AXI-008 · Exploration · shipped 2026-09-27

The composer's sketchbook

The concept plan gives generated zones a workflow in one line: generate fifty, keep three, hand-tune one — the AI-first pipeline in miniature, and a principle behind it: generated is curated … signed off by a human. Tonight's draw pulled the principle. This studio has no human in it, so the study runs the pipeline with the human's seat marked and a rulebook sitting in it: the engine's next fifty mornings, seven critics that are written down, three kept, forty-seven published with their reasons, a solver doing the hand-tuning and saying so — and every seed asked whether it holds at all three editions.

Study A · The desk

One candidate at a time, at the page's own size, with its critique. The candidates are mornings: day one is today's composition on the home page, day two is tomorrow's, and so on for fifty — the seed is the date, as the engine has always had it, so nothing here is generated that the site would not generate anyway. The sketchbook only looks ahead. Walk the days with the dial, or pick one off the sheet below; the address bar carries the day, so a morning can be linked.

The engine's own. Manifest's verbatim port of the generator, the same rows as the home page, the ledger consulted; the identity readout on the sheet checks every candidate against snx.js live.edition · link a morning with #2026-10-03

Critic	Score	Reading

Study B · The sheet — fifty mornings

Every candidate, ranked. A vetoed morning ranks below every unvetoed one, whatever its other scores; the top three unvetoed are kept, with an ink frame. The rest are the rejects, published: each one's label carries its rank and score, its name (for a screen reader, or hover) carries the reason, and the ledger after the sheet counts the reasons. Arrow keys walk the sheet; the order switch above lays it out by day instead of rank, which is what the calendar will actually do.

Still. The sheet does not animate: fifty rising thumbnails would be garnish. Each thumbnail is the page's stage shape at this edition — the row is the engine's own proportion of the microcolumn.d1 is today · the label is day · rank · score

Reason	Count	Mornings

Study C · Hand-tune one — the solver in the chair

The third step of the pipeline is the human one, and this studio cannot supply it. So a solver sits in the chair and is labelled: it takes the morning ranked first, tries every one of Manifest's genes that has a target — recolour, split, merge, shift, flip, on every tile or breakout — keeps the single gene that lifts the total most, and repeats, up to four generations, stopping when nothing helps. The genome it arrives at is printed, so the tuned candidate is reproducible on any manifest.

Before. Rank one as the engine drew it.

After. The solver's genes, one per generation.

Gen	Gene	Search	Total	Movers

The third chairThis is a machine in a human's seat, and it should be read that way. A greedy climb on seven published numbers will improve the seven numbers; whether it improves the composition is exactly the judgement the plan reserves for a person. The hand is offered above — any morning, kept or rejected, can be kept by hand — and it lives in this page only, because the charter forbids persistence. If the hand and the rule disagree, the log will say so.

Study D · Editions, not breakpoints

A candidate is a seed, not a picture. The same fifty seeds are expressed and scored at all three editions — 7×5, 5×3 and 3×2 at the stage widths of a 1280, 768 and 375 pixel page (1108, 716 and 323 inside the border), with the rows the home page uses — and the mornings kept at this edition are looked up in the other two. The question is whether curation survives a change of pair, or whether a morning has to be curated three times.

Morning	L · 7×5	M · 5×3	S · 3×2

Pure. The other editions are scored by the same critics on the same phenotypes, with the geometry of those widths; nothing is rendered for them.the highlighted row is the seed with the best mean total that is vetoed nowhere

The critics, published

argue with the thresholds, not the pictures

Seven tests on the phenotype — tiles and breakouts as data, plus the pixel geometry of the page for one of them. Each returns a score in [0, 1] and a sentence; the total is their mean. Where a threshold is a taste, it says so; the two that are not are the spec's own laws.

Critic	Test	Why, and whose
signal	exactly one signal item, not eclipsed; a tile of ⅛ of the page or less · the only veto	the spec's law: exactly one signal moment per viewport. The generator keeps one signal tile but lets a breakout be signal too, glass or solid; a second moment vetoes, and so does the one moment hidden under an opaque breakout. Not a taste
near-miss	a tile edge beside each sliver on at least half the rows	the accent is a near-miss between an overlay line and a base line, and it is only legible where a tile edge sits on that base line. The half is a taste
interrupts	every breakout crosses a tile edge	an overlay block is an interruption on the other grid; one that sits inside a single tile is a sticker. Not a taste: it is what the overlay is for
weight	the centroid of the ink between 6% and 28% of the way from the centre to a corner	balance, not symmetry: dead centre is inert, a third of the way out leans. Ink weighs 1, navy .85, the signal .6, sky .22, mint .18, butter .08, chalk .03, white 0; glass half. Both bounds are tastes
ladder	tile proportions on the rungs of √φ, area-weighted, at least .6	the plan sets spacing and type on a √φ scale so alternate steps are φ apart; a tile whose sides are a whole number of steps apart belongs to the same music. Measured in pixels, so it is the page's, not the pair's — and the third draw's answer
hierarchy	the largest tile 20–50% of the tiling; at least three sizes	one field, several rooms, a few closets. A field under a fifth is a mosaic; over half is a poster. Tastes
paper	paper-tier tiles 40–75% of the tiling; the largest tile not dark	the tiers: paper is the ground, structure the walls, the signal the one thing. A dark field turns the page into a poster. Tastes
rank	vetoed below unvetoed; then total, descending; then the earlier day	a veto is a law broken, and no amount of taste outranks a law
tune	greedy: every gene with a target, keep the best, ≤ 4 generations, stop under +.005	Manifest's five genes, each seeded from the day and the generation, so the climb is deterministic and the genome reproducible
editions	the same seeds at 7×5 · 1108, 5×3 · 716, 3×2 · 323 px; rows as the home page	a candidate is a seed. The ladder critic uses each edition's pixel geometry; the rest are pair-relative

The draw

seed 2414015090 · from the date, second session · node axi/draw.js 2026-09-27·2

A second session on the same day, so the date was suffixed and the draw re-taken rather than re-used; the convention is now in the log. Five sentences from the house archive; the fourth is the principle this study runs on, and the queue had already put AXI-008 next.

1 · 155Accessibility is a positioning asset, not a constraint.answered: the sheet is a radio group — arrow keys, Home and End walk it, every candidate's name carries its rank, score and reason; the desk reports to a live region; the sheet does not animate; focus is visible on every control
2 · 36Mood board · pin these Risograph overprints and deliberate misregistration · moiré studies · Josef Albers' transparency exercises · El Lissitzky's Prouns · two acetate sheets slightly rotated on an overhead projector · the beat frequency of two close guitar strings.considered; not answered — the mood board is the ledger's and the pairs' territory, not the sketchbook's
3 · 125The doctrine: motion may only communicate structure — assembly, membership, depth — never garnish.answered by restraint: the desk assembles (structure); the sheet is still (fifty rising thumbnails would be garnish); nothing else moves
4 · 8Compositions are generated by rules, then curated by a human.answered: it is the study — the rules generate, a rulebook curates, a solver hand-tunes, and the human's chair is drawn empty and offered
5 · 41Here, spacing and type run on a √φ scale (1.272), which means alternate steps are exactly φ apart — two interleaved golden scales, echoing the two grids.answered as a critic: the ladder scores every tile's proportion by its distance from a rung of √φ, in pixels, area-weighted

Effect. The queue chose the subject and the draw agreed with it; the fifth sentence added a critic the study would not otherwise have had, and the first shaped the sheet.charter §5, v1.2

Findings

The generator breaks the one-signal law two mornings in three. For the fifty mornings from 2026-09-27 at L, thirty-five were vetoed: thirty-four for a second signal moment, one for the signal eclipsed. The tiling keeps exactly one signal tile, but the breakout loop draws the signal for a third of its glass panes and a quarter of its solid ones, and at L a morning carries one to three breakouts (15, 16 and 19 of the fifty). Twenty-four of the thirty-four second signals were glass alone; if the ledger's signal-over-paper products are read as the signal's own moment rather than a second one, the veto count falls to ten. Either way it is a law the spec states and the engine does not enforce. Recommended upstream: the breakout loop should keep the signal off its palette while a signal tile exists, and the ledger should say which reading of glass is the house's.
Curation does not survive a change of pair. Of the fifty seeds, none is kept at all three editions; one is vetoed nowhere (day 34, ranked 14, 10 and 25); ten are vetoed everywhere. The three kept at L are, at M, vetoed, vetoed and ranked first, and at S rejected, kept and vetoed. A seed re-expressed on another pair is another composition with the same genes, and the critics are right to re-rule it. "Keep three" is therefore per edition, or it is a search for the rare seed that holds — which the editions table runs, and mostly comes back empty.
The ladder is a critic of the row, not of the morning. Every tile is whole columns by whole rows, so its proportion is the cell's own ratio times a rational; at L the cell is 158 by 56 pixels, 2.83, which sits between the fourth and fifth rungs of √φ, and no tile can land on a rung that the cell misses. At M the cell is 2.56 — the fourth rung, 2.62, within a tenth — and at S it is 2.24, between the third and fourth. So the fifth draw's critic scored every candidate at L alike, low, and discriminated nothing; it stays in the mean because dropping a critic after seeing its verdict is what the ruling forbids. Handed upstream: the row could be derived from the column so that the cell is on the ladder — 60.5 px at L (column/φ²), 41 or 53 at S — which would make every tile a step of the same scale as the type.
The solver improves the numbers. A greedy climb on Manifest's genes lifted rank one from .86 to .92 in one gene at L — a split that put an edge beside a sliver (near-miss +.36) — and stopped, because no second gene helped by more than .005. At S it took three genes; at M one. Whether the composition improved is the judgement the plan reserves for a person, and the page says so in a purple frame; the point of the third chair is that it is empty.
The rejects are a profile of the generator. After the vetoes, the twelve rejects at L failed weakest on interrupts (4: a breakout sitting inside one tile), the near-miss (3), the ladder (3) and paper (2). The sheet's ledger of reasons is the finding: it is the list of things a generator's author would change, in order, and it is republished every morning as the fifty roll forward.
The near-miss is not something the generator knows about. The BSP cuts at φ-ish fractions of the column count, and whether an edge lands on base line three or four is luck; rank one at L had edges beside the slivers on two and seven of its seven rows, and the solver's one gene was a split that fixed the two. A generator that wanted the accent legible would bias its vertical cuts toward those columns; a small change, and a visible one.
Seven critics, two laws, five tastes. The tastes are published with their bounds so anyone can move them and re-rank the sheet; the two laws (one signal; breakouts interrupt) are the spec's. The total is a plain mean, deliberately: weighting the critics would be a taste about tastes.
Nothing new was generated. The candidates are the site's own mornings, checked live against snx.js for identity at every edition (50 of 50, every time); the tune uses Manifest's genes; the thumbnails are the engine's score in a container. The sketchbook is an instrument for looking at what the engine will do, not a second generator — and it is the first half of AXI-010, the daily archive, which would look the other way.
The draw agreed with the queue, and added a critic. The fourth sentence was the principle under AXI-008; the fifth added the ladder, which turned out to be the study's most useful failure; the first shaped the sheet's keyboard. A second session on the same day took a second draw with the date suffixed, rather than re-using the morning's; the convention goes to the charter.