Runs in your browser. No outside libraries, no network calls, no cookies. Sound is off until you press Play with sound unmuted.

How to read this page

A Stream and brackets

Grant’s extension the stream model and depth as a count · Reported by Math the standard objects (free monoid, Dyck language, tree) · Our remark the interface and the count of bracketings

Grant’s model is a stream of 0s and 1s in which some 0s act as brackets. Click a 0 to cycle it: plain 0, then “[”, then “]”. The page checks whether the brackets balance (a Dyck word, Math’s standard object for matched brackets) and gives every element its depth: the number of brackets enclosing it, a count rather than a height. The ribbon, the arcs and the tree are three drawings of the same bracketing and change together. You can also type 0, 1, [ and ] in the box.

“Grant’s made-up example” is 01010010000110010101000011101, a string he invented to illustrate the model; it carries no data.

1–2. Stream with depth ribbon (numbers are depths), and matched pairs as arcs
3. Nested tree of groups. Click or press Enter on a node to collapse it. Drawing convention: the inside is hidden, not absent.

What this shows. That one string carries many bracketings, and that forgetting the brackets sends every one of them to the same plain string, an element of the free monoid (Math’s point). The count above is ours: we computed it by dynamic programming over the number of open brackets and checked it against brute-force enumeration up to 12 zeros (1, 1, 2, 4, 9, 21, 51, …; these are, as far as we recall, the Motzkin numbers, a name we did not check). It counts all balanced markings, including empty pairs “[]”, which the page also allows.

What it does not show. Which bracketing is right. Nothing in the 0s and 1s picks one; the brackets are added structure. Collapsing a node is not a change of resolution.

B Cues and boundaries

Reported by Music Theory the cue mapping, its reading · Reported by Serialism Grant’s listening report · Verified Dibben 1994 (abstract) · Our remark the toy score

A short note sequence drawn as a piano roll (amber is voice 1, cyan voice 2; brighter is louder). Each gap between notes gets a score: it adds the weight of each cue that fires there, and a gap whose score reaches your threshold is drawn as a candidate boundary. The cues are simplified from the grouping rules Music Theory summarised: a longer gap than both neighbouring gaps (proximity, GPR 2), a leap of 7 or more semitones bigger than both neighbouring leaps (register, GPR 3), a dynamic change of 0.25 or more (GPR 3), the start of a 4-note pitch pattern that occurs more than once (parallelism, GPR 6), and a change of voice. Drag a note, or focus it and use the arrow keys. This is our toy score, an illustration of GPR 2, GPR 3 and GPR 6, not a model from the literature.

Piano roll; magenta lines are candidate boundaries; bars underneath are the toy score per gap

Score for each gap

    What this shows. The two presets use one made-up twelve-tone row (ours). “Strict row” has even onsets, one register and one voice, and no gap reaches the default threshold. “Same row with cues” adds a returning riff and one leading voice, two of the changes Grant reported, via the Serialism bot, as what made a strict row cohere for him, plus a register split that is our addition: one listener, changes made together, not a controlled study. Dibben (1994) found that listeners could match reductions to the source in tonal music but not in atonal music (Verified from the abstract).

    What it does not show. What anyone hears. The score does not predict perception, its weights and thresholds are arbitrary, and real grouping rules are preferences that interact, not a sum.

    C Zeta lobes

    Our remark computed by us, 2026-10-17 · Reported by Math no standard bracket-depth rule known · Grant’s extension zeros as brackets (not tested here)

    Between consecutive zeros of the Riemann zeta function on the critical line, Hardy’s Z(t) makes one lobe, alternately positive and negative. Each bar is one lobe’s extremum, from our saved values for the first 100 zeros. A merge tree joins lobes as a level sweeps through the values: “from below” starts at the negative troughs, “from above” at the positive peaks. A lobe’s depth is the number of merges above it. The window choice exists only to show the window dependence we already reported; it is not a resolution control.

    Window
    Sweep
    View
    Signed Z lobes with the merge tree (or persistence bars; faded bars fall below the cutoff)

    What this shows. The numbers our script printed (tools/signed_z_tree.py): the first lobe’s depth is 3, 5 and 6 from below and 4, 5 and 6 from above for windows of 25, 50 and 100 zeros, so depth depends on the window. In the 100-zero window every persistence is at least 1.309, so cutoffs of 0.5 and 1 prune nothing and 2 prunes 7 of the 48 (from below) or 49 (from above) finite pairs.

    What it does not show. Anything about zeros as brackets. This is one window and a weak test: shuffling the lobe heights (within each sign) nests about as deep. Math knows of no standard rule that gives a zero a bracket depth. The Riemann hypothesis is open, and nothing here bears on it.

    D Kanji part-tree

    Verified decompositions and stroke counts, KanjiVG git 70a0b7ae, IDS cross-checked on Wiktionary · Reported by Kanji reader activation, no single floor · Grant’s extension operator output and a natural parsing floor · Our remark the interface and the bracket string

    Four characters opened as a tree: the whole character, then its components, then the strokes those components are written with. Click a part, or focus it and press Enter, to look inside. Depth is the number of enclosing brackets in the string below, Grant’s count. This panel does not treat kanji as a syntax.

    Click a part to look inside it. That is a reader’s exploration of the tree, not a change of resolution. Grant holds that levels are natural, so there is no coarsen or resolution slider here. Collapsing a node only hides what is inside it.

    Where readers stop

    Reported by Kanji Reading research, from abstracts only (we did not read the full papers), finds that readers activate a character’s radical or component parts, but the split looks task- and component-dependent, and no single natural stopping point was found. A 226-undergraduate survey of how characters are divided, which Kanji reports from Tamaoka et al. (2002), is not in the abstract we opened and was not seen by us. With the overlay on, first-level parts are marked as that component level. The mark is not a floor and not a cutoff.

    Grant’s extension The claim that each kanji is the output of an operator function, and that the parsing floor is natural, is Grant’s, not a research finding.

    What this shows. The part tree of four characters, taken from KanjiVG’s nested groups, and a bracket string for whatever is currently open, with depth as a count of enclosing brackets. 五 and 重 open straight to strokes: KanjiVG splits them with part-groups, a file convention, and we did not treat those groups as components.

    What it does not show. Where a reader actually stops, or that this tree is a linguistic syntax. Expanding a node only shows what is already inside it.

    Sources and credits