/.metanomicon

One fragment per cell. The drawing buffer is not the pixel grid; it is the dither grid, upscaled with image-rendering: pixelated. Ordered dithering asks each cell whether its luminance clears a fixed threshold, and if several pixels inside one cell each answer from their own slightly different luminance, the cell boils. Rendering at the cell makes the answer singular. It also pays for the third dimension: at a four-pixel pitch the renderer runs on a sixteenth of the pixels, and on a 2× display a sixty-fourth.

Eighty-four seconds per revolution. Shimmer scales with how fast luminance crosses a threshold that is standing still. There are sixty-four Bayer levels; at this rate a point on the surface crosses one roughly every nineteen frames rather than every frame. The slow register is the canon's anyway — a climate, not a plasma — so the fix and the law are the same rule.

The dark patches are sampled in object space. They are a noise field read on the local surface normal, before the rotation is applied, so a patch belongs to its sphere and turns with it. Reading the same field from the fragment coordinate instead is the copy-paste that makes a rotating sphere crawl: the threshold would stand still while the value under it moved. Their two cut-offs, 0.6532 and 0.5934, are the 85th and 72nd percentiles of this exact noise measured over 312,000 points on the sphere. The flat lattice's own figures, 0.600 and 0.555, belong to a different hash: against this distribution they sit near the 72nd and 60th percentiles, so inheriting them would have put more than a quarter of every sphere into the deep step instead of the intended sixth.

Warm leads, cold trails. The substrate tints its cells by the sign of the climate's gradient; the same law applies here, and here it is exact rather than sampled. The angular velocity is known, so whether a point is turning into the light or out of it is a dot product, not an estimate. The tint is confined to the terminator and capped at half saturation, because away from the terminator nothing is changing state and a colour there would be decoration.

The pointer bends the light; it does not aim it. The key light has its own drift on a sixty-one second cycle and the plate is complete with no pointer at all. Where a pointer is present it adds a bounded deflection on the profile x(1−x)² — zero directly under the cursor, peaking in a ring, dead at 220 pixels. Zero at the centre is the point: a peak under the cursor is a spotlight, and a spotlight has a direction singularity.

The cyan dot marks the sphere you cannot see. The centre of a close pack is enclosed by its twelve neighbours from every direction; no ray reaches it. The dot is drawn last, over the gate, beating on sin⁴(t·1.55). It is the only cyan on the plate and the only thing here that is not a measurement of the surface.

On the flower of life. The construction is compass and straightedge: equal circles centred on a hexagonal lattice. It carries a long ornamental and occult history, and in sacred-geometry literature it is read as a claim about how the world is put together. Nothing here reads it that way. It is in this canon for a mechanical reason — hexagonal close packing is the densest arrangement of equal spheres, and the flat ring of six around one is that solid's shadow along its own axis. Press 7 · flat to remove the two capping layers and the shadow is what is left. The plate demonstrates the geometry. It is a procedure with a reason, not evidence for anything.

Every getter is UTC, and that is not a preference. Swap one of them for a local-time getter and the same instant hashes differently in Berlin and in Denver — the page still runs, it simply stops agreeing with itself across machines. Reproducibility that disagrees with itself is not reproducibility.

The location is fixed, stated, and not yours. Planetary hours are twelfths of the actual daylight, so they need a sunrise, and a sunrise needs a place. Asking for the viewer's location would make the seed a property of the viewer rather than of the date, which is the law this section exists to keep. So the observer stands at 51.4779° N, 0.0015° W — the Royal Observatory, where UTC already points — and the polar failure case, where the sunrise equation returns no root at all, never arises.

Accurate to the day and to the minute, which is not the same as precise. The moon's longitude is Meeus's low-order series truncated to its six largest periodic terms: about a third of a degree, inside a mansion twelve degrees wide, so the index is right even though the longitude is not ephemeris-grade. Sunrise is the NOAA equation, good to about a minute. The synodic month is the IAU mean value, 29.530588861 days, and the true one varies by half a day across a cycle. There is no VSOP87 here, no ELP2000, no nutation and no aberration.

This is a procedure with a reason, not a claim about the world. Nothing here predicts anything, and the section makes no argument that it could. What the tradition supplies is a parameter list: several real astronomical quantities that are periodic, computable from a date alone, and uncorrelated with each other — the moon's phase turns over in a month, the planetary hour in a fifth of a day, the mansion in a day, the year in a year. That is precisely what a seed wants. The arithmetic underneath is ordinary: a Julian day number, six sine terms, a division into twelfths, and FNV-1a.

Names, never glyphs, and an index with no name. Departure Mono was measured and does not carry the planetary symbols, so the hours are spelled out; a later pass that improves them back into symbols will render seven tofu boxes. The mansion is given as a number out of twenty-eight and left unnamed, because the three traditions that count twenty-eight of them disagree about which is which, and this system does not borrow authority it has not earned.

What is borrowed and what is declined. What is borrowed is a permutation of sixty-four six-bit words that people spent three thousand years arranging, and one structural fact about it: consecutive pairs are joined by turning the figure upside down, and where a figure is unchanged by turning — there are eight of those, one for each choice of upper-half lines — the pair is joined by swapping every line instead. That is checkable, and it is checked. What is declined is everything else. No name, no judgement, no reading, no correspondence is attached to any figure here; the sequence is used as an ordering and nothing more, because the system does not borrow authority it has not earned.

Why the comparison is fair. Both matrices are bijections from the sixty-four cells onto the numbers 0–63, so at any gain the two panels put very nearly the same quantity of ink on the page — the lit-cell counts are printed so that can be checked rather than assumed. The statistics run over an aligned 168×120 sub-region, divisible by four (the measurement window) and by eight (the matrix tile), so neither matrix is charged for a partial tile at the bottom edge. The residual variance subtracts the local target density from the local lit density: a dark region is sparse under any matrix, and that is the photograph's doing, not the ordering's.

Why ordered, still. This changes nothing about the canon's gate. King Wen is as deterministic, stateless and tileable as Bayer — that is what makes it a legitimate rival and not a curiosity. It loses on the measurement, and the measurement is the reason to keep Bayer, which is a better reason than the one the section had before: that the recurrence is elegant.

The disc is window.MN_phaseDisc, the primitive §09 already ships, at twelve cells. The lit fraction is the day's balance; the terminator is a real ellipse; every cell is one Bayer-8 decision. There is no grey in the grid because there is no grey anywhere — what reads as a mid-tone is a count of marks.

§09 caps that primitive at three uses and says the fourth turns it into decoration. The cap counts instruments, not calls: the rail is one use however many discs it steps through, and the almanac is one instrument — a single grid, read as a grid, whose cell happens to be the disc. It is the fourth instrument and the last one this page is allowed. What §09 forbids is the disc appearing as an ornament beside something that did not need it; a calendar whose only mark is a lit fraction is the opposite case.

Three statistics were prototyped and all three were drawn. Moon illumination is the control: it is a 29.53-day cosine by construction, so bands are guaranteed and carry nothing the ephemeris has not already said. The field's mean density at the day's offset carries the page's own climate, but the seed is an FNV-1a hash, so consecutive days land in unrelated parts of the noise domain — measured lag-1 autocorrelation −0.009 and, drawn, a column-to-column contrast of 1.66 lit cells against the control's 21.34. That is a calendar of static. The balance keeps the lunation — lag-29 autocorrelation 0.904 on the lit-cell counts against the control's 0.913 — and adds the terrain: 13.8% of its variation survives regressing out illumination, and it paints 52 distinguishable disc states against the control's 51. It is the only one of the three that is a calendar of this page rather than of the sky alone.

The cursor is the ephemeris's own date, moved by the ephemeris's own controls. It is drawn as a one-cell ring stepped by Bresenham rather than a stroked arc, because a stroked arc antialiases and antialiasing is grey. The ring is sigil, opaque, and the same whether the date is today or one the reader travelled to: instrumentation is sigil, and the page's essence budget is one dot, which the field is already spending. Today is said in words — the [today] bracket in the readout, and the canvas's accessible name — because status is text before it is colour.

The whole year is derived once, at mount, from window.MN_derive, and the grid is sized to the year: ceil((offset + days) / 7) columns, which is 53 most years and 54 for a leap year starting on a Sunday. The grid is painted once into an offscreen canvas and blitted when the cursor moves. Nothing here animates, so there is no loop to start and prefers-reduced-motion has nothing to switch off: the reduced picture and the default picture are the same picture.

Six trains from one counter. The step counter runs 0 to 63 and line i is bit i of it, counting from the bottom line up — so each line is a square wave one octave under the line below it, and the strip is a binary counter you can read. The figure decides what happens to each train: a solid line plays it whole, thirty-two steps of sixty-four; a broken line drops the second pulse of every pair, sixteen of sixty-four. The top line is the exception, and it is stated rather than fudged: inside a sixty-four-step pattern its train is one unbroken block of thirty-two, so there is no pair to drop. It halves that one block instead — the same sixteen of sixty-four, the gap in the middle. Every line's density is halved by a break and no line borrows a bit from another. The two kinds of line differ in density and in nothing else, which is the substrate's own law — a mid-tone is a count of marks — applied to time instead of to space.

The tempo is the hour, and the hour is not an hour. A planetary hour is a twelfth of the daylight or a twelfth of the night, so it is sixty minutes twice a year and never in between. MN_derive(now).hourLengthMin supplies it and the step is that divided by 4096. The observer is the ephemeris's own fixed point — the Royal Observatory, 51.4779°N, in UTC — and it is never the reader's location, because a tempo that followed the reader would make the piece a property of the reader rather than of the date. Over 2026 that hour runs from 36.85 to 83.17 minutes, so the step runs from 539.8 to 1218.2 milliseconds: a longer hour is a longer step, so the piece is 2.26 times faster on a midsummer night than on a midsummer afternoon, and the same figure is a different tempo every hour of the year. Measured over 730 hours in evidence.md.

A one-bit stream carries a tone by pulse density. The sounding lines are combined by exclusive-or into a single bit per sample — plus one, minus one, or zero where no line sounds — and that bit goes through a gentle low-pass. There is no oscillator bank and no level: the filter is what turns a count of pulses into a pitch, exactly as a count of marks turns into a tone on the page. The fundamental is the lunar mansion, which divides one octave from 55 Hz as it divides the ecliptic into twenty-eight, and the lines take its first six harmonics. Underneath, a square an octave below the fundamental is opened and closed by the canonical pulse, sin⁴(t·1.55) — the thump is generated one bit too, and the low-pass is the only thing that makes it round.

Nothing exists before the gesture. No AudioContext, no worklet module, no node is constructed until listen is pressed; pressing it again ramps the master down, disconnects every node and closes the context, and the readout says so. An AudioWorklet is used where one is available and a ScriptProcessor where it is not; both run the same generator function, stringified into the worklet module from the same source the strip is drawn from, so a mark and a pulse cannot drift apart. Output is gain-staged to peak below −18 dBFS and measured there, because a page that makes a sound the reader did not ask for at a level they did not choose has taken something.

The cursor is a line, so it is sigil. It marks a position in a printed record. It is instrumentation and it is not alive; the page's one living dot belongs to the field, and a second cyan would be a second claim. You do not start the piece either — the cursor opens where the planetary hour already stands, and pressing listen joins an hour in progress. Under prefers-reduced-motion the strip is drawn once, at t = 3.0, and held: the whole pattern is legible in that one frame, the audio still plays, and nothing moves.