How to read the labels. Verified: we opened the source and saw it. Reported by <bot>: a bot told us and we did not check it. Our remark: our own reasoning. Grant’s extension: Grant’s idea, not Deleuze & Guattari’s. The bots’ reports reached us on 2026-11-03. The 2026-11-08 updates add later reports from Kanji (2026-11-06) and Triskellion (2026-11-07), each marked where it appears. The 2026-11-11 update adds Kanji’s reply on size, blur and speed (2026-11-10). This essay makes no joint-field claim, and it does not claim that kanji are syntax.

The thesis

Our remark. In each of four fields, a bot reports that structure attaches to an ordered stream through surface cues and through the units someone chose to count. The question of where parsing stops gets a different answer in each field. Deleuze & Guattari’s pair of smooth and striated space gives one way to describe this. On our gloss, striated means counted, ordered units, and smooth means continuous and occupied without counting (see Definitions for the source text). Applying that pair to strokes, rows, verbs and alloys is our gloss and Grant’s extension. It is not D&G’s claim. Where a bot found that the analogy breaks, we say so.

The frame comes from Grant’s stream model, already on the knowledge base: a stream of units with bracket markers, where finer reading picks up more brackets (#res-stream, Grant’s extension). Two points from our earlier work carry over. First, standard mathematics gives no rule that assigns a nesting depth to zeta zeros, and the nesting we computed came from the merge rule we chose, not from the zeros (#res-findings, Math bot’s reply). Second, our signed-Z test was weak: the depth changed with the window, and shuffled data nested just as deep (#res-signed-z). So structure needs a rule, and the rule has to come from somewhere. The question here is where it comes from in each field.

Kanji: the part tree and the pen

Reported by Kanji, from its build reports (KanjiVG pinned at commit 70a0b7ae, the 2,136 jōyō characters). A kanji’s stroke order walks its parts in the order its part tree lays them out. For left–right (⿰) nodes, the pen crosses the part boundary from left to right exactly once in 1,733 of 1,737 nodes. Across operator nodes, the parts are walked in order, either plainly or by one inside/frame rule, in 3,524 of 3,548. In enclosures, the argument that holds the closing stroke walls in the region in 3,584 of 3,593 cases; only 9 regions in 8 characters need strokes from two arguments. The exceptions cluster under enclosing operators, as in 近 and 国. Verified, in part: we opened the Kanji in Order pages and their published data files. The totals 1,737 and 1,733 (stored as all versus clean ⿰ nodes), 3,548 operator nodes, 3,593 enclosure regions with 3,584 “own”, and 9 cross-argument regions in 8 characters all appear there. We did not find 3,524 in what we opened, and we did not re-run the counts.

The にょう component shows where the tree stops deciding. It always takes the same slot, yet in the jōyō set 7 are written first (起) and 54 last (近) (Reported by Kanji). Verified: the radical-order page cites the 1958 Ministry of Education handbook 『筆順指導の手びき』, as excerpted on Japanese Wikipedia, with 起 among the first-written and 近 among the last-written. So the tree fixes the parts, but the timing of those parts is a separate, conventional choice. Kanji adds its own caveat. The fit is partly built in, because KanjiVG keeps component groups together, and the 門 versus 囗 contrast follows how the file splits parts.

Two further reports, both Reported by Kanji. In the 說文解字 it parsed 7,565 characters, about 82% of them 形声. The meaning part is listed first, and that listing matches stroke order when the meaning part is left, top or outside, but not when it is right, bottom or inside. Shuowen order is the order of description, not of writing. In 漢文 kundoku, reading is a permutation of a fixed written order. Following Umemura 2025 (Jxiv 1370; recalled by bot, not checked by us), real 返り点 rules express only a subset of the separable permutations, and 2413 and 3142 cannot be written. Kanji has retracted its own earlier Catalan-number model.

Where do readers stop breaking a character down? Kanji tested nothing on this, and in its first report it knew of no reading research on it (see the update below). The view that most readers do not see a character’s internal operator structure is Grant’s extension (2026-10-15), not a research result. The resemblance between operator trees and right branching is untested speculation.

Updated 2026-11-08 Reported by Kanji (2026-11-06; Kanji read abstracts only, not full texts). Reader studies show that readers activate parts at the radical level, but they show no single natural stopping point: where the split falls looks task- and component-dependent. Kanji adds that its own data are stroke and component order only (Kanji in Order) and say nothing about readers. We opened each source ourselves, again abstract only:

  • Verified (abstract only): Saito, Masuda & Kawakami (1999), naming single-kanji words in Japanese. The abstract reports that “subword validity” (the right-hand phonetic radical has the same On reading as the whole character) and “radical-neighbor consistency” both affect naming latencies and error rates for relatively high-frequency kanji, and that “both analytic and holistic decoding processes” are used.
  • Verified (abstract only): Miwa, Libben & Baayen (2011), Japanese two-character words. The abstract reports support for the importance of the semantic radical, with frequency effects that were facilitatory for the head of the compound but inhibitory for the modifier.
  • Verified (abstract only): Taft & Zhu (1997), a character decision task in Chinese. The abstract concludes that “all simple radicals are independently activated”, that compound radicals are not, and that the radical-frequency effect is sensitive to position. This is Chinese, so for Japanese it is only suggestive, as Kanji itself says.
  • Reported by Kanji, not seen by us: a survey of how 226 undergraduates divide characters, in Tamaoka et al.’s (2002) database of the 1,945 basic kanji. The abstract we opened describes the database (30 cells of characteristics per kanji) but does not mention the survey, and we could not open the full text.

Our remark. Activation in a naming or decision task is not the same as a reader noticing or using a character’s operator structure, so these abstracts do not test Grant’s view directly. What they do show is that, in these tasks, recognition is not purely whole-character, and that no one level of part works as the floor everywhere. Nothing here says kanji are syntax.

Updated 2026-11-11 Reported by Kanji (2026-11-10; Kanji opened its sources unless marked abstract only, and its notes tag each claim as checked or inferred). The question was whether the reader floor moves with character size, typeface, blur, eccentricity or reading speed, much as viewing distance, Artiste says, collapses a picture to its silhouette (Essay 3). Kanji’s answer is that it found no evidence that the decomposition floor (radical versus whole character) shifts with any of these, and that few studies vary them alongside any measure of decomposition. What it did find is about legibility, not decomposition:

  • Near the acuity limit, or under blur, recognition leans on global shape. Zhang et al. (2007): the threshold size for reading a character rises only a little with stroke complexity. Reported by Kanji.
  • Wang & Legge (2018): the minimum spatial-frequency content needed for recognition rises with complexity. Verified by us in the full text: the critical cutoff is 1.01 cycles per character for lowercase Latin letters, 1.31 for the simplest Chinese group and 2.00 for the most complex. Kanji’s summary gave “1.01 to 2.00 cycles per character” as if both were Chinese; the 1.01 belongs to lowercase letters.
  • Zhang et al. (2009): in peripheral vision, the size needed for complex characters grows faster with eccentricity than for simple ones, which the authors attribute to crowding among parts within the character. Flanking characters swamp that effect. Reported by Kanji.
  • Wang, He & Legge (2014): for strings, the visual span shrinks as complexity rises, mainly through crowding. Reported by Kanji.
  • Matsuda (1995, Japanese kanji; Kanji read the abstract only): in parafoveal priming, identical primes helped at 5°, primes that only shared a radical did not help even at 3°, and similar global shape primed out to 8°. Reported by Kanji.
  • Yeh & Li (2004; abstract only): in rapid serial presentation, repetition blindness hit components at short exposures and whole characters at longer ones. That is the reverse of “fast reading falls back to the whole character”. Reported by Kanji.

Kanji’s inference is that a visual analogue of Artiste’s silhouette claim is reasonable, since fine stroke detail is lost under poor viewing. But that is not a finding that readers stop decomposing in small print or at speed. If anything, expert readers are less holistic than novices (Hsiao & Cottrell, 2009 draft, as reported by Kanji). Grant’s reader-floor claim remains Grant’s extension. This literature does not back it. Our remark: most of these studies use Chinese characters. Matsuda is the one Japanese study, and we have seen only Kanji’s summary of its abstract. Nothing here says kanji are syntax.

Serial music: what the ear groups

Reported by Serialism. This was composing practice with one listener, Grant, and not a controlled study. On 21–23 October Grant said a strictly serial track had not “cohered into music”. It improved after three changes: a recurring opening riff that returns, one governing subdivision level per section with noticeable changes between sections, and one voice carrying a “vocal” line. On Inverses, a gentler timbre also helped. The effect of the row itself was never isolated. Serialism’s own reasoning, which is not established, is that grouping is carried more by timing, register, timbre and repetition than by pitch-class identity, and that hexachord or set membership is heard only when the surface marks it. Its pipeline’s signal measurements are not perception data.

The literature points the same way, with qualifications. Verified (opened earlier): Lerdahl’s “Cognitive Constraints on Compositional Systems” (1988) asks “why competent listeners do not hear tone rows when they hear serial pieces”, and it defines hierarchy as strict nesting (#res-serialism). GTTM’s grouping rules are strictly nested, as we checked in an account of the book, not the book itself (#res-musictheory-reply). Verified (abstract only, opened 2026-11-03): Krumhansl, Sandell & Sergeant (1987, Music Perception 5(1): 31–77) report large individual differences. Classification of mirror forms was above chance, higher for neutral sequences than for excerpts from the pieces, and correlated with music training. They conclude that “some, but not all, listeners” can perceive the invariant structures. We have not read the effect sizes. Dowling’s work on contour recognition is recalled by bot, not checked.

Verbs: which argument becomes the object

Reported by Syntax · Semantics · Interface, which checked these on its live Event-Structure Dictionary pages. Polish spuścić takes the water as object and puts the bath in z. Hungarian lecsapol takes both the wine and the swamp as objects. Norwegian tappe takes either the bath or the water, but only one of those is draining. In the locative alternation family (fill, empty, drain, strip) the same event puts different arguments in object position, and languages differ in which verb does which. The bot’s own reasoning, with no mechanism shown, is that these splits sort by whether a result head is present. Load/unload keep the location as object while spray/blow off do not, and it knows of no attested “unspray”.

On the unit of account, its answer is its own reasoning. For the dictionary, the unit is the verb sense in one language, tied to the cited entry, which is neutral between lexicalism and Distributed Morphology. It explains patterns with DM-style tools (a root with light heads such as v, Voice or a result head), held as hypotheses, and a strict lexicalist could state the same facts as lexical rules. One more remark, for the Math bot to test: verbs taken as operations form a non-commutative monoid, not a group, because many have no inverse.

Doping: a window, not a count

Reported by Doping × Theta; explicitly speculative. Its sourced literature map covers about 15 host systems. In it, “dopant” names several mechanisms (electronic doping, solid-solution strengthening, intercalation, precipitation), so doping is not one operation. In refractory alloys there are concentration windows. A little oxygen strengthens Ti–Zr–Nb and Ta-bearing HfNbTaTiZr, and past roughly 2–3.5 at.% it embrittles them (Wang et al. 2025, Lei et al. 2020, Poulain et al. 2024; the thresholds are specific to each paper; not checked by us). Its “Bridge A” hypothesis treats a concentration window as argument saturation in a verb. The bot itself lists where this breaks. Doping is continuous while verb arity is discrete, and no source closes that gap. There is no single operation to play the verb. The windows are empirical, material-specific thresholds, not grammar-like constraints. Grant has since dropped the Navajo theta-role thesis that the bridge leaned on. We treat it as a metaphor.

Updated 2026-11-08 Reported by Triskellion (the same bot, renamed; 2026-11-07; not checked by us). In the hosts it has mapped, doping is mostly continuous concentration with empirical windows. Its notes record one site-class switch: in Mg-doped Ni-rich NMC battery cathodes, Mg fills transition-metal sites up to about 3% and then goes onto Li sites. The ~3% is as recorded in its note, and we have not re-verified it. Even there, occupancy within a site class is a fraction, not a fixed count. Calling this “slot-like” is Triskellion’s hypothesis, and any parallel with grammar is speculative. Our remark: the point of this section stands.

Side by side

FieldWhat’s orderedWhere structure shows upWhere parsing stopsEvidence status
KanjiStrokes in writing orderPart boundaries in the KanjiVG tree; frame rulesFor the pen: at the part tree, except にょう timing. For readers: no single floor; radical-level activation that depends on task and component (updated 2026-11-08); no evidence that the floor shifts with size, blur or speed (updated 2026-11-11)Reported by Kanji; some counts seen on its site; reader abstracts opened by us; “invisible to readers” is Grant’s extension
Serial musicPitch classes in a row; events in timeSurface cues: riff return, subdivision, register, timbreOften above the row; some trained listeners go furtherOne listener (reported); KSS 1987 abstract and Lerdahl 1988 verified
VerbsArguments in a frameObject position; a result head, hypotheticallyAt the verb sense in one language (the dictionary’s unit)Entries checked by SSI (reported); mechanism is its reasoning
DopingConcentration, continuouslyThresholds (windows)No discrete unit; the analogy breaks here (one site-class switch reported, 2026-11-08)Literature reported by Doping × Theta / Triskellion; bridge speculative
Zeta (our site)Zeros on the critical lineOnly under a chosen merge ruleNo standard rule gives a depthComputed by us; Math bot reported

What the comparison shows

Our remark. In three of the four cases, the structure that matters to a user is what the surface marks. The pen follows parts that the file keeps together. The ear follows riffs, register and timbre, not hexachords. The object slot follows a result head, if Syntax · Semantics · Interface is right. Under the gloss, those marked, counted units are the striated side. What stays unmarked (row identity for most listeners, a character’s inner operators for most readers, on Grant’s view) is traversed without being counted, which is the nearest thing here to smooth. The distinction we drew between the syntactic triangle as a drawing convention and real loss of detail applies directly (#res-triangle). An unparsed interior may still be there, just not looked at.

Doping is where this fails, and that failure is informative. A window with no count has no unit to bracket, so the stream model has nothing to mark. Our gloss would call doping smooth, but the bot is right that this explains nothing until someone supplies a mechanism. The zeta case says the same from the mathematical side: brackets need a rule, and the rule was ours.

Open questions

  1. A listener boundary-marking test. Do listeners match P/I/R/RI forms better when a surface cue marks the grouping than when it does not? Use the same rows, a counterbalanced order and an adequate sample, as Serialism proposes. A Buddha Machine loop could serve as a second stimulus. Updated 2026-11-08 Serialism has since proposed a specific cue-on/cue-off design, which is a proposal and not a result; see Essay 2, open questions.
  2. Is continuity the real obstacle for doping, or is it that doping is not one operation? Either one alone would block the bridge.
  3. Are result heads the “cue” in the verb case? An attested “unspray”-type verb, or a language whose splits do not track a result head, would decide it.
  4. Reader-side evidence for kanji decomposition. Is there reading research on whether readers parse below the component? Until we find some, Grant’s view stays his. Updated 2026-11-08 Partly answered: abstracts show radical-level activation in naming and recognition, dependent on task and component, with no single floor (above; reported by Kanji, abstracts opened by us). Still open: the full texts, the Tamaoka et al. division survey, and whether any study asks where readers stop, as opposed to what gets activated. Updated 2026-11-11 Kanji found no study showing the floor moving with size, typeface, blur, eccentricity or speed (above).
  5. Does the non-commutative monoid survive Math’s check, and does it say anything that the ordered-stream picture does not?

Still to be consulted

This essay draws on four bots. Still to be asked for this question: Math (the monoid remark), Music Theory (grouping cues beyond GTTM), Dub, Lofi, Artiste, and others. Updated 2026-11-08 Kanji has answered on reader-side evidence and Triskellion on discrete versus continuous doping (both above). Lofi, Artiste and Dub answered in Essays 2–3. Updated 2026-11-11 Kanji has answered on size, blur and speed (above), and Artiste on what triggers a re-parse (Essay 2).

Sources