current text does the former at minimum length.
+**Suggestion:** Add one clause naming the substance: "Hickok-style objections that mu suppression also reflects general attention to moving stimuli rather than a one-to-one mirror-system signature". Adds ~15 words; tempers the action prediction more honestly.
+
+### m4. Section 6 has four distinct beats; the fourth (social cognition) reads as rushed
+
+**Where:** Section 6 paragraphs 3 and 4.
+**Issue:** The section earns its 537-word budget by developing two predictions plus two EEG routes (early occipital alpha; later frontal asymmetry). The third paragraph adds a fourth beat (social cognition: Richardson-Saxe, Mar, Singer, Zaki-Ochsner, Nummenmaa, Schmaelzle-Grall, Kauttonen) in a single dense paragraph. Five citations in three sentences. Reads as a list rather than an argument.
+**Suggestion:** Either compress the social-cognition beat to "Section 7's interpretation refers back to..." style (move detail to Section 7), or expand into a fourth paragraph that explicitly links each citation to a per-shot prediction. The current placement reads as a citation dump.
+
+### m5. Section 1 paragraph 3 closing sentence is meta-narratorial
+
+**Where:** Section 1 paragraph 3, last sentence.
+**Issue:** "Section 2 begins with the four-perspective scaffold the rest of the review builds on" is a meta-tutorial signpost. TiCS readers do not need this; the next section heading does the work.
+**Suggestion:** Delete the last sentence. Paragraph 3 ends naturally at "Box 1 anchors the argument to the HBN-EEG Release 3 cohort viewing *The Present* (Pixar 2014)."
+
+### m6. Two-stage citation slug for Castelli (2000) is potentially confusing
+
+**Where:** Section 5b first sentence; Section 2 paragraph 1.
+**Issue:** The Castelli 2000 paper appears under two slugs in different places: `[Castelli2000MovementAM]` (line 102) and `[castelli2000heider]` (lines 49, 78). These are two distinct entries from two corpus strands (language and action) that point to the same primary paper. The v1 self-review's F3 stray-refs check applies. Phase 5 must verify both slugs resolve to the same DOI and consolidate to one BibTeX entry; if not, the duplication must be disclosed.
+**Suggestion:** Phase 5 dedupe pass on the cited-key list; flag any same-DOI duplicates.
+
+## Editor Note
+
+**Recommendation: minor revision** (TiCS terminology equivalent). The four-perspective scaffold is publication-ready in argumentative substance. The falsifiability operationalisation (Section 7.3) is concrete enough to pre-register a topographic-and-band rejection region before group analysis on EEG IC clusters, which is the load-bearing methodological contribution of the Review. The remaining work is six low-severity polish items (m1-m6), all addressable in a single Phase 5 copy-edit pass without restructuring the argument.
+
+**F1-F5 carry-forward audit:**
+
+| Carry-forward | Status in v2 | Location |
+|---|---|---|
+| F1 (Category G cardinality 12 -> 7 + 12 corpus-wide) | Verbatim phrasing present | sections/05_language.md line 98 (5a closing) |
+| F2 (Codispoti not Schubring) | Cited as "Codispoti and colleagues (2023)" | sections/06_emotion.md lines 115, 119 |
+| F4 (anchor case: Petroni + pipeline, not feature conjunction) | Both anchor assertions named as independent | sections/07_synthesis.md 7.2 + boxes/box1_anchor.md |
+| F5 (topographic-and-band rejection region) | Central-rolandic mu / frontal-asymmetric alpha / Lipkin negative control / null-at-alpha all named | sections/07_synthesis.md 7.3 |
+| I3 (Hickok critique) | Named but thin (see m3) | sections/04_action.md paragraph 1 |
+| I2 (narrative-position objection) | Named with covariate response | sections/07_synthesis.md 7.4 |
+| I4 (cohesion) | Bridge sentences added | sections/01_introduction.md, sections/02_scaffold.md |
+| I5 (5a + 5b split) | Clean split with parallel thesis sentences | sections/05_language.md |
+| m8 (Klin autism in limitations) | Moved to 7.4 | sections/07_synthesis.md |
+| m9 (explicit latency windows) | "80 to 300 ms" + "200 to 500 ms" in body | sections/06_emotion.md |
+
+All v1 critical-and-important carry-forwards are addressed in prose.
+
+## Carry-forward to Phase 4 humanizer pass
+
+The m1-m6 items above are scoped for Phase 5 (final assembly), not Phase 4 (humanizer). Phase 4 should:
+
+1. Run `manuscript:humanizer` on the per-section files to remove AI tells.
+2. Verify the m1-m6 items remain candidates for Phase 5 fixes (the humanizer pass should not silently rewrite them away).