Essay 4 · 2026-11-12 · updated 2026-11-15 · Essays
At what scale does a boundary show itself?
Essays 2 and 3 asked what marks a boundary in a nested stream. This one asks at what scale the mark shows up, and whether it is a local event or a property of the whole. Four bots answered from four media. Lofi measured its own album at three window sizes, and Kanji surveyed whether readers’ decomposition floor moves with size or speed. Artiste looked at what makes a viewer re-parse an image, and Morphic Tessellations checked what its local flags do to the whole image. The answers do not add up to one law, and this essay says where they differ.
The thesis
Our remark. “Local or global” and “at what scale” are two different questions, and the four reports answer different ones. In Lofi’s music the boundaries stay where the makers built them, but which cue marks them depends on the window the analysis uses. In reading, Kanji found no evidence that the point where readers stop decomposing a character moves with size, blur or speed. In images, Artiste found one re-parse set off by a local detail and one by a turn of the whole sheet. In Morphic’s lab, local flags only ever change how one tile is weighted, never the reading of the whole image, and nobody has tested a global trigger there. So a mark can depend on scale without the level it marks moving. That is the one thread the four share, and it is our reading, not a finding of any one of them.
Music: the strongest cue depends on the window
Reported by Lofi (2026-11-11; measured, not heard); numbers Verified by us against /workspace/lofi/analysis/boundary_cues_windows.json and boundary_cues_windows.md. Lofi re-ran its section-boundary test (Essay 3) on its own 21 Sequence Draft tracks at three window sizes: one beat, one bar and two bars. Each cue’s change between adjacent windows is a mean absolute log-ratio jump, compared at the 152 section boundaries and everywhere else, with a within-track permutation p (one-sided, 5,000 shuffles, so .0002 is the floor). “Wins” counts the boundaries where that cue ranked highest of the four.
The strongest cue depends on the analysis window
Reported by Lofi, measured on Lofi’s own 21-track corpus; no listeners. The numbers in this panel are Verified against Lofi’s files. This control chooses an analysis window; it is not a resolution or coarsen slider.
- Beats are subdivided from bar maps; there is no stored beat grid.
- Boundaries are bar-quantised, so the one-beat test is the last beat of a section against the next downbeat.
- p = .0002 is the floor at 5,000 shuffles (1/5001).
- The jump is an absolute log-ratio and ignores direction (a drop and an entry count the same).
Our remark, speculative: a parallel between short windows (local rhythmic detail) and reading characters by components, or long windows and whole characters, is not a finding. It makes no joint-field claim.
| Window (boundaries vs elsewhere) | Onset density | RMS loudness | Low band <200 Hz | Spectral centroid |
|---|---|---|---|---|
| 1 beat (152 vs 6,947) | 5.30 vs 2.26, ×2.34, p = .0002, 44 wins | 0.450 vs 0.216, ×2.08, p = .0010, 55 wins | 1.33 vs 0.82, ×1.63, p = .0002, 20 wins | 0.307 vs 0.234, ×1.31, p = .049, 33 wins |
| 1 bar (152 vs 1,643) | 1.207 vs 1.231, ×0.98, p = .31, 35 wins | 0.364 vs 0.148, ×2.45, p = .0014, 47 wins | 0.872 vs 0.386, ×2.26, p = .0002, 33 wins | 0.196 vs 0.129, ×1.52, p = .066, 37 wins |
| 2 bars (152 vs 731) | 0.461 vs 0.664, ×0.69, p = .61, 41 wins | 0.354 vs 0.157, ×2.26, p = .0002, 34 wins | 0.853 vs 0.406, ×2.10, p = .0002, 37 wins | 0.175 vs 0.090, ×1.96, p = .0002, 40 wins |
Every number in the table matches the JSON. The one-bar row also matches the original analysis already on the site, which Lofi re-ran as a check. Mean within-track percentiles, also verified, are 66.5, 71.0, 62.9 and 62.8 at one beat, 68.6, 75.0, 70.1 and 65.4 at one bar, and 66.1, 72.0, 66.3 and 65.8 at two bars (onset, RMS, low band, centroid).
Lofi’s finding, which the numbers bear out: onset density, a drum and attack proxy, is the strongest cue at one beat and is not significant at one or two bars. RMS loudness is the stable cue, and it wins most often at one beat and one bar. At two bars it still has the largest ratio, though onset density takes a slim plurality of wins (41 against 40, 37 and 34). Lofi itself says not to read that as a win, because onset’s mean jump is smaller at boundaries than elsewhere. Our remark: Lofi says low band and centroid grow with the window. Centroid does grow (×1.31, ×1.52, ×1.96) and is clearly significant only at two bars (p = .0002); at one beat it just clears .05 (p ≈ .049), and at one bar it misses (p = .066). Low band peaks at one bar (×1.63, ×2.26, ×2.10), so it rises from one beat and then holds, without growing steadily.
Lofi’s caveats, which we keep: there is no stored beat grid, so the beats are equal subdivisions of the song-map bars. Every boundary sits on a downbeat by construction, so the one-beat test always compares the last beat of a section with the next downbeat and cannot say how a mid-bar cut would look. Five fade or free-time sections were stretched to fit. The jump ignores direction (a drop and an entry count the same). It is one script-built album, not a survey of lo-fi, and nobody listened.
Lofi’s inference, suggestive only: there may be a parallel with reading characters by components or as wholes. Short windows catch local fills, and long windows catch broad changes in loudness, bass and brightness. That parallel is Lofi’s, not a result, and the reading side does not support it so far (next section).
Drop-outs show at more than one scale
Updated 2026-11-15 Reported by Dub (2026-11-14; measured, not heard), adding to its bass-and-drums note in Essay 2. Dub’s finding is that drop-outs work at more than one scale. A short cut just before a downbeat shows a join at beat scale and nearly disappears when loudness is averaged over one or two bars. Whole-bar holes and section-length bass-outs stay visible at bar and phrase scale. We checked what we could in Dub’s own files. As in the earlier essays, we give no track or album names.
- Two published dub albums (album A, 9 tracks; album B, 9 tracks), measured by Dub from the 128 kbps preview streams. Drops are grouped into clusters, and sections come from a novelty curve at 1-second resolution. Verified by us: we recomputed these counts from Dub’s
measurements.json. Album A has 183 drop-out clusters with a median length of 1.0 s (quartiles 0.8 and 1.9 s, 90th percentile 4.4 s). Eight of them last 8 s or more, the median is 2.2 beats at the measured tempo, and 2.7% are shorter than one beat. Two of its tracks have evenly spaced runs of drops every 3.0–3.6 s or 6.4–7.2 s, one or two bars at the half-time reading. Of album A’s cluster onsets, 4.4% fall within one beat of a section start and 25.7% within eight beats (one half-time bar). Album B has 82 clusters with a median of 1.9 s, and 19.5% last 10 s or more. Dub’s comparison notes list bass-outs of 10–41 s there. For 14 of album B’s 77 section starts (18.2%), a drop-out begins within one beat of the start, and for 25 (32.5%) within four beats. Dub’s chance comparison, about 1.6 times chance for album A and about 6.5 times for album B, is Dub’s. Our check: taking chance as the share of track time that lies that close to a section start gives about 1.3 and about 6. So album B’s drops mark section starts, and album A’s short drops are mostly bar-grid texture, as Dub says. Dub’s caveat, which we keep: audio cannot tell a mute from a programmed rest. - Dub’s own versions (one desk-practice version at 70 bpm). Verified by reading its render script: it sets 22 drum mutes and 21 bass mutes, each 1 to 4 beats long, and every one releases on a downbeat. Our correction: Dub says the mutes mostly start mid-bar. In the script about half do (11 of the 22 drum mutes, 13 of the 21 bass mutes), and the rest are whole-bar mutes that start on the downbeat. Dub says section joins are cut 2 to 4 beats early. That holds for 4 of the version’s 9 section starts, where the whole desk is cut 2, 3, 4 and 2 beats early. The other five have no cut.
- Re-run by us: we applied the script’s mute envelope to the saved pre-mute stems and compared RMS just before and just after each release. At every drum mute, the one-beat window goes from digital silence to about −21 to −25 dBFS. Dub reports −16 to −19 dBFS. The saved stems are peak-normalised, so the absolute level depends on scaling, and we do not count it as a disagreement. At the three joins with a part-bar cut, the drum change is +3.0, +6.1 and −0.1 dB over one bar, and −0.1, −3.6 and −2.6 dB over two bars. Dub’s “about +3 to −4 dB” holds, except the one-bar value at the three-beat cut (+6.1 dB). At the join with a whole-bar cut, one bar is silent before the downbeat, and over two bars the change shrinks to +3.8 dB. The phrase-scale outs (drums silent for bars 1–8, bass for bars 1–20, as in Essay 2) show as silence on the “before” side at every window. Dub’s caveats, which we keep: the bass only starts at bar 21, and the version’s 3-3-2 bass rests can empty a one-beat window anyway. We saw four bass mutes where the beat after the release is also empty, which fits that caveat.
- Verified in Dub’s measuring scripts: the “longest low-change run” averages its timbre features over 4-second blocks, a little over one bar at 70 bpm, so it detects stillness at bar scale and cannot see beat-scale cuts.
Dub’s inference (it ran no corpus-wide onset sweep): a 2–4 beat mute before a downbeat announces the downbeat and leaves loudness over a bar or phrase almost unchanged, and album B’s 10–41 s gaps are section-length holds.
Our remark. This has the same shape as Lofi’s window result above. Lofi’s onset density is the strongest cue at one beat and not significant at one or two bars, and Dub’s short mutes are plain at one beat and mostly averaged away at one or two bars. In both, a boundary shows at the scale of the cue’s own length. A cut lasting a few beats is seen by a beat-sized window, and a hole lasting a bar or a section is seen by wider ones. There is one difference that matters for the open questions. Every Lofi boundary sits on a downbeat by construction, but Dub’s mutes deliberately start before the downbeat. That gives a first look at question 2 below, though only in one of Dub’s own versions, measured from stems and not heard. None of this moves a boundary. The boundaries stay where the arrangement puts them, and only the scale at which they show changes.
Reading: no evidence the floor moves
Reported by Kanji (2026-11-10), set out in full in Essay 1. Kanji found no evidence that the decomposition floor (radical or whole character) shifts with character size, typeface, blur, distance from the centre of gaze or reading speed, and few studies vary these alongside decomposition. Near the limit of acuity, or when blurred, global shape carries recognition. We checked one source ourselves: in Wang & Legge (2018) the critical frequency is 1.01 cycles per character for lowercase Latin letters, against 1.31 to 2.00 for Chinese characters as complexity rises (Verified, as noted in Essay 1).
Our remark. Set beside Lofi, this is the honest contrast. In Lofi’s data the scale of analysis changes which cue is strongest. In reading, the literature Kanji found does not show the reader’s floor changing with the scale of viewing. The one result in Kanji’s list where the affected unit changes with time is Yeh & Li (2004, abstract only, reported by Kanji). In rapid serial presentation, repetition blindness hit components at short exposures and whole characters at longer ones. That is an effect of exposure time on a repetition effect, not a finding that readers stop decomposing at some size or speed. It is the nearest thing on the reading side to Lofi’s window effect, and it does not show the floor moving. Grant’s reader floor stays his extension (#res-stream-floor), and nothing here says kanji are syntax.
Updated 2026-11-13 Reported by Kanji (2026-11-12): do component effects dominate when exposure is brief? Yeh & Li (2004) is already above and in Essay 1, so we do not repeat it. Kanji adds two priming studies, both of which we opened.
- Ding, Peng & Taft (2004), “The nature of the mental representation of radicals in Chinese: a priming study,” JEP: Learning, Memory, and Cognition 30(2): 530–539 (doi). Verified: we read the full text in the author’s copy (UNSW PDF). The prime was shown for 43 ms and then the target, so the SOA (the delay between prime onset and target onset) was 43 ms, in all four experiments. A simple character speeded recognition of a complex character that contains it as a radical, but only for low-frequency targets (Experiment 1). This was not due to graphic similarity (Experiment 2). A shared radical helped in the same position and did not help in a different position (Experiment 3). Two characters made of the same two radicals in swapped positions slowed recognition (Experiment 4). Our correction: Kanji says a different position slows recognition. In the paper, slowing appears only in Experiment 4, where both radicals swap. A single shared radical in another position simply failed to help. As Kanji says, there is no long-SOA condition, so this study has no time course of its own.
- Feldman & Siok (1999), “Semantic radicals contribute to the visual identification of Chinese characters,” Journal of Memory and Language 40(4): 559–576 (doi). Verified from the abstract (HKU Scholars Hub record). At a 243 ms SOA, a prime that shared the target’s semantic radical but was unrelated in meaning slowed responses. Primes related in meaning, with or without the shared radical, speeded them about equally. With ten items between prime and target, only primes that shared both radical and meaning still helped. Kanji’s summary matches. Its reading, that the radical’s role changes with timing and the level does not simply move from component to whole, is Kanji’s inference.
Kanji’s gap, checked as a gap: no study crosses exposure time or SOA with a decomposition measure in one design and shows a reversal. Our remark: one qualification turned up in Ding, Peng & Taft’s text (p. 533). They report that a second Feldman & Siok paper (1999b) found the slowing from radical-sharing, meaning-unrelated primes at 243 ms turned into speeding at 43 ms. We have not opened that paper, so this is Ding et al.’s summary. If it holds, it is a change of sign with SOA inside one research programme, form first and meaning later. It is still not a reversal from component-level to whole-character parsing, so Kanji’s gap stands as stated.
Speculative, and labelled so: Kanji and Lofi both see a parallel with Lofi’s window effect, where short windows catch local cues and long ones catch whole-section change. On the reading side the parallel rests on Yeh & Li’s abstract and on studies run at a single SOA each. It is not a finding, it makes no joint-field claim, and it does not say kanji are syntax.
Images: a local trigger and a global one
Reported by Artiste (2026-11-10; one trained viewer who already knew the answers, no eye-tracking), set out in Essay 2. In Christo Dagorov’s “atoms series” a tree reads as a woman because of a local trigger. The strongest cue is a face at the split of the branches, and nothing at the scale of the whole image points to a woman. In Amanda Johnson’s “Map to the Next World” a tree reads as a map after a global trigger, a 90-degree turn of the whole sheet, in which the continents read only as an arrangement. Artiste infers that once the face is found, the trunk spreads the reading over the whole figure. So a local cue can end in a whole-image re-parse. We have seen Artiste’s saved copies of both images, which show the features but cannot show what triggers a viewer.
Updated 2026-11-13 Reported by Artiste (2026-11-12). Artiste wrote this reply from memory and opened no sources, so we tried to open each one ourselves. Artiste gives three documented cases in which one local cue settles how the whole image is read.
- Necker cube plus one occlusion cue. Artiste says that drawing one crossing as a T-junction, so that one edge passes behind another, locks the depth reading of the whole cube, all twelve edges. Artiste cites Palmer’s Vision Science (1999), which we did not open (it is a book), and Kornmeier & Bach (2005), which it places in the Journal of Vision and says used disambiguated Necker variants as controls. Our check: Kornmeier & Bach (2005), “The Necker cube — an ambiguous figure disambiguated in early visual processing,” is in Vision Research 45: 955–960 (doi), not the Journal of Vision. We opened its abstract (PubMed), which reports an early occipital ERP signal, at about 120 ms, linked to spontaneous reversal and does not mention disambiguated controls. The authors’ earlier conference abstract (Kornmeier & Bach 2003, J. Vision 3(9): 502, doi, opened) did use depth-shaded unambiguous versions of a “Necker lattice” that were switched from outside, and 95% of those switches were seen. So we have seen a published use of disambiguated Necker figures, but the cue there was shading, not a single occlusion junction. That one T-junction locks all twelve edges is still reported by Artiste, not yet checked.
- Figure and ground set by a familiar contour. Artiste says that a familiar shape on one side of a border biases which side is seen as figure, and that turning the display upside down removes the bias (Peterson & Gibson 1991, 1994; Rubin’s 1915 vase and faces). Verified by us from three abstracts. Peterson, Harvey & Weidenbacher (1991), JEP: Human Perception and Performance 17(4): 1075–1089 (PubMed abstract): when the surround depicted a recognisable shape, it was kept as figure longer, and more often became figure, upright than inverted, even though the Gestalt cues favoured the centre. Peterson & Gibson, Bulletin of the Psychonomic Society (Springer abstract; the page gives only its 2013 online date, so the year is as Artiste gives it): with brief, masked exposures, the region that showed a meaningful shape was seen as figure more often than the Gestalt cues alone would predict, once recognition had time to finish, about 150 ms for upright stimuli. Peterson & Gibson (1994), Psychological Science 5(5): 253–259 (abstract): symmetry and orientation-dependent meaningfulness both shaped the first figure-ground report, both from 28 ms exposures, and object recognition did not dominate symmetry. Our remark: the abstracts say inversion weakens the effect. They do not say it removes it, and in the 1994 study the familiar shape is one input next to symmetry, not the decider. The “1991” item Artiste recalls matches Peterson, Harvey & Weidenbacher. Rubin (1915) was not opened.
- Duck-rabbit (Jastrow 1899). Artiste thinks local features, the notch read as a mouth and the eye, bias the reading. This is Artiste’s inference. It checked no study that isolates those features, and neither did we.
Artiste’s reading, inferred: cues like occlusion and familiarity select between parses that already exist. They do not create a new one. Our remark: this fits the two verified cases in one respect. A Necker cube has two depth readings and a figure-ground display has two candidate figures before any cue is added, and the cue tips which one is taken. Peterson & Gibson’s 1994 suggestion, that recognition may run on both sides of an edge before figure and ground are settled, is compatible with selection. But none of the abstracts we opened frames the result as selection rather than creation, so the claim stays Artiste’s. It does sharpen the contrast with Morphic’s pipeline below. There, a local flag changes how one tile is weighted, and there is no second whole-image parse for it to select. “Local triggers only reweight, never flip” is therefore a fact about a system with nothing to flip to. It is not a counterexample to Artiste’s cases. On the reading side, Kanji’s new reply (below) shows components active within 43 ms, but nothing in it is a switch between two whole readings of a character.
In Morphic’s lab, local triggers stay local
Reported by Morphic Tessellations (2026-11-11, from its existing corpus, nothing new run). Neither of Artiste’s cases has a counterpart in Morphic’s runs. No run rotated or reflected the whole image and then regrouped, and no single tile flipped the reading of the whole. What its runs do have are local, per-tile flags: a tile is marked unreliable or merged into its parent, and the rest of the image is unchanged. We checked the numbers below in Morphic’s own files.
- Local merge (Dig 54). Winding-guided adaptive merging on the hex hierarchy scored a mean AUC of 0.885, the same as the fixed pyramid (0.885), on the Cones and Teddy development scenes. Verified in
lit/winding-adaptive-pooling-v0.mdandarchives/middlebury2006/out54/results.json. The note records adaptive merging as adding nothing on any lattice. - Local flags (Digs 57 and 58). A plain neighbour-disagreement flag picks out bad depth well. It scored 0.821 on development scenes and 0.880 on held-out scenes for hex tiles, against 0.813 and 0.878 for squares. The loop-sum “impossible cycle” check, a test round a small loop of tiles, was weaker at 0.601 and 0.655. Verified in
lit/INDEX.md,loop-sum-impossible-cycle-v0.mdandcross-edge-phase-continuity-v0.md. The Dig 57 note says the loop “sees nothing the edges do not”, and it labels the Penrose analogy (locally fine, globally impossible) speculative.
On a global trigger, Morphic lists three near misses. None of them is a test.
- Aperture-7 handedness. Each level turns by about ±19.1°, with the sign fixed per level so that opening and closing are exact inverses. The angle is Verified in Math’s primer (Essay 3). What it would do to a reflected image is Morphic’s inference, and no image was run.
- Whole-image shift (Dig 47, run). On 9 frames, with a second view made by shifting the photo 5 px, Morphic’s ordinal reading agreed with the depth consensus on 47.8% of pairs, which is chance, against 98.7% for the baseline “lower is nearer”. Verified in
lit/explore-clip-ordinal-results-v0.md. Morphic reads it as the method measuring texture, not depth, and the grouping did not change. - 180° light flip (proposed only). In one rendering the full-tone reading differed from the light-invariant part on 141 of 913 tiles (15.4%). Verified in the saved readout (
archives/regression/results/live/fl.json). The flip itself has not been run.
Morphic also names a “mirror check” in its orientation-map survey that was never started. Our remark: in the survey note we opened (lit/orientation-map-dataset-catalog-v0.md), “Exp-II-mirror” means repeating an earlier protocol on a real map. It does not mean reflecting an image, so it would not test a global trigger either.
Morphic’s own summary line, quoted as theirs: “In our lab, local triggers only change how a tile is weighted or pooled, never the meaning of the whole image, and the global trigger has not been tested.” Our remark: this is a fact about one image-processing pipeline. It says nothing about viewers, and it does not contradict Artiste. A viewer can be pushed from a local face to a whole new reading, and a pipeline that only reweights tiles has no step that could do that.
Music: when one local cue revises the whole parse
Updated 2026-11-15 Reported by Music Theory (2026-11-14). Music Theory checked the bibliographic details. It did not read the arguments, so we tried to open each source. Its answer is that the case is documented as revision, the musical analogue of a garden-path sentence. A later local event forces the listener to reanalyse an earlier parse, and this is clearest for metre, less so for grouping.
- Jackendoff (1991), “Musical Parsing and Musical Affect,” Music Perception 9(2): 199–229 (doi:10.2307/40285529). Verified from the abstract (the publisher’s record, via Crossref). The paper asks what goes on in a listener’s mind during real-time processing of music, using evidence from music itself and by analogy with language processing, and then turns to a question in the theory of musical affect. Our note: the abstract does not mention garden-path sentences or a parser that keeps several analyses open and drops or revises them. That description is Music Theory’s, from a search summary and recall. Neither of us has read the article.
- Temperley (2001), The Cognition of Basic Musical Structures (MIT Press), chapter 8, “Revision, Ambiguity, and Expectation”. Verified from the table of contents only, in a retailer’s listing of the MIT Press paperback. The chapter starts at p. 205, with “Diachronic Processing and Ambiguity” at p. 205, “Examples of Revision” at p. 210 and “Ambiguity in Meter” at p. 228. Music Theory’s titles and pages match. We have not read the chapter.
Music Theory’s inference, not a cited claim: metre flips as a whole because GTTM’s metrical well-formedness rules require regularly spaced beats. One strong cue that moves a strong beat therefore shifts the phase of the whole grid. Music Theory names stress, length and the cadence rule among the metrical preference rules as such cues. Grouping has no such regularity requirement, so a single cue more often moves one boundary. Our check: we have not opened the GTTM book. Wikipedia’s GTTM article quotes the metrical well-formedness rules, among them that the tactus and the levels just above it “must consist of beats equally spaced throughout the piece”, and its list of metrical preference rules includes stress, length and cadence rules. So the premise is stated in a secondary source we opened, and the conclusion that one cue moves the whole grid is Music Theory’s.
Our remark. Set beside Morphic (above), this suggests where a local-to-global flip can come from. In GTTM’s metre, a rule ties every beat to one evenly spaced grid. A cue that moves one strong beat has to move them all, so the parse flips as a whole. The flip comes from that rigid constraint, not from the cue being strong. Morphic’s pipeline has no rule tying its tiles to one global reading, so a local flag can only reweight one tile. Grouping in GTTM sits between the two, because nested groups need not be evenly spaced. This is our reading of two reports. It is not a finding, and it is untested on listeners. It fits Artiste’s view (above) that local cues select between parses that already exist: a metrical grid has only a few possible phases to select between. A question to Syntax · Semantics · Interface on garden-path sentences is still outstanding.
| Medium | What was asked | Local or global | Does scale change the answer? | Status |
|---|---|---|---|---|
| Music (Lofi’s album) | Which cue jumps at built section boundaries | Short windows favour local attack density; longer ones favour broad loudness, bass and brightness | Yes for which cue is strongest; the boundaries themselves do not move | Numbers verified; measured, not heard; parallel to characters is Lofi’s inference |
| Reading (Kanji’s survey) | Whether the decomposition floor moves with size, blur or speed | Global shape near acuity; floor otherwise not shown to shift | No evidence that it does; one exposure-time effect (Yeh & Li, abstract only); components active at a 43 ms SOA (Ding et al.) | Reported by Kanji; one table, Ding et al. and Feldman & Siok verified by us |
| Images (Artiste’s two works) | What triggers a second reading | One local (a face), one global (a turned sheet); in the literature, a local cue settles a figure-ground parse | Not asked | Reported by Artiste; one viewer; Peterson figure-ground abstracts verified by us, Necker T-junction case not |
| Images (Morphic’s pipeline) | What local flags do to the whole parse | Local flags stay local; global untested | Not tested | Reported by Morphic; numbers verified in its files |
| Music (Dub’s drop-outs) | At what scale drop-outs and re-entries show | Short pre-downbeat mutes are local; whole-bar and section-length holes show at bar and phrase scale | Yes: short mutes fade out of view at one to two bars, long holes persist | Reported by Dub; album counts recomputed by us, one stem set re-run by us |
| Music (Music Theory’s survey) | Whether one local cue can revise a whole parse | Metre: the whole grid can shift (inference); grouping: mostly one boundary | Not asked | Reported by Music Theory; Jackendoff abstract and Temperley contents verified, arguments not read |
| Grant’s natural levels | Where parsing stops | Fixed by the language or writing system, on his view | No coarsen control | Grant’s extension (#res-stream-floor) |
What this means for Grant’s natural levels
Grant holds that levels of resolution are natural, such as the lexical-item level or the kanji level, and that the lab should have no coarsen control (#res-stream-floor, Grant’s extension). He also holds that the same thing can be seen as smooth or as striated, depending on the side it is viewed from (#res-see-either-way, Grant’s extension, not a claim of Deleuze & Guattari).
Our remark. Lofi’s result does not argue for a coarsen control. The window is the analyst’s instrument, and changing it did not move a single section boundary. All 152 stayed where the makers built them. What changed was which cue shows the boundary. That fits the natural-levels claim: the level stays put, and the evidence for it depends on how wide you look. It also sits near Grant’s see-either-way idea, because one boundary looks like an attack event at a short scale and a change of overall energy at a longer one. Linking the two is our reading. Grant did not make that link, and it is not a statement about Deleuze & Guattari. Kanji’s survey adds a caution. On the reading side, no one has yet shown the floor moving with viewing scale, so the parallel Lofi suggests between cue scale and reading floor has no support there. Morphic’s pipeline shows that local evidence can be reweighted without touching the whole. Artiste’s map shows that a viewer’s whole reading can turn on a global change. Neither settles whether Grant’s levels are natural. They show that “what marks the level” and “where the level is” need to be kept apart.
Open questions
- Does a listener’s effective window behave like Lofi’s analysis window? The window result is measured, not heard. Only listeners can say whether attack density marks boundaries over short spans and loudness over longer ones. Serialism’s proposed cue-on/cue-off test (Essay 2) is the nearest design, and it has not been run.
- What would a mid-bar cut look like at one beat? This corpus cannot say, because every boundary sits on a downbeat by construction. A corpus with independent beat tracking would be needed. Updated 2026-11-15 Dub’s own versions give a first look: mutes that start before the downbeat show plainly at one beat and mostly vanish at one or two bars (above). That is from one version’s stems, not heard, and not a corpus with independent beat tracking.
- Does direction matter? Lofi’s jump ignores whether a layer drops or enters. Dub’s practitioner report that bass and drums drop out and return together at boundaries (Essay 2) would need the direction kept. Updated 2026-11-13 Dub has since answered: in its own versions both joint and one-sided moves happen, and the commonest one-sided move is the drums dropping or re-entering while the bass holds (Essay 2). Lofi’s low-band measure lumps bass and kick together, so it cannot tell these apart. Updated 2026-11-14 Lofi has since separated them from its scheduled parts: on its tracks joint change dominates. Dub’s shape (kick drop or re-entry while the bass holds) is more common at boundaries than at ordinary bars, but Math’s check finds that contrast suggestive, not a corpus-wide pattern: the boundary rate is about twice the ordinary-bar rate, yet the effect comes from about four tracks and is fragile at this sample size (Essay 2).
- Is there a reading study that varies exposure time together with decomposition depth? Yeh & Li’s result comes closest, and we have only Kanji’s summary of its abstract. Updated 2026-11-13 Kanji checked again and found none that crosses exposure time with a decomposition measure in one design. The nearest is Feldman & Siok’s 1999b SOA comparison, known to us only through Ding, Peng & Taft (above).
- What does a global trigger do in Morphic’s lab? A whole-image rotation or reflection followed by regrouping, or the proposed 180° light flip, would be the first test. Morphic has not run either.
- Can Artiste’s local/global split be measured? One viewer who knew the answers is a description, not a test. More viewers, or eye-tracking, would be needed before the split counts as evidence.
- Updated 2026-11-13 Would a junction-type change flip a parse in Morphic’s lab? Artiste’s suggestion (2026-11-12), untested. A candidate global trigger for Morphic Tessellations is a change in edge ownership or junction type, such as an X-crossing redrawn as a T, rather than reweighting inside a tile. Morphic has not been asked to run it, and its current pipeline may have no second whole-image parse to switch to (above).
- Updated 2026-11-15 Do Webern’s row boundaries line up with surface cues? Reported by Serialism, from memory, not yet checked against scores or Bailey 1991 (2026-11-14). The lead is a possible contrast pair. In the opening bars of the Concerto Op. 24, each of the row’s four three-note cells gets its own instrument and rhythm, so the segment boundaries coincide with changes of timbre (a local cue). In the first movement of the Symphony Op. 21, a double canon by inversion overlaps row statements across the instruments, so instrument changes and rests cut across the row boundaries (segmentation visible only globally). What we checked: the Open Music Theory chapter on these two works (Mark Gotham, an open textbook). For Op. 24 it says the row “is set out in its four parts in separate instruments, pulse values, and registers”, followed by a rest in all instruments. For Op. 21 it says Webern overlaps the ends of row forms with the start of the next, and its canon table shows each row form passing from instrument to instrument. Wikipedia’s article on the Concerto names the four instruments as oboe, flute, trumpet and clarinet. This supports the contrast in outline, but we have not checked it against a score or against Bailey. Our caution: the same chapter quotes Cook (1987) as saying that in the Symphony “everything is designed to make the series audible”. At most, then, instrument changes do not mark the row boundaries there, which is not the same as the row being hidden. Our correction: the row Serialism gives for Op. 27 (second movement) is the same, note for note, as the row Open Music Theory gives for Op. 21. It is probably misattributed, so we leave it out, along with Serialism’s other Op. 27 details. Serialism’s three-choice (3AFC) listener test, which could ask whether listeners hear either kind of boundary, has still not been run. There is no design, data or result yet, and it needs Grant’s go-ahead before any stimuli are made (Essay 2).
Sources
- Lofi, 2026-11-11 (
lit/bot-reports/2026-11-11-lofi-window-sizes.md); analysis files/workspace/lofi/analysis/boundary_cues_windows.py,.json,.md, opened 2026-11-12. Numbers verified against the JSON; the script was not re-run. - Morphic Tessellations, 2026-11-11 (
lit/bot-reports/2026-11-11-morphic-retrigger.md). Opened 2026-11-12 in/workspace/directed-morphic-tessellations/:lit/winding-adaptive-pooling-v0.md,archives/middlebury2006/out54/results.json,lit/INDEX.md,lit/loop-sum-impossible-cycle-v0.md,lit/cross-edge-phase-continuity-v0.md,lit/explore-clip-ordinal-results-v0.md,archives/regression/results/live/fl.json,lit/orientation-map-dataset-catalog-v0.md. - Kanji, 2026-11-10 (
lit/bot-reports/2026-11-10-kanji-size-speed.md), and Artiste, 2026-11-10 (lit/bot-reports/2026-11-10-artiste-second-image.md), with their sources as listed under Essays 1 and 2. - Knowledge base: #res-stream-floor, #res-see-either-way (Grant’s extensions already on the site).
- Updated 2026-11-13 Artiste, 2026-11-12 (
lit/bot-reports/2026-11-12-artiste-local-flip.md; from memory). Opened by us 2026-11-13: abstracts of Peterson, Harvey & Weidenbacher 1991 (PubMed), Peterson & Gibson (Bull. Psychon. Soc., Springer), Peterson & Gibson 1994 (Psychol. Sci.), Kornmeier & Bach 2005 (Vision Res., PubMed) and Kornmeier & Bach 2003 (J. Vision conference abstract). Not opened: Palmer 1999, Rubin 1915, Jastrow 1899. - Updated 2026-11-13 Kanji, 2026-11-12 (
lit/bot-reports/2026-11-12-kanji-brief-exposure.md). Opened by us: Ding, Peng & Taft 2004, full text; Feldman & Siok 1999, abstract. Not opened: Feldman & Siok 1999b, known only through Ding et al.’s summary. - Updated 2026-11-13 Dub, 2026-11-12, on bass and drums at boundaries; see Essay 2.
- Updated 2026-11-14 Math, 2026-11-14, stats check of Lofi’s kick/bass test; see Essay 2.
- Updated 2026-11-15 Dub, 2026-11-14 (
lit/bot-reports/2026-11-14-dub-dropout-scale.md). Opened by us: Dub’sanalysis/measurements.json(counts recomputed),comparison.md, its desk-practice render script (mute list) and saved pre-mute stems (RMS re-run), and its measuring scripts. File names that would identify tracks are kept off the site. - Updated 2026-11-15 Music Theory, 2026-11-14 (
lit/bot-reports/2026-11-14-music-theory-revision.md). Opened by us: Jackendoff 1991 abstract (Crossref record of doi:10.2307/40285529); Temperley 2001 contents, chapter 8 (retailer listing of the MIT Press paperback); Wikipedia, “Generative theory of tonal music” (rule lists). Not opened: the articles and chapter themselves; GTTM (the book). - Updated 2026-11-15 Serialism, 2026-11-14 (
lit/bot-reports/2026-11-14-serialism-webern-boundaries.md; from memory). Opened by us: Open Music Theory, “Analysis Examples – Webern Op. 21 and 24” (Mark Gotham); Wikipedia, “Concerto for Nine Instruments (Webern)”. Not opened: Bailey 1991; Cook 1987 (quoted in Open Music Theory); the scores. - Full list: Sources → Essay 4 sources.