Substrate visualisations

Channel	Model's dominant subject	Share
Type	typography	63%
Fashion	fashion	70%
Luminous	installation	52%
Data	data-visualisation	34%
Character	person	31%
Form	sculpture	27%
Space	interior	23%

After Mehretu

A study in a borrowed visual language, built entirely from measured data. Visual study

A dense storm of thousands of small coloured marks and directional flags over arcing line-work. — 9,000 references · 27,113 neighbour arcs · 4,098 flagged above entropy 6.9

Position is the t-SNE coordinate, so two marks near each other are two references that look alike. The line-work is the nearest-neighbour graph — dense knots are territory the studio returned to, open ground is territory it never occupied. Colour is each reference's measured vivid reading, mark size is edge density, and the flags fall only on the most visually complex references, so the storm gathers where the archive is busiest.

Method and caveats 129 words

This is pastiche of a living artist and should carry her name wherever it appears. Julie Mehretu is in this archive — thirteen references, the oldest pinned 2012-03-30 — which is why she came to mind, and is also why the piece needs attribution rather than quiet absorption. It is a study, not a form FIELD invented. Three further studies follow it below, with their controls.

And the density flatters. A nearest-neighbour graph over thousands of points produces a beautiful mesh regardless of what the underlying data is. Draw random vectors this way and it still looks like weather. The form here is borrowed and the content is honest, which is an easy combination to misread — the resemblance is not a claim that the archive is Mehretu-like.

python3 src/viz/mehretu.py --n 9000 --k 6 --mark-scale 1.6 --flags 6.9 --page a2l --dpi 200 --out mehretu-dense

What these studies are, and what their controls found

The attribution, and the measurement that says how little of the archive nine thousand true marks actually carry. Finding

Method and caveats 84 words

This is pastiche of a living artist, and in print it goes with her name on it. Julie Mehretu is in this archive — thirteen references, the oldest pinned 2012-03-30 — which is why she came to mind, and it is also why the piece needs attribution rather than quiet absorption. It is a study in a borrowed language, not a form FIELD invented. The closer the resemblance gets the more that matters, and every sheet below is closer than the one above it.

The control is the most interesting thing on these sheets

The second study shuffled every measurement across the real positions and re-rendered through the same code path, same parameters, same everything else.

Random vectors are obviously distinguishable, and that is the good news. t-SNE over gaussian noise gives a radially symmetric hairball — no lobes, no islands, no bays of empty paper — and every plan block collapses to the same page-sized rectangle, because a channel whose members are scattered at random has no centroid worth drawing. So the layout is carrying real information.

Shuffled real measurements are very nearly not distinguishable, and that is the finding. Keep every true position and permute only the readings across them, and at reading distance it is the same picture. Measured as mean RGB distance to twelve page-neighbours, the real sheet is 10% more colour-clustered than chance on vivid and 22% on brightness_med. Ten per cent is not much to hang a picture on. The substrate carries the information and the storm carries almost none of it — position, density and the channel plan do the work, while the per-mark bindings are true, are measured, and are individually invisible once there are nine thousand of them on one page.

And then the third study's control found the opposite of what it went looking for

The scribbles came out because Marcus did not like them, so the obvious hypothesis was that they were the noise hiding the signal. They were not. Restricting the same test to the marks actually drawn in colour makes the gap narrower, not wider: 9.2% on the second study's colour set and 9.4% on the third's, against 11.3% for all marks. Taking the scribbles out was the right taste call and it is not the evidence.

What widened was straightening. This sheet is mostly made of direction, and direction is what the second study never measured. At its 55° turn, the correlation between mark directions at 6% of the page width was 0.0058 real against 0.0025 shuffled — both flat zero, so the storm was being combed by a current that was statistically invisible on the page. Turn the residual down to 16° and the same statistic is 0.181 against 0.059, a 3.1× gap that holds at about 2.7× anywhere below 30° of turn. The picture agrees with the number: the third study's shuffle is tellable from the real sheet at a glance, where the second study's needed a loupe and a table.

Shuffle control, n=9,000, 24 permutations, 12 page-neighbours	real	shuffled	gap
A · vivid RGB, all marks	162.9	181.4	11.3%
B · brightness_med, the flow scalar	0.2477	0.3179	28.3%
C · vivid RGB, second study's colour set (q0.62)	170.9	186.6	9.2%
C′ · vivid RGB, third study's colour set (q0.45)	167.4	183.2	9.4%
D · flow-field variance	0.004768	0.0002452	19×
E · mark direction, 12 neighbours, turn 55°	0.4362	0.2779	57%
E′ · the same, turn 16°	0.8358	0.7820	6.9%
F · current reach at 6% of the page, turn 55°	0.005765	0.002548	both ≈ 0
F′ · the same, turn 16°	0.1808	0.05872	3.1×

And the fourth study widened it, by drawing less

Both earlier controls turned on statistics a viewer would have to be told about. The fourth study makes the mark vocabulary carry the confidence instead — a sliver only where the twelve nearest marks agree on an axis, a dot where they do not — so the statistic becomes which mark got drawn. Share of references earning a direction: 66.5% real against 51.5% shuffled at n=9,000, a 15-point gap, and 54.8% against 37.1% at n=4,200. It holds across the threshold (+12.6 points at 0.80, +19.5 at 0.90), and the control sheet is visibly stippled where the real one is combed. Statistics A to F reproduce the table below to four significant figures at n=9,000, which is the check that this is the same metric and not a new one wearing the old name.

Two negatives that come with it. Reducing the grey moved A, B and C not at all — it is a decision about looking, not a gain in evidence. And the bold-set clustering statistic does not discriminate between bolding by evidence and bolding by pigment (30.4% against 31.2% tighter than their nulls), so it says both bold sets are spatially structured and nothing about which rule is better.

Working notes 138 words

D says the field is real

A permuted archive still produces a flow grid, but its cell means are sampling noise, and blurring noise leaves almost nothing behind. 19× is not a marginal result.

E and E′ are not a contradiction

E is local. At a low turn it saturates, because blur any field enough and the immediate neighbours agree whatever is underneath. F is the one that matches what a person standing in front of the sheet sees: how far the agreement reaches.

Two conventions, one measurement

The second study quoted A and B as 10% and 22% because it expressed the gap against the shuffled value; the report here expresses it against the real one, giving 11.3% and 28.3%. Same numbers, same finding, and the arithmetic is checkable from the two columns beside them.

Method and caveats 101 words

Two things the control cannot rescue. The currents reach about 6% of the page and no further — at 18% and 35% the correlation is back at zero, so these are local weather systems and not a composition, and the composition is still the t-SNE layout doing the work. And a 3× gap on a directional statistic is a gap a viewer could see once told, not one they would have noticed unprompted. These sheets should hang next to the plain ones rather than instead of them, and the caption on the wall should say study, not finding.

python3 src/viz/mehretu3.py --n 9000 --report

The strata, opened up

The second study at 4,200 references, sparse enough that the layers are still separable by eye. Visual study

A cream sheet with fine grey architectural blocks, translucent washes, arcing line-work and a sparse storm of small ink and colour marks. — 4,200 references · 8,476 arcs · 1,596 colour slivers · 2,600 ink strokes · 887 flags

The saturation gate is what produces the division of labour: every reference below it becomes ink — a scribble traced along the flow, segment count from entropy, curl amplitude from edge density, darkness from local density — and every reference above it becomes a flat tapered lozenge in its own measured vivid, length from its aspect ratio, width from its edge density. At this count you can still see the plan under the storm.

Method and caveats 61 words

The plan is the one structure in the archive a human authored. 309 nodes three levels deep, each channel drawn centred on the t-SNE centroid of its own members and sized to one standard deviation of their spread, so a tight channel is a small box and a diffuse one is a large one. Everything else on the sheet is derived.

python3 src/viz/mehretu2.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --dpi 300 --paper "#efe9dd" --out mehretu2-open

After Mehretu, in strata

Plan, veil, current and storm at nine thousand references — the version the controls were run against. Visual study

A dense cream sheet: grey architectural blocks and ruled section lines under translucent washes, with thousands of ink scribbles and coloured slivers swirling into near-black knots. — 9,000 references · 22,611 arcs · 3,477 colour slivers · 5,521 ink strokes · 1,898 flags

Eleven layers, bottom to top: the plan, the section rules, the veils, the kNN line-work, the long gestures, the ink storms, the colour slivers, the palette bands, the flags, the punctuation, and the returns — a fine open ring on each of the 106 deliberate re-saves, the only mark here that is about an act rather than an image. The near-black knots and the near-white passages come from the density field; nothing is thinned by hand.

Method and caveats 80 words

The turbulence term is doing more than it looks like. Each mark is turned off the current by how far its own reading sits from the mean of the neighbourhood it landed in, clipped to one standard deviation and scaled to at most 55°. So the crossing marks are the local misfits, and they are the only reason this storm has any turbulence in it at all. The control found that 55° is also what makes the current statistically invisible.

python3 src/viz/mehretu2.py --n 9000 --page a2l --dpi 300 --paper "#efe9dd"

The strata at thirteen thousand

The densest version that still proofs on this machine. Visual study

An extremely dense cream sheet where the ink storm has closed over most of the architectural plan, leaving a few pale bays of empty paper. — 13,000 references · 32,634 arcs · 5,048 colour slivers · 7,944 ink strokes · 2,756 flags

Past about nine thousand the storm begins closing over the plan, and what is left to look at is the bays — the passages of near-white paper the archive never occupied.

Method and caveats 79 words

The whole pool exists but has no proof. A 21,491-reference render with 41,454 arcs is written with --no-proof: the SVG is complete and is the file that would go to print, but this box has no rsvg-convert, so proofs go through macOS qlmanage, which parses a 5.5 MB SVG happily and hangs indefinitely on an 11 MB one at any dpi. 13,000 is the densest variant that proofs here, and brew install librsvg is the fix.

python3 src/viz/mehretu2.py --n 13000 --k 5 --bands 620 --gestures 32 --page a2l --dpi 260 --paper "#efe9dd" --out mehretu2-dense

Control: the strata, shuffled

Every true position kept, every measurement permuted across them. This is the null, not the work. Control

A cream sheet almost identical to the nine-thousand-reference study: the same currents, knots and colour weight, with the local colour patches gone. — 9,000 references · 22,611 arcs · 3,477 colour slivers · 5,516 ink strokes · 1,898 flags · CONTROL: measurements shuffled

Set this beside the real sheet and try to tell them apart. Same currents, same knots, same colour weight. The difference survives only on close inspection, as the loss of local colour patches, and when measured it is 10% on vivid and 22% on brightness_med. Every mark on the real sheet is bound to a real reading, and that is what nine thousand true bindings on one page look like against chance.

Method and caveats 62 words

This sheet is here to be compared, not admired. It is drawn at 130dpi rather than 300 because it is a null and not a print file. The reason it is on the page at all is that a study which looks like evidence should have to show what its own noise looks like, at the same size, in the same section.

python3 src/viz/mehretu2.py --n 9000 --control shuffle --page a2l --dpi 130 --paper "#efe9dd"

Control: the strata, from noise

Gaussian vectors reprojected and measurements permuted — the version that is obviously wrong, which is the good news. Control

A cream sheet with a single radially symmetric disc of uniform density, no lobes or islands, and every architectural block collapsed to the same page-sized rectangle. — 9,000 references · 22,592 arcs · 3,477 colour slivers · 5,521 ink strokes · 1,898 flags · CONTROL: random vectors

t-SNE over gaussian noise is a radially symmetric disc with a uniformly dense core: no lobes, no islands, no bays of empty paper, no vortices. Every plan block collapses to the same page-sized rectangle, because a channel whose members are scattered at random has no centroid worth drawing. Beside the real sheet this reads as a hairball rather than as weather, and that is the evidence that the layout is carrying the shape of the archive.

python3 src/viz/mehretu2.py --n 9000 --control random --page a2l --dpi 130 --paper "#efe9dd"

Note one alone: the ground

The second study's drawing moved to a white ground, with every opacity re-pitched. Visual study

The same wobbly storm of scribbles and washes as the second study, now on a white ground where the washes read harder and the grey line-work sits up off the paper. — 4,200 references · 8,476 chords · 1,596 colour slivers · 0 ink slivers · 2,600 scribbles · 887 flags

Opacity does not survive the move. A wash that reads as atmosphere on cream reads as a dirty rectangle on white, and grey line-work that sank into cream sits up off it, so the veils and the plan run at 1.4 and 1.2. Everything else here is the second study's settings exactly, which is what makes this the isolate for note one.

python3 src/viz/mehretu3.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --dpi 300 --veil-gain 1.4 --plan-gain 1.2 --turn 55 --flow-smooth 2 --bow 0.16 --gesture-bow 1.0 --veil-jitter 0.85 --storm-q 0.62 --colour-q 0.62 --curl 1.0 --stream-steps 110 --out mehretu3-white

Notes one and two: the wobble out

Four parameters, not one — the turn, the flow smoothing, the bow on the chords and the torn edge on the washes. Visual study

A white sheet with rectilinear washes, straight chords and long calm combed traces instead of curling scribbles. — 4,200 references · 8,476 chords · 1,596 colour slivers · 0 ink slivers · 2,599 scribbles · 887 flags

Method and caveats 33 words

What “wobbly” actually was

The turn, fixed at 55° in the second study and 16° here — the most consequential number on the sheet. The flow smoothing, 2 blur passes to 5, giving fewer, longer, calmer currents. The bow on the kNN line-work, 0.16 to 0, i.e. straight chords — and the bow was decoration applied to a derivation, which is the worst place to put it. The veil jitter, 0.85 of a cell to 0, i.e. rectilinear washes, and that last one is where most of the shapes being wobbly actually came from, as against the marks.

The streamlines come down too. A divergence-free field traced far enough closes into a ring, and a ring reads as a wobble. That is a property of the integration, not of the archive.

python3 src/viz/mehretu3.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --dpi 300 --veil-gain 1.4 --plan-gain 1.2 --storm-q 0.62 --colour-q 0.62 --curl 0.35 --out mehretu3-straight

Notes one and three: the scribbles out

The gate inverted — every reference is still a mark, and none of them curl. Visual study

A white sheet made entirely of tapered lozenges, coloured above the saturation gate and graphite below it, still swirling at the second study's turbulence. — 4,200 references · 8,476 chords · 2,315 colour slivers · 1,885 ink slivers · 887 flags

In the second study one saturation gate decided whether a reference got a scribble or a sliver, and it sent 62% of the archive into scribbles. Here there are two: --storm-q, at 0 by default so nothing scribbles, and --colour-q at 0.45, which now decides only what a sliver is made of — its own vivid above, graphite darkened by local density below, drawn finer at 0.80 length and 0.62 width so the ink carries the current without competing with the colour. The division of labour survives, in one family of marks instead of two.

Method and caveats 32 words

This is note three with the wobble deliberately left in, so the two notes can be told apart on the page. It is not the recommended sheet; it is the isolate.

python3 src/viz/mehretu3.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --dpi 300 --veil-gain 1.4 --plan-gain 1.2 --turn 55 --flow-smooth 2 --bow 0.16 --gesture-bow 1.0 --veil-jitter 0.85 --stream-steps 110 --out mehretu3-slivers

All three notes, which are the defaults

The third study as it stands: white ground, calm currents, no scribbles, and a null that is visibly a different picture. Visual study

A white sheet of long combed slivers running in wide calm currents, coloured where the archive is saturated and graphite where it is not, over rectilinear washes and straight chords. — 4,200 references · 8,476 chords · 2,315 colour slivers · 1,885 ink slivers · 887 flags

Method and caveats 88 words

The argument for this version is note two, not note three

Turning the turbulence down from 55° to 16° is what makes the realness of the field visible; taking the scribbles out is what makes it legible. The measurement says the same thing bluntly: dropping the storm does not widen the gap against the null at all, and straightening the marks takes the direction correlation at 6% of the page from 0.0058 against 0.0025 — both flat zero — to 0.181 against 0.059. The full table is in the framing above.

What would make this wrong. The currents reach about 6% of the page and no further, so they are local weather systems rather than a composition, and the composition is still the t-SNE layout doing the work. A 3× gap on a directional statistic is something a viewer could see once told, not something they would notice unprompted. And the flow field is a smooth scalar: integrate any smooth scalar and you get currents, which is why the shuffle control has to be on the page beside this.

python3 src/viz/mehretu3.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --dpi 300 --veil-gain 1.4 --plan-gain 1.2 --out mehretu3

The lowest fifth kept as currents

The scribble layer turned back on at a low curl, where it stops being a curl and becomes a long combed trace. Visual study

A white sheet where the least saturated references are drawn as long trailing traces through the slivers, making the direction of the flow field visible as continuous lines. — 4,200 references · 8,476 chords · 2,315 colour slivers · 1,046 ink slivers · 838 scribbles · 887 flags

The 838 least saturated references become traces rather than lozenges — segment count from entropy, amplitude from edge density, darkness from local density, at a fifth of the second study's curl. It is the one variant where the flow field is directly legible instead of inferred from the way the slivers line up.

Method and caveats 45 words

The layer is kept for the control's sake as much as for this sheet. Leaving the storm in the code is what lets the third study's code path draw the second study's picture, which is the only reason the two controls are comparable at all.

python3 src/viz/mehretu3.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --dpi 300 --veil-gain 1.4 --plan-gain 1.2 --storm-q 0.20 --curl 0.25 --out mehretu3-currents

The same sheet on the house paper

The third study on this repo's own stock, where the second study's opacities are right again. Visual study

The same calm sliver drawing on the warm grey house stock, where the washes sink back into the paper and the grey plan lines settle into it. — 4,200 references · 8,476 chords · 2,315 colour slivers · 1,885 ink slivers · 887 flags

Identical marks, identical counts, no gain applied. --veil-gain and --plan-gain at 1.0 is what the second study's cream values mean, so this is the sheet that shows how much of the difference between the two studies was the paper rather than the drawing.

python3 src/viz/mehretu3.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --dpi 300 --paper "var(--panel)" --out mehretu3-stock

Control: the third study, shuffled

The same permutation as the second study's control, and this time you can see it. Control

A white sheet of slivers with no visible currents at all: the marks scatter in every direction instead of combing into wide flows. — 4,200 references · 8,476 chords · 2,315 colour slivers · 1,885 ink slivers · 887 flags · CONTROL: measurements shuffled

Set this beside the real sheet and the currents are simply absent. That is the strongest claim these sheets have: not that the drawing is beautiful, but that its own null is now visibly a different picture, where the second study's null needed a loupe and a table to separate.

Method and caveats 64 words

It is still only the direction that separates them. Colour clustering on this sheet is 9.4% above chance, which is no better than the second study managed and slightly worse than its own colour set scored there. If someone looks at these two sheets and sees the same picture, they are looking at the colour, and the colour is not where the evidence is.

python3 src/viz/mehretu3.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --veil-gain 1.4 --plan-gain 1.2 --control shuffle --dpi 130

The notes read literally

Dots mixed with slivers, the grey pushed down, and the bold set chosen by pigment. Visual study

A white sheet where combed passages of tapered slivers give way to stippled fields of round dots, with a quiet graphite majority and a small set of saturated marks drawn at full strength. — 4,200 references, none dropped · 2,310 slivers where direction is supported, 1,890 dots where it is not · saturation median 0.14, 68% below 0.25 · 2,604 in the quiet ground · 426 bold, the top saturation decile (≥0.54)

Method and caveats 53 words

Three registers, and the census stays intact

The ground is everything below the colour gate — graphite, at reduced opacity and 0.78 of the previous size, closer to the paper and no longer competing. The middle is drawn in its own measured vivid at normal weight. The bold set is a hard top decile at real strength: 1.7× the length, 1.5× the width, full opacity, and a fine ink keyline on the dots so they sit forward off the ground rather than in it.

Nothing is thinned to achieve this. All 4,200 references are still drawn, and the caption prints the saturation distribution on every sheet — median 0.14, 68% below 0.25 — so a reader is told the quiet ground is the majority rather than left to infer that the archive is more colourful than it is.

The bold set here is chosen by pigment, which is a property of the image rather than of the studio. That is the honest reading of the note as written, and it is the better-looking sheet. The next one keeps the drawing identical and swaps the rule, which is where the argument is.

python3 src/viz/mehretu4.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --dpi 300 --bold-by saturation --ground-by palette --out mehretu4-mixed

The same drawing, bolded by evidence

The loud marks chosen by what FIELD did with a reference, not by how saturated it is. Visual study

The same combed and stippled sheet, with a different tenth of the marks drawn loud — scattered in regions rather than following the saturated colours. — 4,200 references, none dropped · 2,310 slivers, 1,890 dots · 2,358 in the quiet ground · 420 bold by evidence — re-saves, cross-channel reuse, likeness to published work

Method and caveats 92 words

What “evidence” is scored from, and it is three acts rather than three pixel readings

The returns, weight 3.0 — pool.protected(), a pin whose parent pin is also one of FIELD's own, meaning the studio deliberately saved the same picture a second time. 106 are live, and the deduplicator had all 106 in its drop list until src/pool.py put them back. This is the strongest evidence in the archive that somebody came back to something.

The reuse, weight 1.5 per extra channel — a reference filed into more than one channel was carried from one piece of work into another. 1,005 of the pool sit in two or more, 236 in three or more.

The resemblance to what got made, weight 1.0 on the percentile — maximum cosine similarity to any of the 1,086 embedded images of FIELD's own published work. Not proof of influence, but the closest available measurement of “this turned up in the output”. Custody and provenance are deliberately kept out of the score and used for the ground instead, so the two halves of the sheet are not saying the same thing twice.

The measurement cannot settle which bold rule is better. Against a null that permutes the labels across the same positions, the evidence bold set clusters at 30.4% tighter than chance — and the saturation bold set clusters at 31.2%. So the number says both bold sets are spatially structured, which for saturation is unsurprising since CLIP position and colour are correlated, and for evidence is a real finding. It does not say evidence is the better sheet. That argument has to be made on what the marks mean, or not at all.

python3 src/viz/mehretu4.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --dpi 300 --bold-by evidence --ground-by palette --out mehretu4-evidence

The ground carrying a count

The wash inked where the studio inherited more than it found. The print pick. Visual study

The same sheet with a large shaded ground: broad regions inked where most of the references were taken from other people's boards, pale where the studio found things first-hand. — 4,200 references, none dropped · 2,310 slivers, 1,890 dots · 2,358 in the quiet ground · 420 bold by evidence · ground wash = the cells more second-hand than the archive's own 63%, from 2,886 references with readable provenance

Method and caveats 48 words

The one element on these sheets that does not need a control

The ground replaces the second and third studies' veils, which averaged five-colour palettes into atmosphere and, by the second study's own control, carried nothing. Here each cell takes ink in proportion to the share of its Pinterest references that were taken from somebody else's board rather than pinned first-hand from a page, and it takes ink only where that share is above the archive's own rate. Across the pool: 6,969 second-hand against 4,163 first-hand, 62.6% of the pins with readable provenance. So the shaded regions are where the studio collected collections and the pale regions are where it found things itself.

A count survives a shuffle by construction rather than by measurement, which is why this layer needs no null. If one thing in this family is meaningful in the sense the note asked for, it is this — and the honest way to say so is that it stopped being a Mehretu move and became a bar chart with the bars laid over a map. Cells with too few readable pins are left as paper rather than guessed at.

Everything else on the page is still a study in a borrowed language with real numbers underneath it. The composition is the t-SNE layout doing the work, as it has been since the first sheet, and the currents still reach about 6% of the page and no further.

python3 src/viz/mehretu4.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --dpi 300 --bold-by evidence --ground-by secondhand --out mehretu4-secondhand

The same ground, over custody

Where FIELD holds the file itself, on the house stock. Visual study

The same drawing on warm grey stock, with the ground shaded where the studio holds the image files rather than linking to somebody else's server. — 4,200 references, none dropped · 2,310 slivers, 1,890 dots · 2,358 in the quiet ground · 420 bold by evidence · ground wash = the cells where FIELD holds more of the files than its own 19%

Same construction as the second-hand ground, over a different count: whether a reference has a blob_url, meaning the studio holds the file rather than a link to somebody else's server. 3,158 of 21,491, 19.1%. The biography sheet measured what the other 81% is worth — two in five of what was pinned in 2012 no longer answers.

Method and caveats 44 words

A wash needs more gain on the house stock than on white, which is the same lesson the third study learned moving the other way: opacity does not survive a change of ground, and the fix is a gain rather than a redraw.

python3 src/viz/mehretu4.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --dpi 300 --bold-by evidence --ground-by custody --paper "var(--panel)" --veil-gain 1.15 --out mehretu4-custody

Control: the fourth study, shuffled

Stippled where the real sheet is combed — the first control in the family a viewer can separate without being told what to look at. Control

The same sheet with far more dots and far fewer slivers, and the surviving slivers pointing in scattered directions rather than gathering into passages. — 4,200 references, none dropped · **1,643 slivers, 2,557 dots** · 2,358 in the quiet ground · 420 bold by evidence · CONTROL: measurements shuffled

At a fixed support threshold of 0.85, the real sheet earns a direction for 66.5% of its references and the shuffled one for 51.5% — a 15-point difference at n=9,000, and 54.8% against 37.1% at the 4,200 these variants are drawn at.

Method and caveats 89 words

It holds across the threshold too: +12.6 points at 0.80 and +19.5 at 0.90, so it is a property of the field rather than of one lucky cut.

Why that is a gain over the third study rather than a restatement of it. That sheet's best claim was a 3.1× ratio on a directional correlation — true, checkable, and something no viewer would notice unprompted. This converts the same underlying fact into the vocabulary of the drawing: the sheet stops asserting a direction where its own statistics do not support one, and the share of the page where it declines to is itself the measurement. Nothing was added; something was withdrawn, and the withdrawal is what separates.

The rest of the table did not move, and that is stated rather than buried. At n=9,000 statistics A through F reproduce the third study's published numbers to four significant figures, which is the check that this is the same metric and not a new one wearing the old name. Reducing the grey moved none of them. Pushing 2,358 of 4,200 references into a quiet ground is a decision about looking, not a gain in evidence, and colour was never legible as information on any sheet in this family.

python3 src/viz/mehretu4.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --bold-by evidence --ground-by secondhand --control shuffle --dpi 130
# python3 src/viz/mehretu4.py --n 9000 --report   # like-for-like with the second and third studies

Diagnostic: every mark a sliver

The support threshold at zero, which is the third study's vocabulary as a limiting case. Instrument

Every reference drawn as a directional sliver, so the whole sheet reads as combed whether or not the underlying field supports it. — 4,200 references, none dropped · **4,200 slivers, 0 dots** · 2,358 in the quiet ground · 420 bold by evidence · `--support 0`

This is not a study. It is the instrument with one knob at its stop, printed so the vocabulary split can be checked rather than trusted. Every mark asserts a direction, including the half of them whose neighbourhood does not agree on one, which is exactly what the third study did.

python3 src/viz/mehretu4.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --dpi 300 --support 0 --out mehretu4-vocab-0

Diagnostic: every mark a dot

The support threshold past one, which is the first study's vocabulary as a limiting case. Instrument

Every reference drawn as a round dot with no orientation at all, so the sheet reads as a stippled field with no currents in it. — 4,200 references, none dropped · **0 slivers, 4,200 dots** · 2,358 in the quiet ground · 420 bold by evidence · `--support 1.1`

Every mark a dot, none claiming a direction. The first study's vocabulary as a limiting case, against every mark a sliver at the other end.

python3 src/viz/mehretu4.py --n 4200 --k 4 --bands 240 --gestures 20 --plan-min 25 --page a2l --dpi 300 --support 1.1 --out mehretu4-vocab-1.1

Frontier or density

Fourteen years of saving, in the order it happened, asking of every reference how close the nearest earlier one was. Finding

Twelve columns of small images left to right, each column coarse at the top and fine at the bottom; the coarse band does not thin out towards the right. — **11,238 Pinterest saves** · twelve blocks of 1,000 · frontier 16.6% over the whole run · mean nearest-earlier cosine 0.8398

The frontier has not closed. The most recent 238 saves open new ground at 26.9%, against 9.7% for the same references shuffled into a random order. Read the two plates together or not at all: over the whole run they are indistinguishable (16.6% against 16.7%, mean nearest-earlier 0.8398 against 0.8389), and the entire difference is in the shape.

Working notes 210 words

The pool

pool(), aesthetic filter on. This asks what the studio looks at, so a link card pasted into a thread cannot be allowed to count as a save with an eye behind it — it is maximally unlike the photographs around it and would read as frontier every time. 11,779 dated Pinterest saves become 11,238.

Why the control is the whole point

Novelty yield falls by arithmetic alone: the thousandth save has 999 things to resemble, the eleven-thousandth has 10,999. Any “distance to nearest earlier neighbour” curve slopes upward in any growing collection. The shuffled plate is what pure accumulation looks like.

It starts below the null and ends above it

First block: 26.9% real against 33.7% shuffled. Eleventh: 13.4% against 12.0%. Last: 26.9% against 9.7%. Inferred: real collecting arrives in bursts, so the early years produce fewer frontiers than a shuffle; Pinterest capture then collapsed after 2021 and what remained was sparse and various.

Distance is not originality

A frontier save here is unlike the archive. That is a statement about the collection, not about the picture. Nothing on this sheet is ranked by quality. Moving --frontier from 0.78 changes every percentage; the gap between the two plates holds its direction across the range.

Method and caveats 91 words

What would make this wrong. The pool is Pinterest only, because it is the one surface whose dates carry a real chronology — and its capture fell from 2,742 saves in 2020 to 207 in 2026, so the right-hand columns are drawn from a thinning surface. If FIELD's newest looking moved to Basecamp and Figma, and that looking is more familiar than what it pins, this sheet cannot see it. The claim is about the Pinterest habit, not about the studio's whole eye. The last column is 238 saves, not 1,000.

python3 src/viz/d01_frontier.py --dpi 200

The same archive, shuffled

The control plate: identical references, order destroyed, so accumulation can be subtracted from the finding. Instrument

The same twelve columns of small images, but the coarse top band shrinks steadily from left to right until the last column is almost entirely fine. — **the same 11,238 saves, shuffled** (seed 0) · frontier falls 33.7% → 9.7% · whole run 16.7%, mean nearest-earlier 0.8389

Working notes 93 words

This is what a closing frontier actually looks like, and it is what you get for free from a collection simply getting bigger. The shuffled rate falls by two thirds from beginning to end, monotonically. The real archive's does not fall at all. Publishing the finding without this plate beside it would be publishing arithmetic as insight.

Why it is a sheet and not a footnote

The two plates are the same instrument at the same scale, so the comparison is made by eye rather than taken on trust. A reader who does not believe the percentages can still see that one coarse band thins and the other does not.

What is identical between them

The same references, the same 120 k-means regions, the same 0.78 threshold, the same seed. Only the order changes. Territory coverage saturates in both: 120 of 120 regions are occupied by save 10,000 either way.

Method and caveats 56 words

What would make this wrong. One shuffle is one draw. The seed is fixed at 0 so the plate is reproducible, but a single permutation is not a distribution — the block-level percentages here carry sampling noise that the printed figures do not show. The direction of the effect is stable; the individual cells are not.

python3 src/viz/d01_frontier.py --null --dpi 200

The bridges

References whose two nearest predecessors sat in different regions — and the count of them turns out to be an accumulation artefact. Failed hypothesis

A sparse sheet of images in save order, many fewer than the frontier plate, thinning markedly from left to right. — **1,352 bridges** of 11,238 saves · against **1,340** in the shuffled order · 266 in the first block of 1,000, 76 in the eleventh

Nothing about the real sequence produces more connections than chance. 1,352 bridges as collected against 1,340 shuffled, and both decline the same way for the same finite reason: there are only so many pairs of regions to join, and the early saves join them.

Working notes 85 words

The hypothesis that the studio's actual order builds bridges was tested and it is a null.

Why print it anyway

Which pairs got joined is interesting. How many is not. The sheet is kept as a picture of the joins and the count is retracted in the caption rather than quietly dropped from the file.

The weakest of the three classes

A bridge is defined against a k-means partition of a continuum, so it is a claim about an analytical boundary rather than about anything the archive declares. Frontier and infill survive the region count moving; the bridge set does not.

Method and caveats 57 words

What would make this wrong. Beyond the region-boundary problem: the decline is confounded with the block size at the right-hand edge, where the last block holds 238 saves rather than 1,000 (13 bridges real, 22 shuffled — too few to read). Compare the eleventh block instead, 76 against 63, which is the same story at full width.

python3 src/viz/d01_frontier.py --mode bridges

Paths of attention

The archive as the sequence it was looked at in — 385 saving sessions drawn as walks, tunnelling at the top and scanning at the bottom. Finding

Rows of small images running left to right, each image sitting higher or lower than the last; upper rows are almost flat, lower rows zigzag steeply. — **2,665 sessions** at a 60-minute gap · 385 of eight saves or more, covering 6,233 saves · step cosine 0.7024 against a 0.6682 shuffle

The order of saving carries real information, and its grain is minutes. Consecutive saves sit at 0.7024; the same sessions with their own order shuffled sit at 0.6682; random pairs from anywhere in the archive at 0.5971.

Working notes 163 words

356 of 385 sessions beat their own shuffle, sign test p = 5.0e−73.

The pool

pool(), aesthetic filter on, via src/d01_spine.py. A session is a claim about looking, and a run of pasted link cards is a burst of activity with no looking in it. 11,238 saves.

Two effects, different sizes

A session is about something (0.668 against 0.597 for the archive at large) and that is most of the coherence. The order within a session adds +0.034 on top, and that is the part only a walk can produce.

It decays with the clock

Under 10 seconds, 0.7590 (n=513). 10–60 seconds, 0.7314 (n=2,178). One to five minutes, 0.6837 (n=1,854). Five to thirty, 0.6544 (n=1,081). Over thirty, 0.6428 (n=222). Dropping every sub-10-second pair leaves 0.7314, still well clear of the 0.6682 shuffle, so this is not an artefact of batch saving.

What the rows are

Ordered by mean step size and described, never scored: 125 circling, 97 tunnelling, 97 scanning, 46 drifting, 20 mixed. Distance is a change of subject, not a better idea.

Method and caveats 106 words

What would make this wrong. One account is not provably one pair of eyes — Substrate does not record who pinned, so two people working in the same hour are drawn as one walk. The 60-minute gap is a choice with nothing in the data to anchor it: 30, 60 and 120 minutes give different session counts over the same saves, and the median gap between consecutive saves is 67 seconds with no second mode to cut at. And vertical position is a step size, not a direction: a session oscillating between two territories and one walking steadily away both draw as a jagged line.

python3 src/viz/d02_walk.py --dpi 200

One session, whole page

A single hour of saving at a size where the individual pictures can be read. Visual study

One row of larger images across a landscape page, rising and falling, drawn from a single saving session. — **session 3** of 385 · one row across A3 landscape · the same measurement as the full sheet, one row of it

Method and caveats 38 words

The aggregate sheet carries the effect and hides the thing itself. At this size you can see what a step of 0.70 actually looks like between two pictures, which is the only way to judge whether the measurement means anything.

What would make this wrong. It is one session, chosen by index rather than by argument, and it carries no statistical weight at all. Nothing on this page should be generalised; the numbers live on the full sheet.

python3 src/viz/d02_walk.py --one 3 --page a3l

Diverge, then converge

Twenty-three project boards drawn as breathing bands, testing the received story that research opens wide and then narrows. Failed hypothesis

Small multiples of horizontal bands filled with tiny images, each band swelling and narrowing along its length; no consistent shape across the panels. — **23 project and pitch boards** of 100+ usable Pinterest references · narrowing at p < 0.05: 2 · widening: 0 · mean correlation **−0.039**

Project research does not systematically converge. Against 300 shuffles of each board's own saves, two boards narrow significantly where about 0.6 would be expected by chance, none widen, and the mean correlation of spread against position is −0.039 — a tilt nobody could act on.

Working notes 200 words

Diverge-then-converge is a story the archive declines to tell.

The pool

pool(), aesthetic filter on, via src/d01_spine.py. The board count fell from 26 to 23 when the filter was applied: three boards no longer clear 100 usable references, which is itself worth knowing about how much of some project research was screenshots.

The two that do narrow

242 Swisscom Energy Air — Dance! (r = −0.856, p = 0.007) and PG2 (r = −0.657, p = 0.033). Real individual shapes, not evidence of a studio process. A third, P108 UAE Expo Pavillion, lands at p = 0.050 and is not counted.

The second test pushes it further

How often does a board reach into a region it has not used yet, first third against last? That rate falls by +0.292 across a project — and falls by +0.366 in the shuffled order. Any accumulating set runs out of unused regions whether or not anyone was converging, and the arithmetic accounts for more than the whole observed decline.

A smooth curve is not a better process

Some of these projects needed late disruption and some needed sustained focus. No panel is scored, and --sort number puts them in plain project order for anyone who finds the shape-ordering suggestive.

Method and caveats 98 words

What would make this wrong. A Pinterest board is not a project's research — it is the part that got pinned, and a band that thins may mean the looking moved somewhere Substrate cannot see. Save order is not necessarily thinking order: boards get reorganised and old pins get added to new boards, which trimming to the densest 80% of each date range reduces but cannot remove. And 23 boards is a small n; a larger corpus could turn this null into a weak finding in the narrowing direction. It could not turn it into the textbook curve.

python3 src/viz/d03_breathes.py --dpi 200

Persistent, dormant, returning

Forty visual regions drawn as pulse lines across fourteen years, asking which ideas died and which were only quiet. Finding

Forty horizontal rows, each a broken thread of small images across a timeline, most running nearly continuously with occasional bare gaps. — **11,238 dated Pinterest references** in 40 of 40 regions, by quarter · continuously active **31 of 40** · flashes 0 · returning after a year of silence 25

FIELD does not abandon visual ideas, and when it comes back it comes back to the same pictures. Not one of the 40 regions flashes and dies. 25 go quiet for a year or more and return — and the centroid of what was saved after a dormancy matches the centroid before it at a median cosine of 0.912 (range 0.705 to 0.980).

Working notes 133 words

The poetic reading — a dormant idea returning transformed — is not what this measures. The returns are not transformed.

The pool

pool(), aesthetic filter on, via src/d01_spine.py. Persistence is a claim about a continuing interest; a region kept alive by recurring screenshots is not one. 11,779 dated saves become 11,238.

The fifth temporal null

This sits alongside four already established in the repo — no colour drift, no complexity drift, stable CLIP year-centroids, no subject or mood trend — and says the same thing from a fifth direction. The eye did not change, and it did not drop anything either.

A burst needs a floor

At the default floor of 1, a region with a single save in a quarter is drawn as active. That is the maximally generous setting about persistence, which is the conservative direction for a sheet arguing that things stop. --floor raises the bar.

Method and caveats 98 words

What would make this wrong. The clock is Pinterest's, and Pinterest capture fell from 2,742 references in 2020 to 207 in 2026 — so the right-hand edge of every row thins for a reason that has nothing to do with the idea. A region going quiet after 2021 may be a region going quiet, or the studio's collecting moving to Basecamp and Figma where this sheet cannot follow it, and nothing here separates those. Regions are k-means cuts through a continuum; --k moves every lifespan on the sheet. Nothing here is ranked: a flash is not a failure.

python3 src/viz/d05_halflife.py --dpi 200

The same, with every region's biggest board removed

The guardrail plate: if persistence is really the shadow of a few large jobs, this is where it collapses. Instrument

The same forty rows, visibly sparser, but with the same continuous threads and the same gaps. — **9,994 references left** after deleting each region's largest board · continuously active **31 of 40** · flashes 0 · returning 26 · centroid cosine across a dormancy, median **0.907**

Working notes 116 words

The persistence survives the surgery. A region lighting up in 2018 and again in 2023 looks like a recurring interest and could just as easily be two unrelated jobs that needed similar pictures. Delete every region's dominant board and nothing moves: 31 still continuous, still no flashes, one more region returning, and the before-against-after cosine falls from 0.912 to 0.907. Only one region in forty is half-supplied by a single board in the first place (X Activist Aesthetics, 63%).

Why this is not a footnote

It is the honest version of the sheet, and any claim made from the first plate has to survive it. In this archive it does. A row that emptied here was one project all along and the recurrence was never a cycle.

A known reporting bug, stated

In this plate the terminal's largest board column divides the removed board's original count by the reduced total, so it prints impossible shares (X Activist Aesthetics reads “169%”). Nothing on the sheet uses that column and no number in this caption comes from it — the guardrail figure quoted above is from the full run. It is recorded here rather than fixed silently.

Method and caveats 46 words

What would make this wrong. Removing the largest board removes about 11% of the references and does not remove the second-largest, so a region carried by two big jobs still survives this cut. It is a test against single-board dominance, not against project-driven persistence in general.

python3 src/viz/d05_halflife.py --without-top --dpi 200

Research personalities

Thirty-three boards with their size divided out, so what is left is the shape of the research rather than the amount of it. Finding

Thirty-three rows of equal width, each made of runs of similar images; upper rows are long uniform runs, lower rows are mosaics of many short runs. — **51 boards** of 100+ · 21,491 references in 60 regions · entropy at n=100 runs **2.36 to 3.44** against 4.09 for a perfectly even board

Research structure varies far more than volume does. Two boards of almost identical size sit at opposite ends of the range: 242 Swisscom Energy Air — Dance! at 105 references occupies 24 regions with 46% of itself in one, and Culture at 113 references occupies 41 with no region above 11%.

Working notes 203 words

The pool

pool(), aesthetic filter on, via src/d01_spine.py. A board's breadth is a claim about the eye, and a board padded with link cards would read as broad for a reason that has nothing to do with looking. 26,214 references become 21,491; 63 qualifying boards become 51.

The size trap, and what survived it

Entropy rises with sample size for pure arithmetic reasons, so every board is subsampled to the same 100 references before it is measured. Raw, entropy against board size is r = +0.444; controlled, r = +0.177. The control removes about 60% of the relationship and does not remove all of it. Read the ordering as “structure, partly confounded with scale”.

The regions are archive-wide

One k-means over all 21,491 references, not one per board. A board's breadth has to be measured against the archive's territory, not against a partition drawn around the board itself.

The modes are not a ranking

Exploratory 13, fragmented 7, obsessive 5, tightly resolved 3, panoramic 2, bifurcated 2, concentrated 1. The names print to the terminal and appear nowhere on the sheet; rows are ordered by a measurement, and nothing computes a score. Obsessive is not worse than panoramic.

Method and caveats 101 words

What would make this wrong. Regions are cut, not found — --k changes every number here, and while the ordering of boards is stable from k = 40 to 120 the individual region counts are not. A board is not a project but the pinned part of one, and several channels under Projects are places rather than jobs (Beirut, Hong Kong, Munich, Seoul), which is a known pollution in this repo's channel data. And repetition is measured on what survived deduplication: the genuinely duplicated saves, which would be the strongest evidence of obsession, are the ones the pool throws away first.

python3 src/viz/d04_personality.py --dpi 200

The borderlands

Rows running from one subject through the seam into another — and the finding that “ambiguous” is not one property but four unrelated ones. Finding

Twelve rows of images, each running from one clear kind of picture on the left through an unreadable middle to a different clear kind on the right. Nothing is labelled. — **18,113 references** with a subject and both CLIP models · twelve confusable subject pairs · neighbourhood entropy 1.75 bits against 3.32 on shuffled labels

“Ambiguous” is four different sets, and one of them shares nothing with the others. Of the top 48 by each definition: neighbourhood entropy and self-disagreement share 37, entropy and cross-model disagreement share 27, self and cross-model share 33 — and “sits on a boundary between two clusters” shares 0 of 48 with all three. A reference near a seam in the partition is not the same object as a reference whose visual company cannot agree what it is.

Working notes 239 words

The pool — and why this one keeps the screenshots

pool(aesthetic_filter=False). This is the one sheet in the set that is about the method as much as the pictures: its stated job is to separate productive ambiguity from analytical failure. Filtering the failure cases out upstream would let it claim a cleanliness it had not earned, and the guardrail would stop being checkable. The filter off, the population goes up — 19,209 to 19,315 — because the 106 deliberate re-saves come back and nothing is removed. 18,113 of those carry a subject label and both models.

Explicit refusal measures nothing

The vision model returns unclear on 40 subjects, 0.22% (and 340 mediums, 1.88%). That is a measure of how reluctant a language model is to refuse a multiple-choice question, not a measure of ambiguity. It is reported and set aside.

What the archive actually confuses

The closest subject centres: typography | graphic 0.980, texture | abstract 0.977, object | sculpture 0.974, interior | installation 0.973. The first pair only exists because the aesthetic filter is off here, which is the decision above doing visible work.

Two more numbers worth having

A reference's own label differs from its neighbourhood's majority 44.0% of the time, and the two CLIP models' neighbourhoods vote for different subjects 27.4% of the time. Nothing is labelled on the sheet: either the two ends look like two different things and the middle looks like neither, or the row has nothing to say.

Method and caveats 97 words

What would make this wrong. Neighbourhood entropy is bounded by the label set — with 27 subjects of wildly uneven frequency, a reference sitting among the three biggest scores higher than one among the three smallest for no reason to do with the picture. That is why the shuffled control (3.32 bits, against a 4.32 ceiling) is printed beside the real 1.75. And a borderland between two model-assigned labels is not a borderland in the world: these are the seams of one taxonomy chosen by one model, and a different list of subjects would put them elsewhere.

python3 src/viz/d08_ambiguous.py --dpi 300

Ambiguous references are not reused more

The most attractive sentence in the brief, tested against the archive's own filing — and the answer is a null on a test that could never have been strong. Failed hypothesis

Forty-eight large images, each one hard to name: surfaces that could be landscape or material, objects that could be sculpture or product. — **the 48 most ambiguous** by neighbourhood entropy, after the thinness floor · 17,113 references sit in at least one channel, **mean 1.08 channels each**

Null, and in the opposite direction to the one predicted. The claim was that references resisting a settled identity get reinterpreted across more briefs. Correlation between neighbourhood entropy and channel count: r = −0.050.

Working notes 106 words

The most ambiguous tenth sits in 1.05 channels, the least ambiguous tenth in 1.12. Filed in more than one channel at all: 3.7% of the most ambiguous tenth against 8.3% of the least, 5.8% overall.

And the test was underpowered before it started. The mean reference sits in 1.08 channels. There is almost no reuse in Substrate to correlate anything with, so no version of this test could have produced strong support in either direction. Both halves of that have to be said: the sign is against the claim, and the instrument was too blunt to have settled it.

Why the sign still counts for something

It is consistent across three different cuts of the same data — the correlation, the mean channel count by decile, and the share filed more than once. A weak test that points the same way three times is not evidence for the claim.

Filing is not generativity

The deeper problem, and it is not fixable from inside this archive. A reference that appeared in one board and shaped a whole job looks identical here to one nobody opened twice. What can honestly be said is narrow: this archive holds no evidence for the claim, and a little against it.

Method and caveats 87 words

What would make this wrong. Channel membership is the only reuse signal Substrate records, and it is a filing act rather than a use. If FIELD reuses references by looking at them again rather than by re-filing them, this test is measuring the wrong thing entirely — and at a base rate of 1.08 channels per reference, that is the more likely reading. The correlation is also computed over references that carry a subject label and both models, which is 18,113 of the 19,315 in the pool.

python3 src/viz/d08_ambiguous.py --plate fog --dpi 300

Thin, not ambiguous

The guardrail made visible: the references the thinness floor throws out before anything is called interesting. Instrument

A sheet of near-blank scans, blown-out photographs, crushed dark frames and tiny low-resolution thumbnails. — **the thinness floor catches 23.9%** of the population · **12 of the top 48** by raw entropy are thin rather than ambiguous · bottom-decile greyscale entropy, or flat, or blown, or under 200px

Working notes 98 words

A blank scan is genuinely unlike everything and genuinely unclassifiable, and it is not what direction 08 is about. Without this floor, a quarter of the “most ambiguous” list is bad files. The floor is drawn rather than asserted, so the judgement can be checked by eye instead of trusted.

Why it is drawn at all

The brief's instruction was to separate productive ambiguity from analytical failure “where possible”. A filter that removes a quarter of the candidate set is a large intervention, and a sheet that applied it silently would be making an aesthetic judgement while claiming to make a measurement.

What it costs

Four crude tests stand in for one hard judgement. A deliberately flat, low-contrast photograph is indistinguishable to them from a bad scan, so some real references are on this page. That is the direction the error runs, and it is the conservative one.

Method and caveats 69 words

What would make this wrong. The four thresholds are chosen, not derived — bottom-decile entropy, near-zero edge density, blown or crushed median brightness, under 200px on the short edge — and moving any of them moves the 23.9%. The floor also cannot see a well-exposed, sharply-focused image that is nonetheless a screenshot of text, which is the case the aesthetic filter would catch and this sheet deliberately leaves off.

python3 src/viz/d08_ambiguous.py --plate thin --dpi 300

Twelve compositional sentences

Everything a picture is of thrown away, leaving only where the energy sits — then clustered. Finding

Twelve rows, each starting with a small grey mass diagram and continuing with unrelated photographs and renders that share the same arrangement. — **17,499 references** reduced to 8×8 mass diagrams · k=12, clusters of 1,050 to 2,626 · **AMI 0.035** against 24 subjects, **0.006** against 11 mediums

Arrangement is near-orthogonal to content. A low AMI is the good outcome here and the opposite of what it means in clusters.py: at 0.035 against the vision model's subjects and 0.006 against its mediums, the compositional axis is carrying information the subject axis does not.

Working notes 254 words

A photograph of a face and a render of a turbine land in the same group if they are arranged the same way.

The pool

pool(), aesthetic filter on. Composition is a claim about how the studio's material is arranged, and a screenshot of a text post has a strong compositional signature — horizontal bands, high emptiness — that would take one of only twelve rows for a non-aesthetic reason. 19,209 references become 17,499.

Collected against made, and it weakened

The 999 published works are distributed across the twelve sentences differently from the 17,499 collected references: chi-square 19.7 against a null median of 9.6, permutation p = 0.037, largest enrichment x1.35. On the unfiltered population this test read chi-square 27.5, p = 0.004, x1.42. The effect is roughly half as strong on the pool and now sits just inside 0.05. It is a lean, not a house style.

And the enriched sentence is described by measures that carry nothing

The over-used sentence (81 of 999 published against 1,050 collected, x1.35) is characterised as mass_x +1.5 · sym_lr −1.0. mass_x scores 0.002 on the validity table and is flagged do-not-quote; sym_lr scores 0.032. The clustering is on the full 64-cell diagram, so the cluster is real — but the words describing it are not load-bearing, and this caption will not pretend otherwise.

Nothing in the input knows about anything

Not subject, colour, client, year or board. If the rows work, twelve unrelated things look like variations of one arrangement. --k varies the count and the AMI prints each time; if the headline moved with k, the finding would be about the parameter.

Method and caveats 107 words

What would make this wrong. Gradient energy is a poor stand-in for attention: a quiet subject against a busy ground has its measured mass in the ground, so some rows are rows about where the texture is. Eight by eight is a coarse diagram — it cannot tell one object from two touching ones, and anything below a fiftieth of the frame is invisible to it. And the diagram squashes where the sheet crops: d09_compose.py resizes to 128×128 without cropping, so a panorama is measured as if it were square, while the page centre-crops. A very wide reference is where the picture and the row diverge most.

python3 src/viz/d09_grammar.py --dpi 300

Four measures that measure nothing

The validity check, including the two the brief named most explicitly — rule of thirds and vanishing point — both of which fail. Failed hypothesis

Pairs of panels: a real photograph beside the grey eight-by-eight mass diagram derived from it, at a size where the two can be compared. — **16,124 labelled references** across 21 subjects · variance between subjects, of sixteen measures: best **0.105**, worst **0.002**

Four of the sixteen compositional measures carry no signal, and the two the brief leaned on hardest are two of them. thirds scores 0.004 — the rule-of-thirds measure separates nothing at all.

Working notes 171 words

mass_x scores 0.002, horizon 0.010, vp_k (the vanishing-point proxy) 0.016. All four are flagged do not quote on every run.

They stay in the file rather than being deleted, because removing a measure after watching it fail, without saying so, is how a method quietly becomes a story.

What does work, and it is predictable

empty 0.105 (product +0.57 / texture −0.68), blobs 0.098 (texture +1.00 / product −0.42), spread 0.091, border_share 0.091 (both texture-high, sculpture-low), layers 0.051. Texture running to many small blobs filling the border and sculpture to one centred mass are the directions a person would predict, which is the point of the check.

The horizon estimator's tell

It scores highest on graphic (horizon_k, 0.026), which have hard horizontal bands and no horizon in them. That is not a weak measure; it is a measure of something else wearing the name.

The guardrail is still not met

The brief asked for a check against a human-curated sample and there has been no human check. This plate exists so that one can actually be made: images beside their own diagrams at a size where the two can be compared. The validity table is the machine substitute, and it is scored against a different model's labels rather than against the one that produced the clusters.

Method and caveats 89 words

What would make this wrong. Between-subject variance is a proxy for validity and a crude one: a measure could capture something real that happens not to vary by subject, and would score zero here. It is evidence that a measure is uninformative about content, which is exactly the argument being made about thirds — but it is not proof that thirds measures nothing at all. The table is computed on the 16,124 pool references the vision model labelled, over the 21 subjects with at least 100 of them.

python3 src/viz/d09_grammar.py --verify --dpi 300

Eye level is one per cent

The archive laid out by where the picture puts you — elevation across, proximity down — instead of by what is in it. Finding

A dense grid of images; across a row the viewpoint rises from under the subject to above it, down a column the viewer moves from pressed against a surface to looking across a distance. — **17,499 references scored**, 13,031 placeable · 384 cells, no gaps · surrounded **31.0%** · confronted **12.9%** · **eye level 169, 1.0%**

FIELD's archive puts its viewer either inside a field or in front of one object, and hardly ever where a person actually stands. Surrounded 5,428 (31.0%) and confronted 2,260 (12.9%) against eye level: 169 references, 1.0%.

Working notes 258 words

Whatever this studio collects, it is not photographs of the world at head height.

The pool

pool(), aesthetic filter on. Where a picture puts you is the plainest possible taste question. It changes one headline honestly: indeterminate falls from 28.3% to 25.5% once the screenshots and pure typography are out, which is the filter removing exactly the material that had no viewpoint because it was not a view. 19,209 references become 17,499.

The check that matters more than the finding

Camera position is scored by CLIP. Subject is labelled by a different model that never saw these prompts. Eight agreements at x4.2 or better: landscape → distant overview x15.8, architecture → outside x11.2, interior → inside x9.6, landscape → above x6.6, interior → threshold x6.2, person → eye level x6.0, texture → extreme close-up x5.9, architecture → below x5.8. Inferred: agreements that strong between two models sharing no training and no vocabulary are not a coincidence.

Four categories are under one per cent — including the headline one

Eye level (169), threshold (149), distant overview (109) and looking through (104) all fall under 1%, and the script names them on every run as categories the model can barely tell apart from their neighbours. That includes eye level, which is the number this sheet is built on. The finding is that the category is nearly empty; the exact figure is not a strong measurement.

There is no key

None is needed. The top of the page is close and the bottom is far, and you can see that it is. Each row is filled by taking a proximity band and spacing it evenly along the elevation ranking, so every cell is a real image.

Method and caveats 120 words

What would make this wrong. Zero-shot proportions are not calibrated: 31.0% surrounded is a share of this prompt book, not of some absolute truth, and changing the fifteen sentences moves every share. What the sheet supports is the ordering and the relative reading between slices; what it cannot support is “a third of the archive is immersive” as a bare fact. Elevation and proximity are contrasts, not measurements — each is a softmax over two or three of the fifteen categories, so an image with no elevation cue at all still gets a position near the middle, indistinguishable from one genuinely at eye level. And the grid equalises what the archive does not; the counts are printed for that reason.

python3 src/viz/d10_viewer.py --dpi 300

Enclosure and aperture

The same references and the same instrument, turned onto a second pair of axes. Visual study

A grid of images running from open exposed space to fully enclosed interiors across, and from wide framing to a narrow aperture down. — **the same 17,499 references**, 13,031 placeable · enclosure across, aperture down · 384 cells

Method and caveats 52 words

Elevation and proximity are the pair with the strongest external validation, so they carry the finding. Enclosure and aperture are printed because the axes are cheap to swap and the archive reads differently under them — and because a grid that only ever works on one pair of axes is a grid that was fitted to its answer.

What would make this wrong. Nothing here is validated against the second model the way elevation and proximity are, so this page is a study and not evidence. Both axes are softmax contrasts over a handful of the fifteen prompted categories, which means an image with no enclosure cue still lands somewhere.

python3 src/viz/d10_viewer.py --axes enclosure,aperture --dpi 300

The quarter with no viewpoint — and the version of it that failed

The part of the archive that is a surface rather than a place, plus the obvious way of detecting it, which does not work. Failed hypothesis

A dense page of flat images: gradients, graphic marks, abstract fields, diagrams — pictures that are surfaces rather than views into a space. — **4,468 references** with no locatable viewpoint · **25.5%** of the pool · drawn on their own page because they are not on the grid

The confidence-margin implementation of “indeterminate” does not work, and it fails in the wrong direction. Tested on 3,154 flat graphics against 3,547 photographs of places: the median top-two margin is 0.0070 for the flat graphics and 0.0057 for the photographs — a working measure would be the other way round.

Working notes 130 words

At every threshold it flags more photographs than graphics: 16% against 20% at margin < 0.002, 31% against 38% at 0.004, 44% against 52% at 0.006, 66% against 71% at 0.010.

The explicit category does work. Given indeterminate as one of the fifteen sentences the model can choose, it picks it for 59% of the flat graphics and 14% of the photographs. Asking the model outright beats inferring from its uncertainty, and the failed version is printed here rather than dropped, because the margin trick is what anyone would try first.

The boards that lean into it

Against the archive's own average: FIELD-OS Branding +46% indeterminate (137 references), Type +43% (149), Graphic +36% (411), 360 IMI +36% (111). The categories that lean in are the ones that should — graphic x2.6, data-visualisation x2.1, abstract x1.5 against a different model's subject labels.

Arguably the more interesting page

A quarter of the archive is a surface rather than a place, and it is invisible on the main grid by construction. It gets its own sheet rather than a footnote.

The pool moved this number

On the unfiltered population indeterminate was 28.3%; on the pool it is 25.5%. Some of what had no viewpoint had no viewpoint because it was a screenshot. What is left is material that is genuinely flat by intent.

Method and caveats 98 words

What would make this wrong. The control uses the vision model's own labels to define “flat graphic” and “photograph of a place”, so it tests agreement between two model outputs rather than against ground truth — a systematic error shared by both would pass it. It is also computed over all 23,036 embedded ids rather than over the pool, deliberately: it is a calibration of the scorer, not a fact about the evidence. And indeterminate being a prompted category means the model is being asked to name its own confusion, which is a different act from being confused.

python3 src/viz/d10_viewer.py --plate indeterminate --dpi 300
# the margin control quoted above was recomputed from the cached scores in
# data/d10-camera.json, which is what src/d10_camera.py prints on a fresh run
# (--force). It is not re-derived by the command that draws this sheet.

The smallest honest edition

How far fourteen years can be compressed before it stops being the archive — six editions, and what each one loses. Finding

Six horizontal bands, each split by a gutter: on the left a small set of representative images, on the right a set of screenshots and text captures the edition fails to cover. — **17,499 references** · six editions from 10 to 5,000 · greedy set cover at cosine 0.80 · 64 k-means territories, smallest 4, largest 575

Ten images cover a third of this archive. 35.0% of 17,499 references sit within cosine 0.80 of one of ten pictures; 250 cover 63.2%; 5,000 — less than a third of the archive — cover 99.3%.

Working notes 236 words

The centre is enormously compressible, which is an unflattering and useful thing to know about fourteen years of looking.

The pool

pool(), aesthetic filter on. An edition is a selection to show, so it has to be drawn from the material saved as reference. It also matters for the number itself: a large block of near-identical link cards is trivially covered by one representative, and counting them would overstate how compressible the studio's looking is. 19,209 references become 17,499.

The two numbers move against each other

Territories with no representative at all, of 64: 54 at an edition of 10, 45 at 25, 20 at 100, 6 at 250, 1 at 1,000. So the 100-image edition reaches 55.0% of the archive while holding nothing from 20 of its 64 regions. A small edition reproduces the centre and erases the edges.

What the right-hand panel is

A random draw from the 240 references that edition covers worst. Read down and the left side fills with the archive's actual material while the right side stays what it is. Even with the aesthetic filter applied, what an edition never reaches is not the archive's rarest beautiful thing.

Representatives, not a canon

Every method here optimises for covering the space and none has any opinion about quality. A picture chosen at scale 10 is chosen because nothing else looks like it. cover is nested, so the first cell of every band is the same picture — that is the ladder showing, not a printing fault.

Method and caveats 95 words

What would make this wrong. “Represented” means “something within 0.8 cosine of it was kept”, and CLIP cosine is not perceptual adequacy; two pictures at 0.85 can be a photograph and a render of unrelated things. The territories are k-means regions imposed on a space that is neither round nor evenly dense — and on the filtered pool the smallest of the 64 holds just 4 references, so “lost territories” is a coarser measure here than the count suggests. Nesting is a property of cover and fps only; the kmedoid ladder is six unrelated selections.

python3 src/viz/d06_edition.py --dpi 300
# the full method comparison behind the caption:
# python3 src/viz/d06_edition.py --sweep

Farthest-point sampling loses to random

The textbook method for choosing a diverse subset, beaten by picking at random at every scale — and the usual explanation for why turns out to be wrong. Failed hypothesis

Six bands whose left-hand selections are visibly odd and marginal images, with dense ordinary material on the right that they fail to represent. — **farthest-point sampling**, six editions · share of 17,499 references within cosine 0.80 of something selected: **1.1% at 10, 1.7% at 100, 8.1% at 1,000**

Random beats it at every scale, by a factor of seven at the small end and still by 5.8 points at 5,000.

Working notes 220 words

Within cosine 0.80, fps against random: 1.1% / 7.5% at 10 · 1.2% / 18.3% at 25 · 1.7% / 35.3% at 100 · 2.6% / 43.2% at 250 · 8.1% / 60.8% at 1,000 · 75.0% / 80.1% at 5,000. Greedy set cover, at the same scales, runs 35.0% to 99.3%.

A correction to this repo's own explanation

The earlier account of this failure was that fps “walks straight out to the archive's junk — screenshots of text posts, blank scans, broken thumbnails”. That explanation is no longer sufficient. This run is on the pool with those excluded, and fps still loses to random at every scale, by almost exactly the same margins. The failure is the method's, not the archive's.

What is actually happening

Each pick is whatever is currently worst covered, so the set grows outward from the edges inward and spends the entire edition on outliers. Inferred: what fps finds when the screenshots are gone is simply the next most isolated material — which is still, by construction, the material that represents least.

Over-representation

The swell figure — the largest ratio between a region's share of the edition and its share of the archive — runs x33.3 at 100 and x12.8 at 1,000 for fps, against x2.8 and x1.9 for cover. Read it from 100 upward; below that it is dominated by granularity rather than by bias.

Why print a losing method

It is worth printing precisely because it looks wrong. This is what “maximise diversity” means when it is applied literally to a real archive, and it is the standard answer to the question this direction asks.

Method and caveats 85 words

What would make this wrong. Coverage at cosine 0.80 is one figure of merit and it is the one cover optimises directly, so the comparison is not neutral between the methods — a set-cover method winning a set-cover metric is close to a tautology. The result that survives that objection is fps against random, which optimises nothing. Note also that fps closes most of the gap at 5,000 (75.0% against 80.1%), so this is a failure at small edition sizes, not at every size.

python3 src/viz/d06_edition.py --method fps --dpi 300

Same archive, different answers

Every headline number in this project re-run under other models, other seeds, other k, other projections and five surgeries on the dataset — and a plain list of what did not survive. Finding

Eighteen columns. Each has a probe image at the top, four neighbours below it from one model, and four more below those from a second model. On the left the two blocks are the same pictures; towards the right they are completely different pictures. — **17,499 references** in the pool · four representations, five seeds, six values of k, four surgeries · 18 probes ordered by how much the two CLIP models agree about them

The cluster structure survives everything. AMI against FIELD's own boards holds at 0.160–0.162 across five seeds (between-seed ARI 0.76), 0.170 on a CLIP trained on entirely different pictures, and 0.143–0.176 under four surgeries — the pool filters removed, Pinterest removed, the largest board removed, board-balanced sampling.

Working notes 237 words

It is the strongest finding in the project and it is now stronger.

The loneliness result survives everything. Two definitions of “alone”, zero overlap in their top 48, at k = 4, 8 and 16, on both models. Six measurements, six zeros. That disagreement is a property of the archive, not of a parameter.

Board coherence's numbers are a fact about one model. 0.713–0.843 on OpenAI CLIP against 0.532–0.714 on LAION — the whole range shifted down by about 0.18. The ranking is safe (cross-model rho +0.94, split-half rho +0.88 ± 0.02), so “Culture is the loosest board in the archive” holds. “Board coherence runs 0.71 to 0.84” must never be quoted without the model beside it.

Median cluster purity is not a number at all. 0.10 to 0.11 across seeds; 0.06 to 0.55 across the inclusion rules. Cap every board at 60 references and it is 0.06; remove Pinterest and it is 0.55. An order of magnitude, from choices nobody would think to write down. Purity should be reported with its inclusion rule or not reported.

No 2-D map of this space can support the claim a reader makes of it. A PCA plane keeps 2.3% of a reference's ten nearest neighbours. t-SNE keeps 29.0% — far better, and still seven of ten neighbours on the page are not the neighbours — and two t-SNE seeds of the same data agree with each other on only 55.4% of what they each draw. A 2-D atlas can support “there are regions and here is roughly where they sit”. It cannot support “these two things are near each other”.

“Similar” is model-specific, and not marginally. Two CLIPs of identical architecture differing only in training corpus return the same top-10 neighbourhood 33.5% of the time. The control that makes that readable: the same model asked over half the archive returns 100% the same neighbours, so the 33.5% is the model and not the noise. Substrate's own text vectors agree with the image vectors 2.7% of the time and eight non-learned measurements 1.7% — above a chance rate of about 0.05%, and useless.

The pool

pool(), aesthetic filter on, through src/d07_reps.usable_refs. This one is forced rather than chosen: the findings being audited — clusters.py, coherence.py, loners.py, flex.py, atlas.py — are all computed on the pool. An audit run on a different population is not an audit, it is a sixth answer. 19,209 references become 17,499; 16,499 of them carry all four representations.

The point is not to discredit the analysis

It is to say which questions this material can answer and which it only appears to answer. A finding that survives four representations is worth more afterwards than it was before; one that does not needs saying out loud, and four here do not.

Why it is a sheet and not a chart

One column per probe, ordered by cross-model agreement. Top cell is the probe, then its four nearest neighbours in this repo's model, then its four nearest in the other one. On the left the two blocks are the same pictures. Somewhere across the page they stop being the same pictures. That is the finding rendered rather than plotted: the archive is real, its neighbourhoods are constructed.

The k-means check, on every run

This repo's own numpy k-means is verified against sklearn's on the same data: inertia 6,098 to 6,098. That check exists because a convergence bug here once returned unfitted k-means++ seeding and produced a false null in direction 09.

Method and caveats 131 words

What would make this wrong. Two CLIPs are not two models — same architecture, same objective, same rough era — so agreement between them is evidence against a corpus artefact and nothing stronger. The text and measured representations are included precisely because they are not CLIP, and they disagree with it far more. A moved number is not a wrong number: board coherence on another model is a different measurement, not a correction of the first, and the claim under test is only ever “would the sentence in the docstring still be written”. The surgeries are not independent — removing Pinterest removes most of the archive's colour, dates and boards at once. And t-SNE runs on a 4,000-reference subsample, because the question (do two seeds agree) is answerable at that size.

python3 src/viz/d07_stability.py --dpi 300

The scale ruler

Where the studio looks, in metres. A third of what it saves has no size at all. Finding

Ten horizontal bands of small images stacked down a tall page, the vertical gaps between them uneven, with two thin bands isolated near the top and bottom by long stretches of empty paper. — ten bands drawn, 220 cells · 17,359 in the evidence pool · vertical gaps are real log-metre distance

Scale-indeterminate is the largest band in the archive: 5,782 references, 33.3%. The vision pass returns a scale band and the cue it used, and where the cue is none the record is forced to indeterminate at parse time rather than believed.

Working notes 175 words

So a third of what this studio collects is material an eye cannot put a size to.

Of the 11,584 that do carry a cue, the mass sits in four adjacent bands — object 4,608 (26.5%), room 2,714 (15.6%), body 1,855 (10.7%), landscape 944 (5.4%). That is 58.3% of the pool inside roughly 0.1 m to 10 km. Human and near-human scale.

The two ends the direction is named after are close to empty. Molecular 5, cellular 48, material-surface 60, planetary 62, cosmic 30 — 205 references, 1.2% of the pool across both extremes combined. Molecular does not clear the twenty-image floor and is not drawn at all.

Why it is a real ruler

The gap between two bands is proportional to the log-metre distance between them, so the crowding near a metre is the archive's own crowding and the long empty runs are decades nobody looks at. Within a band, images run by cosine distance to that band's CLIP centroid — most typical first.

The pool decision

Default pool. This is a question about what the studio looks at, and the headline number is the indeterminate band. A link card or a screenshot of a chart has no depicted scale and lands there by construction, so counting them would be counting the answer into the question. 1,812 references dropped, and indeterminate fell from 36.0% to 33.3%.

The metres are mine

The log positions are order-of-magnitude stand-ins chosen so the ruler can be drawn. “Room” is not 5 m in general. cosmic at 1e20 m is the worst of them and it dominates the page: half the sheet is the gap beneath it, which is truthful about astronomy and arbitrary in its exact size.

Method and caveats 98 words

What would make this wrong. The cue column is the model's claim about its own reasoning, which is weaker than a record of it — a model can report human-figure on an image with no figure, and nothing here audits that. A recheck over 500 images asked twice puts scale at 90.8% and the cue at 91.6%: stable, which is not the same as right. And indeterminate is doing two jobs — an abstract gradient genuinely has no scale, a tight crop of a building has one this image cannot show, and both land in the same band.

python3 src/publish.py d11-scale-ruler
# python3 src/viz/d11_scale.py --page a2 --dpi 300

The cross-scale wormhole

The premise was that a cell, an aerial and a nebula sit visually adjacent here. The measurement says the opposite. Failed hypothesis

Forty side-by-side pairs of photographs in a four-column grid, each pair visually alike but labelled with two different scale bands. — the 40 closest nearest-neighbour pairs three or more bands apart, closest first · 80 cells

Take every reference with a real scale cue, find its nearest CLIP neighbour, and ask how often that neighbour is three or more bands away. 5.42%. Shuffle the bands and the same count is 24.84%.

A ratio of 0.22: images that look alike here are about a fifth as likely to cross scale as chance would put them.

Working notes 104 words

Visual similarity in this archive is strongly scale-preserving, which is the inverse of the direction's hypothesis, not a weak version of it.

The pairs that do cross are real and there are not many. Cellular against body at 0.949 cosine, landscape against room at 0.936, landscape against body at 0.933. One of the closest pairs was dropped as a perceptual duplicate — the same picture labelled two ways, which is a model inconsistency wearing a wormhole's clothes.

Why the twins are dropped from the drawing and kept in the rate

A duplicate pair scores 0.99 cosine and crosses three bands, so it would sit at the very top of this sheet as a labelling error rather than a finding. It stays in the 5.42% because there the shuffled baseline controls for it. Near-duplicates are caught at Hamming distance under 6 on the 64-bit pHash.

What survives

Forty pairs out of a pool of 17,359. The direction's idea works — it is just rare rather than characteristic, and a sheet of forty is the entire population of it above this similarity floor.

Method and caveats 59 words

What would make this wrong. Both bands in a pair are the model's, so a wormhole can be a mislabelling in either direction, and the shuffled baseline cannot tell those apart from real ones. The rate is also sensitive to the three-band gap: a two-band gap counts “object next to body” as a crossing, which it plainly is not.

python3 src/publish.py d11-wormholes
# python3 src/viz/d11_scale.py --wormholes --page a2l --dpi 300

What the light is doing

One row per structural role of light. More than one image in five is a subject that is itself a light source. Finding

Six horizontal rows of twenty photographs each, sorted dark to light along every row: lit objects, glowing screens, light beams, silhouettes, near-black frames, and glowing surfaces. — six roles, 120 cells · 16,360 in the evidence pool · each row sorted left to right by median brightness

The distribution is lopsided in a way worth saying out loud. 72.8% is the ordinary case — light falling on a subject, 11,902 references.

Working notes 151 words

21.6% is a subject that is itself a light source: 3,530 references of screens, neon, LED, projection and glowing form.

Everything the direction is most interested in is together under four per cent. Light-as-surface 302 (1.8%), volumetric 140 (0.9%), silhouette 123 (0.8%), darkness-with-signal 27 (0.2%) — 592 references, 3.6% of the pool.

Emitting at better than one image in five is the number that carries the story. Screens and glowing form are not a minor register here; they are the archive's second subject. Their median brightness is 0.22 against illuminated's 0.55, which is physics rather than a labelling artefact: a light source is usually photographed against a dark surround.

The pool decision

Default pool. A question about what light is doing is a question about taste, and a screenshot of a web page is lit by nothing. 1,784 references dropped from the 18,144 that carry both a perception record and a photometric measurement.

No time axis, deliberately

The brief's interesting story is a progression over fourteen years from light as treatment to light as material. Substrate has three date columns and none of them is “when FIELD found this” for more than about a tenth of the archive. A curve drawn on created_at would be a curve of when other people posted things, so it is not drawn and the progression stays open.

One role per image

A frame can be a silhouette against a volumetric beam. The model returns one label, the rows are not a partition of anything, and a picture doing two things is filed under one.

Method and caveats 74 words

What would make this wrong. darkness-with-signal has 27 references in sixteen thousand. Its row is a handful of pictures, not a register, and the counts print on the sheet so a thin row cannot pass as a thick one. The recheck is stability rather than accuracy: light_role comes back the same on 93.4% of 500 images asked twice, but light_source only 83.6%, which is why the source cut is the weakest of the three.

python3 src/publish.py d12-light-role
# python3 src/viz/d12_light.py --field role --page a2l --dpi 300

Exposure held constant

The obvious defence — that these families are not just brightness — is false. So the control was done properly instead. Failed hypothesis

Eight stacked pairs of image rows down a tall page, running dark at the top to bright at the foot; within each pair the upper row is lit objects and interiors and the lower row is screens and colour fields at the same brightness. — eight brightness bands, illuminated above and emitting below in each · 320 cells · exposure equal within a band

The guardrail asks that brightness not be treated as meaning. The obvious way to satisfy it is to show that brightness does not predict the roles.

Working notes 159 words

It does. Median brightness separates illuminated from emitting at AUC 0.775, a long way from the 0.5 of a number carrying no information.

So the claim “these families are not exposure” would have been false, and an uncontrolled version of this piece would have printed it.

What is drawn instead holds exposure fixed. Eight brightness bands; inside each, a row of illuminated above a row of emitting. Down every pair the pixels are equally bright and only the role changes — and the two halves still look nothing alike. In the bright bands the emitting half is screens, interfaces and colour fields while the illuminated half is objects and interiors. Not “brightness is irrelevant”, but “the role survives brightness being held fixed”.

Why AUC and not accuracy

These groups are 77/23, so “always say illuminated” scores 77.1% while knowing nothing at all, and a threshold adds 2.8 points on top without the number carrying much. AUC is the probability that a random member of one group outranks a random member of the other, and 0.5 is no information regardless of how lopsided the classes are.

The other three readings carry almost nothing

Entropy 0.624, median saturation 0.584, brightness spread 0.549. The one photometric number that works is the obvious one, and it works by about the margin physics predicts.

The bands are not equal populations

Emitting is front-loaded into the dark end — 1,141 against 730 in band one, 110 against 1,662 in band eight. The rows are sampled evenly through each cell of the matrix, so the sheet argues about what the two kinds of frame look like at equal exposure and never about how many of each there are.

Method and caveats 76 words

What would make this wrong. Four photometric numbers tried one at a time is not every photometric number. A trained classifier over all of them, or a feature closer to what “emitting” means — the shape of the top percentile of the histogram, say — would likely beat 0.775. That still would not make the labels meaningless; it would mean the model is naming something a photometer can partly see, which is what one would expect.

python3 src/publish.py d12-light-matched
# python3 src/viz/d12_light.py --matched --page a2 --dpi 300

The physics the studio likes to break

A catalogue of broken laws, split photographs above renders below, because uncontrolled it would be a CGI detector. Provocation

Eight labelled blocks of images, each split into two rows: objects hanging in mid-air, bodies dissolving into particles, frozen splashes, matter passing through matter. — eight conditions, 189 cells · 16,198 in the pool carrying both a condition and a medium label · photographs above, renders below in every block

6.41% of the pool shows any impossible condition at all — 1,039 references. The rest is none.

Working notes 158 words

Whatever this studio's relationship with physics is, it is not that most of its references break it.

Two conditions carry most of it: levitation 519 and no visible support 258, which are close to the same observation made twice — an object in the air, and an object that ought to fall. They should probably be read as a single count of 777, 4.8%. Third is body-dissolving-into-field at 119, fourth frozen fluid at 68. Everything else in the brief's list is in the low tens or fewer: matter-through-matter 28, liquid-solid 13, impossible-scale 13, transparent mass 12, sideways gravity 7, endless depth 2.

What the impossible images are of: abstract 22.0%, installation 16.7%, person 11.7%, object 11.3%, sculpture 10.2%.

Why this is a provocation and not a finding

The brief calls human review essential and it has not happened. What the repo can do instead is put every image on the page beside its claimed law at a size where a reader can reject it. The blocks are visibly coherent, which is itself some evidence the labels mean something, and it is not review.

Confidence is a dead control

Raising the floor changes almost nothing: 1,039 at conf ≥ 0, 1,039 at conf ≥ 2, 987 at conf ≥ 3. The model answers 3 nearly every time it answers at all. impossible_conf is 97.6% stable across a recheck, which is stability without content, so it filters nothing here.

The pool decision

Default pool. The medium control divides photographs from renders, and the non-aesthetic material — link cards, pure typography, screenshots — is labelled graphic or illustration rather than either, so leaving it in dilutes both denominators and moves every lift. 1,783 references dropped.

Method and caveats 73 words

What would make this wrong. The condition label is the model's poetic reading, exactly as the guardrail warns, and it is the most stable field in the pass at 96.6% over 500 images asked twice — which is still not review. The medium label is also the model's, and a photograph of a screen showing a render is a photograph; the lift column inherits that error in the direction that inflates the finding.

python3 src/publish.py d13-impossible
# python3 src/viz/d13_impossible.py --page a2l --dpi 300

The medium control

Renders break physics 4.1 times more often than photographs. Most of what looks like a taste for the impossible is a taste for 3D. Failed hypothesis

Eight rows of photographs only, no renders: suspended liquid, splashes stopped mid-air, objects held in the air, figures reflected into transparency. — the same catalogue restricted to what a camera made · 87 cells across 8 rows · 10,567 photographs in the pool

The pool is 65.2% photograph against 27.1% render. A render can break physics for free; a photograph mostly cannot, and when it does someone built a rig, waited for a splash, or lied in post.

Working notes 82 words

So a count of impossible conditions with no medium control is a CGI detector wearing a poetics costume.

Measured: photographs 3.43%, renders 14.14%, an overall lift of 4.1×. The warning in the brief was right, and the uncontrolled version of this piece would have ranked a taste for the impossible by how much 3D work the studio happens to have saved.

What survives the control is small and specific. Two conditions have a lift near 1, meaning the medium does not explain them — transparent mass at 1.4× and sideways gravity at 1.8×, both tiny at 12 and 7 references. The one with both a low lift and real weight is frozen fluid at 2.7× over 68 references: splashes, suspended liquid, a pour stopped mid-air. Endless depth is the only condition that appears exclusively in photographs, and it has two.

So the direction does not support “FIELD repeatedly seeks worlds where physical rules become negotiable” as a measured claim about the archive. It supports a much narrower one about frozen fluid.

Lift, stated

How many times more often a render shows a condition than a photograph does. High lift means the condition was made. Near 1 means it was sought, and those are the only ones this direction can honestly claim.

Every count reported twice

Once inside each medium, on the same population, so the control denominator and the catalogue numerator can never drift apart. That is why the pool filter had to be applied before the split rather than after.

Method and caveats 62 words

What would make this wrong. Transparent mass at 12 references and sideways gravity at 7 are the two results that survive the control, and neither is a population you can conclude anything from. The honest reading of this sheet is that the control removed the finding and left a hypothesis worth testing on more material, not that a smaller finding was recovered.

python3 src/publish.py d13-impossible-photographs
# python3 src/viz/d13_impossible.py --photographs --page a2l --dpi 300

The wind map

Two towers where the archive runs flat and upright. They are frames, not forces. Failed hypothesis

twelve orientation bins of fifteen degrees, column height is the real share · 423 cells · 17,392 in the pool

36.3% of the pool has a dominant axis within fifteen degrees of horizontal — 3,111 in the first bin and 3,200 in the last, against a 16.7% baseline for those two bins together.

Working notes 145 words

24.8% is within fifteen degrees of vertical, against the same 16.7%.

Together: 61.1% of the archive is orthogonal against a 33.3% baseline. That is a strong, real measurement, and it is what horizons, skylines, table edges, window ranks, columns of type and the rectangle of the frame itself produce. It is almost certainly not about motion.

And the pixels cannot detect motion at all. Median coherence of frames the model calls still is 0.215; of frames it calls moving, 0.231. Directional structure does not distinguish the two. The radial fit is worse than null — 0.070 for still frames against 0.063 for the ones labelled radial, the wrong way round.

The filter did not explain the towers

The obvious suspicion is that the orthogonal peaks are screenshots and pages of type, which are the most perfectly orthogonal objects in any archive. The default pool removes 1,816 references and the towers barely move — 62% orthogonal before, 61.1% after. So the frames are the archive's, not the filter's leftovers.

What a column is made of

Its bin's own most directional images, most coherent at the bottom. Which is close to circular for the towers: the columns select for streaks and strong parallel structure, and coherence is what measures those. The thousands beneath the drawn cells are counted and not drawn.

Coherence is low everywhere

Median 0.217 of a possible 1.0, ninetieth percentile 0.545. Direction 14's premise — that this studio selects stills already carrying behavioural potential — is not what the archive says. It selects stills.

Method and caveats 69 words

What would make this wrong. A still image has an axis and never an arrow. Rising and falling are the same measurement to a structure tensor, and the sign exists only in the model's answer. Orientation from gradients is objective and measures the wrong thing — a brick wall scores as a strong horizontal — which is the whole reason both readings are on the page instead of one.

python3 src/publish.py d14-wind-map
# python3 src/viz/d14_latent_motion.py --page a2l --dpi 300

Where the pixels and the model agree

The archive is 87.8% still. On the one class that names a direction, two independent methods land on the same angle. Finding

Ten rows of images grouped by kind of implied movement: streaks and trails, spirals, radiating bursts, rising and falling forms, and a long final row of still frames. — ten families, 257 cells · 17,358 references carrying both a pixel reading and a model reading · row order is by count, still held back to the foot

The archive is still. 15,248 references come back still — 87.8%. Counting indeterminate as well, only 12.1% carries any implied motion at all, and the largest moving class is linear at 868 (5.0%).

The one genuine positive is the agreement test. Of the 867 references the model calls linear and gives a direction, 54.4% have a pixel axis within thirty degrees of it against 33.3% by chance, with a median error of 22.5° against 45°.

Working notes 144 words

That is far outside noise. When the model says an image moves left to right, the gradients agree that it runs left to right — two independent methods, one made of arithmetic and one of language, landing on the same angle.

Falling holds up at 51.2% over 172; rising does not, at 40.9% over 171 with a median error of 42.3°.

So the pixels can confirm a direction the model reports and cannot detect motion the model reports. Those are different claims and the sheet keeps them apart.

The one family where coherence separates

linear at 0.313 median against the archive's 0.215. Which is close to circular: a linear-motion image is largely defined by having a streak in it, and a streak is what coherence measures. The vortex fit is the only other positive — 0.103 for the 322 rotational frames against 0.073 for still ones, weak.

Compared mod 180, on purpose

Rising and falling collapse to vertical for the comparison because a structure tensor has no sign. Scoring the model mod 360 would be scoring it against a number that does not exist.

88% still may be caution, not stillness

The prompt offered still as a correct and common answer, which is honest and also makes it the cheap answer. The motion label is 95.8% stable over 500 images asked twice; motion_conf only 76.8%, so the confidence field filters nothing here.

Method and caveats 53 words

What would make this wrong. Neither reading is trustworthy alone, and the agreement test is the only ground they share. It can only be run on the three labels that name a direction — 1,210 references of 17,358 — so the positive result covers seven per cent of the material the sheet draws.

python3 src/publish.py d14-motion-families
# python3 src/viz/d14_latent_motion.py --families --page a2l --dpi 300

Ten oppositions, and six of them were never opposed

The shape of each field is the argument. Most of them are round blobs. Failed hypothesis

Ten small square fields of tiny images arranged across a wide page, most of them roughly round clouds, one of them a clear diagonal band. — ten oppositions, each an 18×18 field over ±3.4 sd on both poles, one image per cell · 2,656 pasted · 17,522 in the pool

Of the brief's ten oppositions, exactly one behaves like a strong opposition in this archive: soft/engineered at r = −0.409. Three are weak — ancient/futuristic −0.184, serene/violent −0.131, minimal/dense −0.080.

The other six are not oppositions at all. Precise/chaotic −0.073, physical/digital −0.002, intimate/monumental +0.009, organic/synthetic +0.046, familiar/alien +0.091 — and controlled/emergent comes out positively correlated at +0.158, meaning the two words describe the same images here.

And the corner is not over-populated. Excess — the count of double-high images against what each axis's own correlation predicts — runs 0.79 to 1.24 across the ten.

Working notes 197 words

Nothing near a doubling. Only ancient/futuristic at 1.24 and familiar/alien at 1.18 sit above the noise, and neither by enough to claim.

So the archive does not demonstrably over-collect images that hold opposing qualities. What it shows instead is that most of the proposed oppositions were never oppositions — the guardrail's own question answered rather than the brief's.

The finding underneath the finding

Scored the obvious way, with no debiasing, all ten axes came out positively correlated at r between +0.55 and +0.83 — which reads as “every one is a false binary”. That was CLIP's text cone, not FIELD. The gap between those two runs is the largest single effect measured anywhere in this direction, and it was produced entirely by the method. It is why the numbers print on the sheet rather than being summarised.

Fixed limits, not quantiles

Quantile bins fill the square evenly and would hide the very thing the field is drawn to show — that a real opposition leaves its top-right corner empty. Occupancy runs 75% to 87% of 324 cells, and soft/engineered is the emptiest at 75%.

The pool decision

Default pool, and every statistic recomputed over it rather than inherited. d15_poles.py scores the whole embedding set with only the duplicate filter, because encoding CLIP text directions is the expensive half. Here the z-scores are re-standardised over the 17,522 that survive the pool and r and the corner excess computed again from those — otherwise a correlation measured on one population would sit beside cells drawn from another.

Method and caveats 81 words

What would make this wrong. CLIP's word is not the studio's word: “monumental” is trained largely on captions about buildings, so an axis can sort subject matter while appearing to sort quality. And absence of excess is not absence of contradiction — this measures contradiction as two text directions both scoring high, which is a thin proxy for an image that genuinely sustains two readings. A picture can be unmistakably both serene and violent to a person and score unremarkably here.

python3 src/publish.py d15-field-contradictions
# python3 src/viz/d15_contradictions.py --page a2l --dpi 300

Soft against engineered

The one axis that is really an axis, drawn full page. It is about making, not about mood. Finding

A tall square-gridded field of small images leaning as a diagonal band from the lower left to the upper right, with the top-right corner nearly empty paper. — one opposition, a 30×30 field over ±3.4 sd · 609 of 900 cells filled (68%) · r −0.409, corner excess 0.83

r = −0.409, the only strong opposition among the brief's ten, and the field leans as a diagonal band because scoring high on one pole really does cost you the other.

Working notes 70 words

Its corner — both poles above +1 sd — holds 93 references, which is 0.53% of the pool against the 0.64% that this axis's own correlation predicts. An excess of 0.83: emptier than chance, not fuller.

The axis is about making, not about mood. Asked to propose its own pairs rather than take the brief's, the archive returns natural/machinic (−0.49), soft/machinic (−0.46), soft/industrial (−0.42), organic/machinic (−0.39). The recurring pole is machinic, and the tension these references hold is hand-and-body against machine — not serene against violent, and not any of the temporal or scale pairs.

The corner's own subject mix backs that: product 24%, object 15%, person 9%, and 73% photograph.

Why the emptiest field is the strongest one

68% occupancy against 75–87% for the others. A real opposition cannot fill its corners, so the empty paper is the measurement rather than a gap in the data.

One image per cell throws away density

A cell holding four hundred candidates and a cell holding one look identical on the page. The counts print to stdout; the picture does not carry them.

Method and caveats 65 words

What would make this wrong. The ten pairs are the brief's guess and the search bank is mine — neither is the studio's vocabulary. Every pair in this piece is a proposal a person should accept or throw out, and the fact that the archive's own best axis is natural/machinic rather than anything the brief named is the clearest sign the vocabulary came from outside.

python3 src/publish.py d15-field-soft-engineered
# python3 src/viz/d15_contradictions.py --axis soft/engineered --page a2 --dpi 300

The paradoxical corners

The images that sit deepest in each top-right corner, one row per opposition. Visual study

Ten labelled rows of fourteen images each, every row the pictures that score highest on two supposedly opposing qualities at once. — ten corners, 140 cells · ranked on the weaker pole, not the sum · both poles above +1 sd

Working notes 79 words

What the corners actually hold

Depth is the largest guaranteed score on the weaker pole. Ranking on the sum would let a huge left score carry a mediocre right one, which is the opposite of what a contradiction card wants.

Corner populations run from 93 (soft/engineered) to 696 (controlled/emergent). The subject mixes are the honest reading of them:

intimate/monumental, 355 references — person 18%, interior 16%, installation 14%, and 90% photograph. Ancient/futuristic, 334 — architecture 22%, product 15%, and 46% render, the most render-heavy corner on the sheet. Controlled/emergent, 696 — installation 25%, abstract 15%, which is the corner of a pair that was not an opposition to begin with.

Why the subject mix prints

A corner that is 80% one subject found a category, not a tension. None of these are that concentrated — the strongest is installation at 25% — which is mildly reassuring and not a result.

Read this beside the excess column

Every one of these corners is at or below what its axis's own correlation predicts. The rows are worth looking at; they are not evidence that the studio seeks the images in them.

Method and caveats 60 words

What would make this wrong. This is a visual study and it is labelled one deliberately. The corners exist for every axis including the six that are not oppositions, so a beautiful row here is not evidence of a contradiction — for those six it is just the far end of two independent qualities, which any pair of numbers has.

python3 src/publish.py d15-contradiction-cards
# python3 src/viz/d15_contradictions.py --cards --page a2l --dpi 300

Two kinds of rectangle

The one the machine handed over, and the one somebody cut. Where the studio crops, it is extracting information off a page. Finding

Two large fields of small images side by side, each picture at its own true proportions rather than squared, the left field noticeably paler and flatter than the right. — 4,281 cut by hand (23.4%) · 14,007 came with the image (76.6%) · 420 drawn each side, 17×25 per field · no image centre-cropped anywhere on the sheet

Where this studio crops, it is extracting, not reframing. Compare the 3,996 hand-cut frames carrying a vision label against the 13,142 machine-shaped ones that do, one two-proportion z-test per label:

data-visualisation 6.88% cut against 4.17% (z +7.0) · sculpture 7.36% against 4.86% (z +6.1) · made = drawing 8.51% against 4.41% (z +10.1) · made = diagram 3.38% against 1.48% (z +7.6).

And in the other direction: person 6.18% against 7.43% (z −2.7), texture 4.03% against 5.04% (z −2.6), machine 1.20% against 1.89% (z −2.9), photograph 58.56% against 61.59% (z −3.4), render 20.80% against 24.88% (z −5.3).

Measured off the pixels rather than the labels, the cut set is brighter (median 0.5412 against 0.4804) and less saturated (0.1094 against 0.1392) with more edges (0.1884 against 0.1752) — the signature of a white page with a diagram on it.

So cropping here is an act of extraction: pulling a chart, a drawing or an object off the page it was printed on.

Working notes 186 words

It is not the act the brief imagined, of reframing a photograph to remove people or context. People go slightly down in the cut set. The studio takes photographs the shape it finds them.

The guardrail is the measurement

The brief says do not assign intention when the platform created the crop — and on this material the platform created most of the crops, so it cannot be honoured by a footnote. frame_origin() splits every reference by where its proportions came from before anything is drawn.

The split validates against itself

By source, cut rates run instagram 2.4% (the platform enforces the shape at upload), manual 14.4%, creativeapplications 19.1% (a WordPress theme crops every header to 3.545:1), pinterest 28.2%, basecamp 34.5% (files a person dragged into a project thread). Instagram at the bottom and Basecamp at the top are the two ends the method predicts before looking.

Why the fields are drawn at equal counts

The right-hand field is three times the left in reality. Drawn at true proportion the eye reads the density difference and stops looking at the pictures, and that difference is already in the caption. Both fields draw 420, sampled evenly through each set after sorting by aspect. The sheet argues about what is inside the two kinds of frame, never about how many there are.

Method and caveats 71 words

What would make this wrong. The modal test is a threshold and 40 is a choice; --modal-min moves it and the counts move with it, which is why both scripts print the split rather than only drawing it. It cannot see a hand-crop that lands on 3:2, and it will call a rare camera format a hand-crop. It is a population measure and no single image's label should be trusted alone.

python3 src/publish.py d16-crop-fields
# python3 src/viz/d16_crop.py --page a2l --dpi 150

Saved against source

The brief asked for before-and-after crops. In 26,214 references there are eleven, and every one is inside a single platform. Failed hypothesis

Eleven pairs of nearly identical photographs stacked down a portrait page, each pair differing only in how tightly it is framed. — every verified re-crop in the archive · 11 pairs over 18 distinct references, 22 images drawn · aspect ratios 1.13 to 1.26 apart, CLIP 0.905 to 0.988

The brief asks for saved-against-source: retrieve the original, align it, show what was left outside. That comparison is not available in this archive, and the failure is specific enough to be worth recording.

Every duplicate family was searched for members whose aspect differs by more than 10%: 1,885 candidate pairs.

Working notes 91 words

Those are candidates and not crops — the perceptual hash resizes to a fixed 8×8 grid, so it is aspect-blind and matches structurally similar pictures freely. Confirming each pair against its two CLIP image embeddings, which are not aspect-blind: 21 pairs at CLIP ≥ 0.85, 11 at ≥ 0.90, 8 at ≥ 0.95.

Eleven. In 26,214 references there are eleven cases of the same image saved twice at two different framings, and not one crosses from one platform to another — all eleven are pinterest to pinterest.

That is a result rather than a failed run. The studio does not re-crop what it already has, and the archive holds no before-and-after pairs to argue from. The sheet draws all eleven at once, which is the entire population and fits on one page.

Why this belongs in public

A null with a population of eleven is more useful than the sheet the brief asked for would have been, because it says the question cannot be answered here rather than answering it on eleven cases. Ten of the eleven keep the same subject label on both sides.

The shared population

Read the 26,214 alongside the d19 sheets: d16_crop and d19_frames both call d19_formats.load(), so this is one population reported twice and a change to it moves both. See the note under the frames sheets.

Method and caveats 78 words

What would make this wrong. The candidate list comes from perceptual-hash families, so a re-crop tight enough to change the structure of the picture would never enter the 1,885 and could not be confirmed. Eleven is a floor, not a count — but the platform result holds regardless of the floor, because the mechanism that would produce cross-platform pairs is a person re-saving the same picture from two places, and none of the 1,885 candidates does that either.

python3 src/publish.py d16-recrops
# python3 src/viz/d16_crop.py --mode pairs --page a3 --dpi 120

First-hand against second-hand

Most of this archive was collected by somebody else first. The studio started out finding things and ended up collecting collections. Finding

Two fields of very small images side by side at unequal widths, the left about a third of the sheet and the right about two thirds, both running oldest at the foot. — 11,779 drawn at 4.22mm cells · first-hand 4,374 (37.1%) against taken-from-elsewhere 7,405 · fields at true proportion, date order, oldest at the foot

Pinterest's export carries parentPinId on every pin — the pin it was repinned from. Of the 11,902 pins live in Substrate and surviving the pool:

taken from somebody else's board 7,407 (62.2%) · pinned first-hand from the web 4,389 (36.9%) · re-saved from the studio's own pin 106 (0.9%).

A pin with no parent was saved from a page.

Working notes 183 words

A pin with a parent was saved from another pinner's board — already collected, already framed, already captioned by somebody else, and the studio took the whole package.

And it moves, in one direction, across fourteen years. 2012: 71.3% first-hand (on 195 pins, so read it as a direction). 2016: 45.0%. 2018: 32.3%. 2023: 21.3%. 2025: 20.6%. 2026: 19.6%.

Where the two kinds of pin point is a second, independent signal that they are different acts. Reached first-hand: vimeo 386, instagram 225, youtube 193, google 157, behance 151. Inherited from other boards: behance 283, flickr 236, instagram 144, designboom 101, vimeo 97, archdaily 97, dezeen 81. What the studio found itself is moving image and portfolios. What it inherited is architecture and design blogs.

Project boards, the numbered ones opened for a live job, are more second-hand than the taste boards: 30.1% first-hand over 5,226 against 42.2% over 6,676. Under a brief the studio takes what somebody else has already gathered.

The pool decision

Aesthetic filter off. This is a sheet about the archive's own provenance, and a screenshot has a parent pin like anything else. Filtering to reference material would answer a different question. What the pool does change here is the 106.

The method's own mistake, printed

1,089 pins in the export name another of the studio's pins as their parent. 106 are live — and all 106 were in dedupe.redundant(). Not most. All. A picture deliberately saved a second time was, to every measurement in this repo, a duplicate to be dropped, and the third category of provenance was empty in every deduplicated analysis including this one. pool.protected() puts them back: 106 dropped by the raw duplicate filter, 0 dropped by the pool.

Why the fields are unequal

Unlike the crop sheet, these two are drawn at true proportion — the first-hand field is 37.1% of the ink because it is 37.1% of the archive. Inside each field images run in date order, so the drift toward second-hand is visible as the fields change density up the page.

Method and caveats 101 words

What would make this wrong. parentPinId records the pin a repin was made from, not the human act — a pin with no parent could have been saved with the browser button from a page, which is what “first-hand” is meant to mean, or created by an upload or a paste. And createdAt is the pin's date, which for a repin is the date the studio took it and not the date the picture entered the world. This covers Pinterest only; the other live references have no equivalent field, so nothing here is a claim about the archive as a whole.

python3 src/publish.py d17-provenance
# python3 src/viz/d17_biography.py --mode provenance --page a2l --dpi 150

Link rot, drawn

Two in five pages pinned in 2012 no longer answer. The picture is here; the page it came off is not. Finding

Nineteen columns of small images across a wide page, one per year, with many pictures in the older columns faded almost into the paper and the recent columns nearly solid. — 1,758 checked references drawn in year columns · 323 dead (18.4%) ghosted back toward the paper · 921 servers that refused a script excluded

Of the references drawn here, 323 point at a page that is gone or a host that no longer answers — 18.4%. Split by the year the material was filed:

2012: 41.3% gone, over 63 references.

Working notes 249 words

2015: 39.8% over 113. 2018: 19.4% over 134. 2021: 9.2% over 130. 2024: 3.3% over 123. 2026: 0.0% over 112.

Read against the survival side of the same check — the Pinterest subset of the link sample, 1,436 links, where the destinations are a genuinely diverse web, and slightly larger per year than the drawn columns because it does not require a cached picture: 2012 54.1% still answering over 63 links, 2015 59.6% over 123, 2018 78.5% over 139, 2021 88.1% over 110, 2024 90.4% over 59, 2026 100% over 11.

This is the argument for owning your own data, measured rather than asserted. Substrate holds the picture. The page it was taken from, for two in five of the 2012 links, does not exist at that address any more. A studio that had kept only the link would have lost two fifths of its 2012 research and all of the context around it.

The sampling method, because the strata are the whole story

A stratified survival check is only as good as its strata. d17_linkcheck.py takes up to 150 links per calendar year, 2008–2026, seed 7, from live references with a public http(s) URL, minus the duplicate filter and minus private hosts. Every year with enough material contributes exactly 150, so the sample is deliberately not proportional to the archive — it is built to make thin early years comparable with thick recent ones. 2,766 links checked, HEAD then a ranged GET, four outcomes kept apart.

Why refusals are excluded, and what that costs

refused — 401, 403, 429 — is a live server declining a script, not a dead page, so it is counted separately and left out of every survival rate. It is 33.3% of the whole sample, and 92.2% of CreativeApplications' 880 links. On the full sample the early years would be a measurement of one publisher's rate limiter, which is why the survival curve is quoted on the Pinterest subset (1,436 links, 68.6% still answering).

Where the strata are too thin to read

After refusals and the requirement of a cached picture, 2009, 2010 and 2011 fall to four references each and 2008 to 45. Those columns are on the sheet at their real size and they carry no information; the curve should be read from 2012 forward. This is the cost of a publisher-dominated early archive, and it is visible rather than smoothed.

Method and caveats 87 words

What would make this wrong. A soft 404 — a dead article redirected to a live homepage — is not detected and counts as alive, so every survival figure here is an upper bound and every decay figure a floor. The check also ran once, from one network, on one day; a site down for an afternoon reads as gone. And the sample is of URLs the archive recorded, which for Instagram (391 links, 98.7% still answering) means a platform permalink rather than a page anyone published.

python3 src/publish.py d17-decay
# python3 src/viz/d17_biography.py --mode decay --page a2l --dpi 150

The archive has no weather and no hour

Seven in ten references depict no time of day at all, and the light is not the sun. Finding

One long horizontal band across the page made of thousands of tiny images in segments of different widths, each segment running dark at the foot to light at the top. — 16,017 of 16,017 drawn at 3.54mm cells, 109 rows · one column per cell width, so a segment's area is its real count · ordered by measured brightness, dark at the foot

16,017 references, asked what time of day they depict: no-cue 7,871 (49.1%) · indoor 3,504 (21.9%) · day 3,043 (19.0%) · night 921 (5.8%) · golden 397 (2.5%) · dusk 236 (1.5%) · dawn 23 (0.1%).

71.0% of what this studio has collected depicts no hour of the day. Only 4,620 references, 28.8%, carry a readable outdoor time.

Working notes 301 words

Season is thinner still: 90.8% is seasonless and the four seasons together account for 1,233 images.

Dusk beats dawn ten to one — 236 against 23 — a real asymmetry and not an artefact of anything the model was asked.

And the light is not the sun. Artificial 48.1%, sun 15.2%, none 14.9%, screen 8.2%, sky 6.8%. The archive is lit by lamps, LEDs and screens: a record of interiors, studios, galleries and rendered space.

Asked instead about the age of the thing shown: contemporary 66.1%, timeless 28.2%, c20 3.0%, future 1.7%, historic 0.8%, ancient 0.2%, prehistoric 0.2%. 1.1% depicts anything made before about 1900. The hoped-for story of an archive in conversation across radically different historical times is not what the pictures show.

The pool decision, and this was the one left for a human

This script was still calling dedupe.redundant() directly. The decision is not close: the headline is what share depicts no hour, and a link card, a screenshot and a page of typography are all no-cue by construction — pictures of nothing at any time. Counting them counts the answer into the question. Default pool, 1,716 references dropped, 17,733 to 16,017. The finding survives and softens: no-cue 52.0% to 49.1%, no-hour 72.5% to 71.0%. The clearest evidence the filter did the right thing is the lighting column, where none falls from 19.6% to 14.9% — the dropped material was disproportionately the material lit by nothing.

The validity test, printed before the sheet draws

A model asked “is this night?” will answer, so the answers are checked against something that was not the model. measured.json holds brightness_med for every one of these images, measured off the pixels months before this pass existed: night 0.172 · dusk 0.339 · indoor 0.416 · golden 0.474 · day 0.514 · dawn 0.533 · no-cue 0.557, against 0.484 for the whole set. Monotone from night through dusk and golden to day, on an independent measurement, over 921 and 236 and 397 and 3,043 images.

The form is the proportion

One band across the page, each segment one column per cell width, so a segment's area is its count and nothing is scaled anywhere. Dawn is one partial column, 23 cells in a field of 16,017, because that is what dawn is here. Inside each segment images run by measured brightness, so the validity check is not only printed but visible: the night block reads dark and the day block reads light, and nothing arranged them that way except two independent measurements agreeing.

Method and caveats 81 words

What would make this wrong. Brightness agreeing with the label does not prove the label — a dark studio shot and a night street both read dark. The check rules out the failure mode where the model answers from the prompt rather than the image; it does not certify any individual record. Dawn at 0.533 on 23 images is the one bin too small to trust, and it is drawn at that size, which is the honest way to show it.

python3 src/publish.py d18-daytime
# python3 src/viz/d18_daylight.py --mode daytime --page a2l --dpi 150

The archive changes shape completely

Read as taste it is dramatic. It is not taste. Failed hypothesis

Nineteen groups of concentric rectangle outlines, one per year, each a stack of every frame from that year, changing from wide and flat in the early groups to tall and square in the later ones. — 19 year groups, 18,215 frames of 26,214 drawn · every reference as nothing but its own outline, stacked concentrically · 73 outside the 0.25–3.6 aspect clamp left out

The shape of this archive changes completely across nineteen years:

2011 — ultrawide 74.3%, square 0.4%, tall 0.0% (245 references)
2015 — ultrawide 0.8%, square 9.6%, tall 31.9% (492)
2021 — ultrawide 0.8%, square 23.8%, tall 19.8% (1,097)

Read as a story about taste that is dramatic: the studio abandons the panorama, discovers the portrait, then the square.

Working notes 179 words

It is not a story about taste, and the source column explains all of it. See the sheet below.

The modal shape per year is the plainest version of it: ultrawide in 2009, 2011 and 2012; wide in 2010, 2013, 2014, 2016, 2017, 2019–2022 and 2026; tall in 2015, 2018 and 2023–2025. Those are not aesthetic eras. They are which pipe the intake ran through.

Constant area, not constant height

Every frame is drawn at the same area, so an ultrawide and a portrait carry the same ink and shape is the only difference the eye can see. Scaling to constant height would make wide images look important and tall ones look small, which is a claim about attention this data cannot support.

Why the stack, not a chart

Outlines pile up at a year's modal shape and fade at its spread, at an opacity that falls with the count so a thin year and a thick year stay comparably legible. It is the same construction as an exposure stack, and it is the only one of the three forms that makes nineteen years comparable at a glance.

The year is the material's, not the studio's

references_.created_at is the pin's own date on Pinterest and the article's date on CreativeApplications. For a blog archive imported in bulk that means the early years are the blog's publishing history rather than anyone's browsing history — which is exactly why the ultrawide era looks so clean.

Method and caveats 62 words

What would make this wrong. Resolution is deliberately absent: the cached files are Pinterest derivatives — 9,333 of them exactly 282px wide — so a resolution axis here would describe this repo's downloader. Aspect survives the derivative and was checked rather than assumed, against Pinterest's native dimensions on 14,516 pins, matching to a median relative error of 0.00043 with 100% inside 2%.

python3 src/publish.py d19-windows-year
# python3 src/viz/d19_frames.py --by year --page a2l --dpi 150

The same frames, grouped by where they came from

Five sources, five shapes. This is a history of the windows the studio looked through. Finding

Five groups of concentric rectangle outlines: one strongly wide, one tall, one almost perfectly square, and two mixed. — 6 source groups, 25,886 frames · pinterest 11,885 · **figma 7,851** · instagram 2,636 · creativeapplications 2,249 · basecamp 1,174 · manual 177

The finding

creativeapplications — wide 56.9% + ultrawide 32.1%, so 89.0% landscape or wider. Its WordPress theme crops every post header to 3.545:1, and 648 references sit on that exact ratio to the pixel.

pinterest — tall 27.7% + portrait 18.9%, 46.6% upright. A portrait feed.

instagram — square 49.8%. Square by decree until 2015 and square by habit after.

basecamp is wide 38.8% and manual 63.2%, which is what files dragged into a project thread look like.

So the years the archive is ultrawide are the years its intake ran through a blog, and the years it is square are the years its intake ran through a phone. The guardrail — avoid claiming a platform from dimensions alone — is answered by never having to: every reference carries a recorded source, so this sheet shows the attribution directly rather than inferring it from the shape.

Working notes 196 words

The pool decision, for all four frame sheets and the crop sheet

Aesthetic filter off. This is format archaeology and a screenshot has a shape like anything else. Filtering down to reference material would remove exactly the rows that explain the shape — the CreativeApplications header crop, the Pinterest derivative widths — and turn a finding about publishing formats into one about taste. What the pool does change is that the studio's 106 deliberate re-saves are no longer discarded as duplicates: 26,214 references before, 26,214 after.

One population, reported once

d19_frames.py and d16_crop.py both call d19_formats.load(), deliberately, because both turn on the same question — where did this rectangle's proportions come from — and neither should be allowed to answer it differently. The 26,214 and the 23.4% hand-cut figure are the same numbers on both sheets, and a change to d19_formats moves the crop sheet too. They are coupled by design; the coupling is stated here rather than discovered later.

Named ratios in the population

1:1 2,798 · 16:9 1,888 · 2:3 1,527 · 3:2 1,505 · 4:5 977 · 4:3 933 · 3:4 833 · 9:16 787. Nobody arrives at 1.7778 by eye.

Method and caveats 57 words

What would make this wrong. Five groups cover 17,965 of 26,214 — the small sources are not drawn, and github, linkedin and upload are tens of references each. Source is a recorded field rather than an inference, which is this sheet's strength, and it records where the reference entered Substrate, not where the picture was made.

python3 src/publish.py d19-windows-source
# python3 src/viz/d19_frames.py --by source --page a3l --dpi 150

The wall

Every reference in the archive, once, as nothing but its frame, in date order. Visual study

A dense field of twenty-six thousand empty rectangle outlines in a regular grid, wide and flat across the top rows and progressively taller and squarer down the page. — 26,214 frames, 163×113, date order reading left to right and top to bottom · all at the same area, hairline outlines only

Working notes 61 words

What it shows

The brief's literal request, and the texture change is visible with no grouping at all: the wide band across the top is 2008–2013, and the field turns upright and then square as it descends.

It is the same population as the two window sheets — 26,214 references, aesthetic filter off — drawn once each rather than stacked, so nothing is hidden behind anything.

Why no images

An outline at constant area carries exactly one variable. Put the photographs back in and the eye reads subject matter, which is a different sheet and is the one below.

Date order, with the same caveat

The date is the material's, not the finding's. The clean band at the top is a blog's publishing history arriving in bulk.

Method and caveats 35 words

What would make this wrong. This is a visual study and makes no claim on its own. Every measured statement about it lives on the two window sheets, which is where the source attribution is.

python3 src/publish.py d19-wall
# python3 src/viz/d19_frames.py --mode wall --page a2l --dpi 150

The wall, refilled

The same twenty-six thousand frames with the photographs back inside them, at true aspect. Visual study

The same dense grid of twenty-six thousand cells, now each one filled with its actual photograph at its own proportions rather than cropped to a square. — 26,214 of 26,234 frames filled, 195×135 · every picture at its true aspect, nothing centre-cropped · rendered at 200dpi

Working notes 65 words

What it shows

The only sheet in this repo that does not square-crop, and it is that way on purpose: squaring the cells would destroy the one variable the sheet is about.

All 26,214 cells filled — no cache misses, no skipped frames. Read beside the empty wall above, it is the same argument with the subject matter switched back on: the shape change is still legible underneath the pictures, which is the point.

Drawn in pixels, not vectors

Unlike the other three, this mode composites directly into a raster at 200dpi rather than emitting an SVG. Eighteen thousand embedded images in one vector file is not a printable object.

The gaps are the shapes

Where a cell's picture does not fill its allotted square, the paper shows. That negative space is the aspect distribution, drawn as an absence.

Method and caveats 47 words

What would make this wrong. Visual study, no claim. At this scale an individual cell is about eight pixels on the published JPEG and the sheet is honest only as a texture — the print file is 4,677px wide and is the one worth looking at closely.

python3 src/publish.py d19-wall-filled
# python3 src/viz/d19_frames.py --mode filled --page a2l --dpi 200

Two senses of similar, and they barely touch

Five references in six have a visual neighbourhood and a linguistic one with nothing in common. Finding

Eighteen rows, each starting with one photograph, then six pictures that look like it, then a gutter, then six pictures described like it — and the two halves of every row are visibly different kinds of thing. — 18 rows · anchor, then its six nearest by CLIP image vector, a gutter, then its six nearest by voyage-3 text vector · 234 images · rows chosen from those sharing nothing, ordered most-disagreement first (centroids 0.485 to 0.624 apart)

The archive carries two independent descriptions of every reference, built for different purposes by different models. CLIP image vectors — 512-dimension, computed in this repo from the cached picture; “similar” means looks like.

Working notes 251 words

voyage-3 text vectors — 1024-dimension, built inside Substrate from each reference's title and scraped description; “similar” means is described like.

Over the 16,569 references carrying both, comparing the top eight neighbours under each metric:

mean overlap@8: 0.253 of 8 · share nothing at all: 13,724 (82.8%) · share exactly one: 1,989 (12.0%) · share four or more: 116 (0.7%)

Five references in six have a visual neighbourhood and a linguistic neighbourhood with no member in common. Substrate's own search runs on the text vectors. This sheet is what that search is not showing you.

Both halves of every row are real pictures, which is the only way to make the point: the left half looks like the anchor, the right half is about the same words, and they are different sets of things.

How the two senses were held apart

They were never mixed. Two matrices, two metrics, computed independently and never averaged into a single “similarity”: a 16,569 × 16,569 cosine pass under each, top-8 taken from each, and the only quantity compared is set overlap — how many of one list appear in the other. That is a measure with no free parameters and no shared axis, so neither space can flatter or contaminate the other. The guardrail is that neither side is the truth: this is a comparison between two models, not between a model and a person.

The pool decision, and what it proves

Default pool. The temptation was to keep everything and call this archive structure — but a link card is the one object where text-image disagreement is true by construction and means nothing, since its picture is a publisher's thumbnail and its text is a headline. Filtered, the finding is about reference material. And it barely moves: overlap@8 is 0.253 either way, and share-nothing goes 83.0% to 82.8% over 16,569 references instead of 18,272. So the disjointness is a property of the two embedding spaces, not an artefact of link cards.

Whose language it is

Direction 17 measured that 62.2% of the Pinterest archive was taken from someone else's board, and the description usually comes with the pin. So a good deal of the “linguistic neighbourhood” is other people's language about this studio's pictures. That does not weaken the disjointness; it changes what it is a finding about.

Method and caveats 71 words

What would make this wrong. Overlap@8 is sensitive to k, though not enough to change the story — the disjointness holds because the two spaces are built from different inputs, not because eight is a small number. A reference whose description is empty still gets a text vector from its title alone, and about 2,265 references have no description at all, so those are noisier on the linguistic side by construction.

python3 src/publish.py d20-neighbourhoods
# python3 src/viz/d20_naming.py --mode neighbours --page a2l --dpi 200

Some words are pictures

Light is a thing you can photograph. System is a thing you can only claim. Finding

Eighteen rows of images, one word each, the upper rows visibly consistent runs of one kind of picture and the lower rows looking like unsorted noise. — 18 rows, 612 images, 16.53mm cells · one word per row, its own references sampled evenly from most typical to least · ordered most visually coherent at the top · no score drawn anywhere

For every word appearing in the title or description of fifty or more references, the mean cosine of that set to its own CLIP centroid — how much the images sharing a word look alike — against a size-matched random baseline.

Working notes 195 words

834 words qualify.

Most visually coherent: installation 836 refs z +15.8 · light 653 z +13.6 · generativeart 119 z +11.2 · abstract 315 z +10.4 · touchdesigner 115 z +9.5 · glass 181 z +9.2 · sculpture 319 z +8.1.

At or below baseline, and these are the studio's own vocabulary: system 228 z +0.3 · technology 282 z +0.2 · human 211 z +0.7 · experience 299 z +0.8 · spatial 61 z +1.0 · brand 150 z −1.9 · future 258 z −2.5.

It is not abstraction that separates them. abstract is among the most visually coherent words in the archive at z +10.4. The dividing line is whether the word names something an eye can check. Light is a thing you can photograph. System is a thing you can only claim.

Of the five words the brief nominated — organic, intelligent, spatial, alive, system — intelligent and alive appear in fewer than fifty references and cannot be measured at all, spatial and system sit at baseline, and only organic (81 refs, z +3.3) shows anything.

The method checks itself at the bottom of its own list

The least coherent words are have (494, z −8.0), friends (102, −7.2), even (113, −7.1), now (380, −7.1), today (164, −7.0), world (565, −6.8) — caption filler and function words, exactly the vocabulary that should carry no visual signal, and does not. screenshot at 126 refs and z −5.6 is a word about the file rather than the picture.

What the pool changed, stated plainly

Under the old duplicate-only population, system scored z −2.7 and spatial −1.4 — measurably below a random handful of the archive. Under the pool they are z +0.3 and +1.0. The honest claim is now “at baseline” rather than “below”, which is weaker, and it is the number the sheet is drawn from. The rank order and the argument survive; the sharpest version of the sentence does not.

The ordering is the score

No number is drawn anywhere on the sheet. A coherent word reads as a run of one kind of picture and an incoherent word reads as noise, and the row is the evidence for its own position. Rows run most coherent at the top to least at the foot.

Method and caveats 101 words

What would make this wrong. Word sets are matched on raw substring tokens with a short stoplist, so light the noun and light the adjective are the same set, and homonyms are never separated. Coherence rewards a set that is tight anywhere in the space — it is not a quality score and must never be read as one. A low score is not a criticism of a word: system failing to predict a picture is exactly what you would expect of a word naming an intention rather than an appearance, and that difference is the argument, not anyone being wrong.

python3 src/publish.py d20-words
# python3 src/viz/d20_naming.py --mode words --page a2l --dpi 150

The motion work, laid flat

418 published videos across 107 of 176 works, and every sheet in this repo used to skip them. This is a film as a row you read left to right. Visual study

Twenty-six horizontal rows, each a film sampled into twenty-four frames butted together, some rows narrower than others because of their format. — 26 rows · 24 frames each · an even walk down 414 measured pieces · sorted by edge slope, −0.284 (loses structure) to +0.237 (gains it)

Working notes 232 words

Why this sheet comes first

The video work is the only material in the archive that can carry a verb. The studio's own brief argues that FIELD's signature is a set of verbs — dissolve, refract, fragment, swarm, reveal — and a verb is temporal. CLIP on a single frame cannot tell dissolve from refract: same pixels, different moment. 418 of Canopy's published media are .mp4, across 107 of 176 published works, and until these were downloaded none of it was on disk.

The two sheets after this one are reductions. A reduction is only trustworthy next to the thing it reduced, so this is the one that lets you check them: same pieces, same order, whole pictures.

The sample, plainly

418 published mp4s; 416 cached (two failed — one asset 404ing under two work records); 132.5 minutes of footage; 24 evenly spaced frames from each, 9,984 frames in all. A sheet drops anything under 1.5 seconds, which leaves 414: the two dropped are 1.0s and 1.24s. Pieces run 1.8 seconds to 453.9 seconds, median 12.5.

The ordering is the argument

In arbitrary order this is a contact sheet. Sorted by edge slope it runs from the pieces that lose edges over their length to the pieces that gain them, which is as close as measurement gets to sorting dissolve from assemble. Of the 414: 148 lose structure, 163 gain it, 103 sit flat inside ±0.005. When more pieces qualify than the page holds, the sheet walks evenly down the whole sorted order rather than taking one end of it and implying it is the range.

Not the same archive

These come from Canopy's published media, not from the reference archive, so the evidence pool in src/pool.py does not apply and never filtered them. What they are instead is curated by the client: these are the pieces FIELD chose to show and someone approved. And Cloudinary re-encoded them at w_480,q_auto:eco, so fine grain is gone and edge density reads systematically low — consistently across all 416, so comparisons hold and absolute values do not travel.

Method and caveats 145 words

What would make this wrong. The rows are not on a common time axis. A four-second loop and a seven-minute film are both stretched to the same 24 cells, so cell position compares within a row and never between two. That is deliberate — a real time axis would reduce most of the sheet to a sliver — but it means nothing here measures duration. And 24 samples is a coarse net. The longest piece in the set runs 453.9 seconds and is read once every 19 seconds; anything that happens between two samples did not happen as far as this sheet is concerned. Frames are letterboxed rather than cropped, so a square loop draws a thinner ribbon than a 16:9 film: 50 distinct aspect ratios across the 416, of which 301 are 16:9. That raggedness is real information about the formats FIELD publishes in.

python3 src/viz/filmstrip.py --sort structure --dpi 300

The motion signature

Manovich's slice, applied to FIELD's own motion work: every frame reduced to one thin column, one band per piece. Finding

Fifty-four horizontal bands stacked down the page; the upper bands are smooth colour fields, the lower bands break into hard vertical stripes. — 54 bands · centre column of every frame · an even walk down 414 pieces, sorted continuous to edited · 0 to 10 cuts · 72 to 1,200 real source frames per band, median 300

217 of the 414 pieces show no cut at all, and the two most heavily edited show ten each. The published motion work is overwhelmingly continuous — things that move, dissolve and drift inside a held frame — rather than edited.

The same fifty-four band construction, reordered so bands that thin out over their length sit at one end and bands that thicken sit at the other. — the same instrument, sorted by structure · −0.284 to +0.237 · median 450 real source frames per band

Working notes 226 words

That is a claim about FIELD's house grammar that no still-image measurement could have made, and it is legible without a key: the flat fields at the top of the sheet are the pieces that hold, and the hard vertical stripes at the bottom are edits you can count.

The caveat travels with it. A cut is only a cut if it lands between two of the 24 samples the ordering was measured from. A fast cut back and forth reads as no cut at all, so 217 is zero cuts at this sampling and never zero cuts. The number to distrust is the low one.

The silhouette is the verb

A flat field is a piece that holds. Hard vertical stripes are cuts. A smooth horizontal gradient is a dissolve or a slow drift of palette. Wavering horizontal streaks are something moving through a held frame — a camera push, a swarm, a flow. None of that needs a legend and none of it needs playback, which is the whole reason the form survives print.

The bands are dense; the sort is coarse

Worth separating, because the two halves of this sheet have different sampling. A band is built from real frames, 72 to 1,200 of them, capped at the video's own frame count so ffmpeg never duplicates its way to a smoother-looking piece. The order they are stacked in comes from the 24-frame measurement. So the picture is fine-grained and the ranking that arranges it is not, and a piece can be in slightly the wrong row while its own band is honest.

Why two sorts and not one

The sort is the piece. Ordered by cuts, the sheet separates continuous work from edited work. Ordered by structure it runs from pieces that lose edges to pieces that gain them. Same 54 bands drawn the same way both times: the only thing that changed is the question, and the eye finds the break point on its own in each.

Method and caveats 114 words

What would make this wrong. The centre column is an arbitrary sample of the frame. A piece where everything happens at the left edge reads as a flat field here and is not one — a real failure mode, and the reason this sheet is only trustworthy read against the filmstrip above it. Columns are not seconds: a band spans the whole piece whatever its length, so stripe spacing compares within a band and never between two. And a short piece has fewer frames than the band has pixels, so its columns are stretched and its stripes are correspondingly fatter; the thinnest band here rests on 72 real frames against another's 1,200.

python3 src/viz/motion_signature.py --sort cuts --dpi 300 --columns
python3 src/viz/motion_signature.py --sort structure --dpi 300 --columns

Average, sum, difference

The whole of a piece collapsed into three still images. The average of a piece that holds still is sharp; the average of a piece that moves is a smear. Visual study

Thirty triptychs in three columns; each triptych shows a smeared average, a bright accumulation and a dark panel where only the moving parts glow. — 30 triptychs, 3 across · 24 frames each · an even walk down 414 pieces, sorted by structure · difference exposure fixed at gain 3, gamma 0.6, identical for every panel

Working notes 281 words

What the third panel does that a frame cannot

Difference is the mean absolute frame-to-frame change per pixel: black is where nothing moved, bright is where everything did. It separates a locked camera with action inside it from a piece where the whole frame travels, which is exactly the distinction a single frame cannot make and the one the verb question turns on. Read down the right-hand panel of each triptych and the studio's quiet work prints dark while the busy work prints as a lit rectangle.

The sum panel is not a sum

Adding 24 frames and clipping at white was tried first. Re-measured on a fresh random 40 of the 414: the median piece clips every one of its pixels, and 38 of the 40 clip more than half. A literal sum of this corpus is a white rectangle 414 times over, and scaling it back down to fit recovers the mean image exactly, which is already panel one. So the middle panel is the brightest value each pixel reached — light accumulated the way a long exposure accumulates it. Calling it “sum” without this paragraph would be the kind of small lie that prints.

One exposure for all of them

The difference panel uses a fixed linear gain and gamma, never a per-video normalisation. Normalising each piece to its own maximum would make every panel equally bright and destroy the only thing worth comparing across the sheet. The curve is there because the corpus has a 531:1 range in frame difference — 0.0006 to 0.3188 — and any linear exposure that keeps the busiest pieces off white leaves the median almost black. It is monotonic and identical for all 414, so it compresses the distance between ranks and not their sequence. A quiet piece is meant to print dark.

Motion and cuts are the same ink

A hard cut lights the entire difference panel; so does a whip pan; so does a swarm filling the frame. The panel says the pixels changed and the filmstrip says why. Nothing on this sheet is a verb label, and the gap between a descriptor and a verb is exactly where a person has to do the interpreting.

Method and caveats 118 words

What would make this wrong. All three panels are built from 24 evenly spaced samples, not from every frame, so the average is an average of a sample and the difference is a difference between samples that can be half a second or nineteen seconds apart depending on the piece. Raising the sample count costs time linearly; it does not change what any panel means, but it does make the smear smoother and the difference brighter. And a cut in a dark scene reads quieter than a cut in a bright one, because this measures absolute pixel change and not perceived change: two identical edits, one at night and one in daylight, print at different weights.

python3 src/viz/difference.py --sort structure --dpi 300

Typography, unfiltered

The same question asked of two different populations: 875 references, or three. Both answers are correct. Finding

A grid of two hundred and thirty-six typographic references: black and white letterforms, grids, moiré type, numerals, poster layouts. — made: typography, unfiltered · 875 in the pool, 236 drawn, 18×14 · ordered most-typical first · typicality 0.250 to 0.914

A vision model labelled 1,119 images in this archive made: typography. Dropping near-duplicates leaves 875.

Working notes 206 words

Then the evidence pool's aesthetic filter runs and three survive.

That is not a bug, and it is the clearest worked example on this site of what the filter is for and what it costs. src/pool.py exists to answer what is FIELD's eye, and for that question a wall of type specimens is material the studio saved to read rather than material it saved because of how it looked — so removing it is right. For the question what typography does this archive hold, a three-image answer is useless. The filter is not a quality judgement, it is a scope judgement, and a sheet that quietly inherits the wrong scope answers a question nobody asked.

The two populations, exactly

1,119 labelled → 875 after de-duplication → 3 after the aesthetic filter. The sheet draws the 240 most typical of the 875; four of those have no cached image, so 236 are on the page. Near-duplicates are dropped on both sides of the split — --unfiltered removes the aesthetic filter and nothing else.

The three are not a summary of the 875

Scored against the same centroid as everything else, the three survivors rank 37th, 180th and 665th of 875. So the filter is not keeping the most typographic typography and shedding the edge — it is keeping three images that happen to pass a test about taste, and their position inside the category is close to arbitrary. Dominant moods across the 236: modern, structured, clean, bold. Across the three: bold, modern, dynamic, urgent.

Ordering, as everywhere else

Cells run by cosine similarity to the CLIP centroid of the facet, most typical first, so you watch the category assemble and then dissolve at the edges. The labels are the model's, not FIELD's: it never saw the channel names, and it still reproduced the studio's own fourteen-year-old taxonomy. That makes the categories real. It does not make every individual label correct.

Method and caveats 84 words

What would make this wrong. The population here is defined by a model's one-word answer to “how was this made”, and typography and graphic are not cleanly separable — a poster is both. A different labelling pass would draw a different 875. What survives that objection is the ratio: whatever the exact boundary, the aesthetic filter removes about 99% of whatever is inside it, and any sheet about the archive's type has to be run unfiltered or it is answering a different question.

python3 src/viz/facet.py --made typography --unfiltered --dpi 300
python3 src/viz/facet.py --made typography            # the three that survive the filter

The Atlas, unfiltered

The same map with the screenshot continent put back: 26,214 references instead of 21,491. Visual study

A denser grid of twenty-six thousand small images; one side of the field is visibly made of white cards, interface screens and diagrams. — **26,214 references** · 162×113 lattice · 56px cells at A0, 200dpi · the evidence pool without its aesthetic filter

Working notes 168 words

Why this one exists next to the other four

The published atlases all draw from the 21,491 references the evidence pool keeps, because the atlas is a sheet about taste and the screenshot continent that used to sit at one edge was a media-type artefact rather than a region of the eye. This is the same construction over the 26,214 the pool holds before that filter runs, and the extra 1,788 are not scattered evenly — they clump, and you can see the clump. It is the map of what the archive holds rather than the map of what the eye is, and the pair is the only honest way to show what the filter does to the shape of the thing.

What the 1,788 are

By how they were made: 861 typography, 328 diagram, 265 photograph, 134 graphic, 116 render. By subject: 489 graphic, 403 typography, 345 data-visualisation, 139 interface. Type specimens, chart studies and interface captures — things the studio saved to read, not to look at.

Not the same layout

This is a separate t-SNE run over a different set, not a zoom of the published atlas: 88 seconds over 26,214 points at perplexity 40, then a greedy snap onto a 162×113 lattice. Adding 1,788 points changes every neighbourhood a little, so regions here do not line up cell for cell with regions there.

Same two limits as the others

The layout is not stable across runs unless the seed is fixed, and distance across the sheet is not proportional to distance in the embedding — snapping to a lattice deliberately distorts scale to remove gaps. Adjacency is meaningful; measured distance is not. Near-duplicates are dropped here exactly as they are in the filtered atlas.

Method and caveats 85 words

What would make this wrong. Nothing here says the filter is a mistake — every sheet on this site that asks what FIELD's eye is should keep it, and this one would be the wrong sheet for that question. What it says is narrower: the filter removes about 10% of the archive and that 10% is a place, not a scatter. If a later pass moves the boundary, this sheet moves with it, so it is dated by its pool the way the corridor is.

python3 src/viz/atlas.py --unfiltered --max 30000 --page 1189x841 --dpi 200 --out atlas-dense

One territory on its own

The best-evidenced client in the translation field, drawn alone at the composite's own scale. Visual study

A single tall territory: a wide cloud of grey reference marks hanging below a horizon line, a chain of five image thumbnails dropping through it to a tighter cloud of published work below. — Adidas · 306 references against 32 published images · gap 0.038, null 0.892, lift +0.070, rank 1 of 40 · 556 mm per unit of cosine distance, the same scale as the composite

Working notes 176 words

What an isolate is for

On the composite sheet, one territory is a few centimetres wide and the force paths inside it are hairlines. Here the same territory has a page, and the scale does not change: 556 mm per unit of cosine distance on both sheets, because both are 420 mm tall. So every length here can be laid against the composite with a ruler, which is the one thing a zoom usually destroys.

Adidas is the strongest case in the set and the sheet is drawn solid rather than as a ghost: 0% of its 306 references sit closer to another client's published work than to its own, against 93% at the other end of the table. Its four force paths carry 8, 24, 30 and 44 references and none of them is coloured, meaning no part of this relationship is worse explained by its own research than by an unrelated studio's output.

Every length is a measurement

Vertical position is a real 512-dimension cosine distance at the stated millimetre scale. Horizontal position is local density and nothing else — the horizontal separation of two marks measures nothing at all. There is no t-SNE on this sheet and no length you can put a ruler to that is not a real number.

The two rules across it

The solid rule is the client-level gap, 0.038 — the whole-relationship claim. The dashed rule is where the same 306 references sit against an unrelated client's published work, 0.108. The distance between those two lines is the entire evidential content of the territory, and drawing them together is what stops a high pull from being read as a finding on its own.

The circles at the foot

Twenty-eight clients have published work on field.io and no reference side that joins it. They are not decoration: they are the null pool every territory is tested against, and printing them keeps the test visible next to its result rather than in a footnote.

Method and caveats 104 words

What would make this wrong. Adidas ranks first of 40 at p = 0.025, which is the permutation floor rather than a small p-value — one client cannot do better than 1/40 on this test, and twelve tests would want 0.0042 to survive Bonferroni. So this is the best-evidenced territory in a collective signal, not a licence to make a claim about Adidas specifically. The isolate also inherits everything the composite says: client-level only, because channel titles match published work on 2 of 166; and no direction, because direction needs a shared two-dimensional basis, which is the projection the sheet exists to avoid.

python3 src/viz/translation_field.py --client Adidas --dpi 200

Twelve territories, one panel each

The translation field broken apart, with each client's verdict printed over it. Half of them are ghosts. Finding

Twelve panels in a grid, each a client territory with its name and verdict above it; six are drawn solid and six are pale. — 12 territories · 3,441 references against 853 published images · 167 mm per unit of cosine distance · 6 specific, 3 marginal, 5 indistinguishable, 3 inverted

Side by side in one row on the composite sheet, twelve territories read as one gradient and the eye takes the whole thing as evidence. Given a panel each, with the verdict set over the name, the sheet says out loud that half of it is not evidenced: Adidas, OPPO, Diesel and Fairlight Cycles clear p ≤ 0.05; Ailsa Bay and Barbican clear p ≤ 0.10; Meta, Microsoft, Google and Theatre X are indistinguishable from an unrelated client; VASHI and Fairlight sit on the wrong side of their own null, matching other studios' published work more closely than their own, at lift −0.026 and −0.051.

The unmatched share tells the same story more bluntly.

Working notes 179 words

It runs from 0% at Adidas to 93% at Fairlight — the proportion of a client's references that sit closer to at least half of the other 39 clients' published work than to their own. Across all twelve it is 455 of 1,825.

Ghosts, not deletions

The six territories that do not clear their null are drawn pale rather than dropped. An unevidenced territory is a finding about the join and not an embarrassment to hide, and a sheet that silently kept only its four good cases would be the dishonest version of this one.

The scale is not the composite's

Twelve small panels get 167 mm per unit against the composite's 556 — the same geometry drawn 3.3× tighter, because a panel is a third of the page height. Lengths compare panel to panel on this sheet and must not be measured against the main sheet or the isolate. Both sheets print their own scale in the caption for exactly this reason.

What a panel still shows

Everything the composite has, at panel size: reference cloud, force paths, output cloud, the solid client gap and the dashed unrelated-client baseline. The published cloud stays visibly tighter than the reference cloud in 11 of 12, up to 2.94×. The work is a narrowing of the material, not a move away from it.

Method and caveats 121 words

What would make this wrong. Small panels flatter a weak signal: at 167 mm per unit a 0.03 lift and a 0.10 lift look more alike than they are, which is why the verdict is printed as text over every panel rather than left to the drawing. The joins are thin under the two-image threshold — six clients rest on 6 to 9 published images — and only 25 of 66 client pairs have 95% intervals that separate, so the left-to-right order of the panels is not a ranking anyone should quote. Cutting every client to six output images shifts the levels by up to 0.036 and leaves the order at Spearman +0.91: the ranking survives, the absolute values do not.

python3 src/viz/translation_field.py --small-multiples --dpi 200

The atlas, with a legend

It read as a colour field because colour is its loudest variable and carries no meaning. Position carries the meaning. Here it is, named. Finding

The image atlas with forty regions outlined and labelled by their dominant subject, purity and lift, border weight scaled by how pure each region is. — 26,214 references · 40 regions from the full 512-dimension CLIP space · 27,761 vision labels read · border weight, plate opacity and type size all scaled by purity

The atlas was a map without a legend, and that is why it read as a colour field. Colour is its loudest variable and the one carrying no meaning: across the archive, CLIP similarity against colour distance is r = −0.005, and CLIP neighbours are only 13.6% closer in colour than chance.

Working notes 172 words

Colour and meaning are near-independent here. Position is what carries the meaning. Cut the CLIP space into 40 regions and each holds a dominant subject at a median purity of 46% against a 12.6% baseline and a 3.7% uniform chance across 27 subjects — a median lift of 7×, reaching 28× for vehicles and 27× for fashion.

What licenses drawing a border at all

Two images in adjacent lattice cells share a region 58.4% of the time, against 2.7% by chance — 22×. Without that the outlines would be decoration over a projection. With it, a region is a real neighbourhood and the border is where it stops.

Six identical names, six different places

Six regions share the subject installation and would have been interchangeable. Naming each by its two most over-represented tags separates them: public art, suspended · performance, projection · art gallery, contemporary art · corridor, light art · structure, grid · digital art, exhibition.

The empty ground, answered honestly

t-SNE equalises density by construction, so a white patch on it is a property of the projection and not a gap in the archive. Instead each region carries a thinness bar — median cosine distance to the tenth nearest neighbour in the full 512 dimensions — and nine of 27 subjects hold no territory at all: diagram, illustration, infrastructure, light, material, nature, sky, water, unclear.

Method and caveats 112 words

What this sheet can and cannot support. PCA-2 retains 2% of a reference's true ten-neighbourhood, t-SNE 28%, and two t-SNE seeds agree on only 52% of what they draw. So it supports "there are regions and here is what they contain". It does not support "these two images sit near each other, therefore…". Purity is drawn three ways and printed as a number so a region at chance draws nothing; smoothing and the patch threshold are drawing operations and cannot move a figure. Captions grew from 26,296 to 27,761 during the render, so a re-run prints slightly different numbers — every figure is computed live and stamped on the sheet.

python3 src/viz/atlas_regions.py --dpi 300

The territory diagram alone

The same forty regions on bare paper, without the pictures. Instrument

The same forty labelled regions drawn on blank paper, without the underlying images. — the same 40 regions, drawn without the mosaic — for use small, or where the pictures would fight the labels

Method and caveats 25 words

Identical geometry and identical numbers to the sheet above. It exists because at a small size the mosaic wins and the labels stop being readable.

python3 src/viz/atlas_regions.py --no-images --dpi 300

The images we reach for again

1,636 references sit in more than one channel. Counted across actual jobs rather than shelves, 760 travelled from one piece of work to another. Finding

Reused references drawn as tiles sized by the number of projects they reached, above a base band of everything used once. — 30,528 live references · 1,636 in more than one channel · 760 reaching more than one job, of 16,786 that reached any · tile size is project count

Of 1,636 references in more than one channel, 876 never reach a second project. Being filed under Form and Graphic is tagging, not reuse. Counted across actual jobs, 760 references travelled from one piece of work to another — 4.5% of the 16,786 that reached any project — and the ceiling is five projects, not ten.

Method and caveats 145 words

Two ways this inflates, both corrected. The raw channel count mixes projects with taxonomy shelves: the most-reused object, a pin called KAV—K, is in 10 channels but 4 jobs and 6 shelves. And sibling channels double-count — 265 Theatre X and 265 Theatre X — Act 1 are siblings, not parent and child, so one pin read as a five-project object when the work was two. Jobs are keyed on number plus first word, because the number alone merges 238 Microsoft x SAP with 238-2 City + Guilds. The second-placed raw object, a pin called network at 9 channels, is in one project.

And the honest limit: channel membership is a proxy for deck membership. Substrate holds no record of the slide itself, so every count here means "somebody put this in that project's research", not "this survived to a slide".

python3 src/viz/reuse.py --dpi 300

The raw pyramid, with its top rows named

Every reference drawn once, stacked by how many channels hold it. This is the count before the correction, which is why it is worth looking at. Reference

A pyramid of image tiles: a single reference at ten channels on top, widening through rows of two and three, to a dense field of everything filed once. — 19,272 drawn of 30,473 in the pool, from 30,528 live · tile edge is the number of separate channels the reference was filed under

Method and caveats 64 words

The top seven rows are 24 references out of thirty thousand, and naming their channels is what shows the count to be soft. KAV—K sits in ten: four of them jobs, six of them shelves called Character, Data, Form, Graphic, Surface and X future minimalism. network, at nine, is in one job. 908 references are in exactly two channels and 18,121 in one, so the distribution is not a long tail with a fat head. It is a floor with a few things standing on it.

Why keep an uncorrected sheet on the page. The corrected version folds shelves out and siblings together, and it is the one to quote. This one is the reason to trust that correction: the six shelf names are printed beside the pin, so the inflation is visible rather than asserted. A sheet that only showed the fixed number would be asking to be believed.

python3 src/viz/reuse.py --count channels --out reuse-channels --dpi 300

Not calmer colour — calmer surface

Do we come back to a particular kind of image? Almost every apparent difference is age. One thing survives. Finding

Measured readings for reused references against the archive, with the year-stratified permutation result drawn beside each. — year-stratified, 4,000 permutations, Pinterest only · two of six readings survive

Naively the reused references are quieter, calmer and older — and almost all of that is age. A 2014 pin has had twelve years of decks to appear in, and its reuse rate is 15.4% against 0.9% for 2022 material.

Method and caveats 56 words

Holding year constant and shuffling 4,000 times, two readings survive: edge density 0.151 against 0.173, p = 0.003, and entropy 6.74 against 6.86, p = 0.023. Saturation, brightness and hue spread do not (p = 0.23, 0.56, 0.32). The studio does not return to calmer colour. It returns to calmer surface.

What would make this wrong. Both survivors held across the re-caption onto a different vision model, which moved the evidence pool underneath them — that is the strongest thing that can be said for them here. But two survivors out of six readings is a narrow result, and it is a statement about Pinterest material only.

python3 src/viz/reuse.py --figure character --out reuse-character --dpi 300

The working life of a reference

Not how many projects a reference reached, but over how many years. One was found in 2016 and used again a decade later. Finding

Each long-lived reference drawn once at every year it was used, with a grey rule running back to the date of the material itself. — 484 references used in more than one year, of 16,024 · only 115 span four years or more · the image repeats at every year it was used

484 references were pulled into work in more than one year. Only 115 span four or more. The longest life belongs to a pin of Nicolas Bernier, found in August 2016, used in Dolby that year and in 400 Kia Milan Design Week’26 a decade later.

Method and caveats 152 words

A reference used three times in one year is a mood; one reached for across a decade is a conviction.

Dating this needed three attempts and the first two were contaminated. reference_channels.first_observed_at is on all 38,020 rows and is the bulk-import date, not when a reference joined a channel — only 273 references carry more than one value. Channel creation date fails too: 46 of the project channels were created on 25 August, during the Figma ingestion, so anything touching them would read as 2026 work. So would anything filed directly into the Projects and Pitches roots, which carry the import date. The sheet uses a ladder — trustworthy channel date, then project number (year = 0.048n + 2006.9, median residual 0.20 years), then Figma median, then dropped — and prints which rung each job used. The number deliberately outranks the Figma date, because 113 Spring’s frames all carry 2026 from an old file being re-opened, which would have manufactured a ten-year life that never happened.

python3 src/viz/reuse_life.py --page a2 --dpi 300

Red threads, and the things that only look like one

A red thread is not a common tag. It is a theme leaned into again in separate jobs. Finding

Themes ranked by how many projects over-represent them, with a lower band of themes that are everywhere and nowhere more than chance. — 97 projects · over-representation counted per project against the archive rate · the lower band is the null

Across 97 projects, abstract is over-represented in 26 where chance reaches 4, installation in 25 against 5, geometric in 29 against 10. The lower band is the more useful half. texture, vehicle, organic, warm, industrial, botanical and urban are everywhere in the work and nowhere more than chance predicts.

Method and caveats 60 words

They are what the archive is made of, not what the studio keeps reaching for.

Why presence beats frequency. A tag appearing 5,000 times in one project is that project, not a thread. This counts how many separate jobs lean into a theme harder than the archive does, and draws the chance rate beside every bar — so a theme that looks common and isn't distinctive is visibly not a thread rather than quietly omitted.

python3 src/viz/threads.py --dpi 300

One person's archive, and the studio's

Substrate records what one person collected. The team's collecting went into Figma, where nothing indexed it — and most of what is in there now was never collected at all. Finding

Three stacked registers: the archive's intake by year built from its own references; beneath it Basecamp's Pinterest and Figma link counts by month on the same axis, crossing in early 2022; and at the foot two blocks of real images at a six-to-one ratio, generated against collected. — 22,069 live references by their own date · Basecamp outbound links by month · Figma images at their measured ratio · nothing file-dated is plotted on time anywhere on this sheet

Substrate holds 22,069 live references, and most came off Pinterest — which was Marcus's own habit rather than the studio's. The archive peaks at 2,605 references dated 2020 and falls to 571 by 2022, a 78% drop across exactly the window when Figma overtook Pinterest in Basecamp: February 2022 is the first month Figma led, and by August 2023 it was the default.

Working notes 132 words

The team never stopped collecting. It collected inside a tool that keeps no index, filed under whoever happened to be working — 27.9% of the images in FIELD's Figma files sit on a page named after a person. And most of what is in there now was not collected at all: for every image carrying a trace of having come from the world there are six carrying a generator's signature, and in 378 Kia Brand Space it is closer to fourteen.

Why the third register has no time axis

References recovered from Figma are dated by when their file first appeared in Basecamp, not by when the image was found. Plotted on a timeline they would manufacture a dramatic recent surge that is an artefact of the import. So the top register uses each reference's own date and nothing file-dated appears on a time axis anywhere on this sheet.

Numbers that moved

Three figures from the earlier survey are restated here from a larger sample: person-named pages 34.3% → 27.9%, generated share 18.2% → 26.4%, collected share 9.45% → 4.4%. The survey counted 9,017 distinct layer names; this counts every classified placed-image record. Bigger sample, different unit — and the ratio it implies is worse, not better.

Method and caveats 143 words

What would make this wrong. The 2026 column is drawn with a flat block where its 2,087 machine-captured references sit, so its height is honest and its content is visibly not looking — the recent rebound is a capture agent, not behaviour. The Basecamp Pinterest series is carried from the migration survey rather than re-pulled; the Figma series was re-derived row by row. The Figma corpus was still growing while this was drawn (52 → 57 projects during the run), so every ratio is computed at draw time and printed in the caption. Four Figma projects belong to clients whose published work is flagged confidential — their images are counted and never drawn. And the correlation is not proof of the cause: Google A Space for Being and On Running NYC have no references and no Figma links, because they predate Figma entirely.

python3 src/viz/two_archives.py --dpi 300

Multi-coloured in, on-brand out

IBM's references are no bluer than the rest of the archive. Its published work is half as blue again as anything else FIELD makes. Finding

Three stacked hue histograms whose bars are built from the images themselves: a wide spread of references above, a size-matched sample below it, and a narrow blue-heavy band of published work at the foot. — 379 hue-bearing IBM references · one size-matched draw of 136 · 136 published images · 910 real pictures · hue read from the `vivid` measurement on both sides

The brand colour lives in the output, not the input. IBM's references are 18.2% blue against 20.5% for the rest of the archive — very slightly less blue.

Method and caveats 75 words

Its published work is 43.2% against 29.4% for all published work. Hue dispersion falls from 0.874 to 0.418, a narrowing of +0.456. The studio does not collect IBM blue in order to make IBM work; it collects something else, and the blue arrives in the making.

What would make this wrong. Achromatic images are counted in their own column rather than dropped — IBM has 461 of them against 379 carrying a hue — because a sheet about colour that silently discards the greys is measuring its own filter. 85 published IBM videos are unmeasured. And this is one colour system throughout, the vivid reading; the repo measures colour three different ways and they must never be mixed on one sheet.

python3 src/viz/convergence.py --sheet ibm --dpi 300

Not a law, a tendency

Twelve clients, references above a shared hue axis and published work mirrored below it. The studio-wide version of the convergence claim does not survive. Failed hypothesis

Twelve paired hue histograms, one client per row, references drawn upward and published work mirrored downward from a shared axis. — 12 clients · 2,330 pictures · references above the axis, published work below

8 of 12 clients narrow — and 8 of 12 still narrow after the correction, so the count is not the story. The magnitude is. Mean narrowing falls from +0.093 to +0.040 once the reference side is subsampled to the published side's size.

Method and caveats 94 words

Adidas keeps a fifth of its apparent narrowing, OPPO a quarter. Only IBM (+0.456), Envision (+0.381) and Barbican (+0.155) are large. On (−0.314), Diesel (−0.144) and Kia (−0.108) run the other way and expand.

What would make this wrong. Nike looks like the strongest counter-example and cannot be used as one: its channel holds 36 references, 18 after the pool, and 5 carrying any chroma at all — a dispersion over five colours is not a measurement. It falls below the floor and is excluded on the same rule as everyone else. The claim this sheet supports is not "FIELD converges on brand colour" but "where a client has a strong colour identity the output converges on it while the research does not, and the effect is client-specific".

python3 src/viz/convergence.py --sheet clients --dpi 300

Why the correction was needed

The published side is always the smaller sample, and hue dispersion is biased low at small n. Half the apparent effect was that. Instrument

Four clients shown as three histograms each — full reference set, size-matched sample, published work — above a single axis carrying every client's measured and corrected span. — four clients as full / size-matched / published · every client's measured span against its corrected span on one axis

Method and caveats 61 words

Dispersion estimated from a small sample runs low whatever the underlying distribution, and the published side of every client is smaller than its reference side. Subsampling the references to the published n, 200 draws each, removes that bias. More than half the studio-wide effect disappears with it. The correction is drawn rather than described, because the uncorrected sheet is the one a reader would otherwise believe.

Why this is on the page at all. This project has published two findings that turned out to be artefacts of how they were measured — a hole in the archive that was the smoothest point on its sheet, and a null whose test could not have detected the thing it declared absent. Drawing the correction is cheaper than another retraction.

python3 src/viz/convergence.py --sheet correction --dpi 300

The looking comes first, by about ten months

Seventeen projects with dates on both halves. Every one of them was looked at before it was made. Finding

Seventeen horizontal rows of small images, each row a project, the references clustering into a dense burst well to the left of a mark showing when the work was dated. — 17 projects · 877 references · 4,224 pictures · time is the only axis, and one tile is one reference at 1.56mm

17 of 17 lags are positive, median 10 months, range 1–18. Only 2 of 877 references (0.23%) were saved after their own project's work was dated — so the archive feeds the work rather than documenting it, and that is a measured number rather than a reassurance.

Working notes 133 words

A board is a sprint: the median project puts 93.8% of its references inside a single three-month window, and in event time the mass peaks at −10 months with 95.2% inside the eighteen months before.

The null that matters

Permuting works across pairs gives p = 0.0003, but that destroys the calendar trend and two series both advancing with time will correlate whatever their relationship. The honest test is the blocked permutation, swapping only between works dated within ±24 months, which preserves the trend: p = 0.0149. That is the number to quote.

What it cannot do

Asked which of the 17 boards sits closest in time to a given work, the work's own board wins only 4 of 17 — mean rank 3.94 against a chance 9.0. Real, and far weaker than the collective claim. Several jobs run at once and their bursts overlap, so timing alone cannot tell you which board fed which brief. The sheet is drawn as a distribution, not as 17 verdicts.

Method and caveats 96 words

What would make this wrong. Canopy's date is Sanity's completedDate, hand-entered at month precision — 71 of 75 usable values carry no time, 16 calendar days hold two or more works — and it is ambiguous between finished and put on the website. So "ten months from inspiration to work" is not supported; only that the looking is over roughly ten months before the recorded date. Everything is rounded to months for that reason. And 66 of 141 works cannot be timed at all, because two bulk-import days carry 66 of them between them.

python3 src/viz/incubation.py --dpi 300

Fourteen years, both halves on one calendar

Everything the studio saved and everything it published, on a shared timeline. Visual study

A long calendar band, references above and published work below, thickening and thinning across fourteen years. — about 5,000 pictures · 1 tile = 5 references · confidential work drawn as an empty outline — position kept, picture withheld

Method and caveats 39 words

The intake collapse is visible here without being argued: the archive falls from 3,168 references dated 2020 to 688 dated 2022, across exactly the window in which the studio moved from Pinterest to Figma. February 2022 is the measured crossover, and the 2026 rebound is machinery — a capture agent and a one-off import — rather than behaviour.

Confidential work is named but never shown. Five flagged works have usable dates and appear as empty outline marks, counted separately from the four genuine cache failures so the caption does not conflate a policy with a broken image.

python3 src/viz/incubation.py --sheet calendar --out incubation-calendar --dpi 300

The claim these sheets support

The explorations

The whole thing at once

The Atlas

Why

Why gridded, not a scatter

The data

The Atlas, other ways

A prediction I got wrong

The translation field

No projection, anywhere

Pull alone is not evidence

Two territories are inverted

Reliability is drawn, not captioned

The corridor

Why this one is different

Slerp, not lerp

Nothing repeats

The number worth having, corrected

Migration — the one thing that did change

What we look at

Facets

The ordering is the argument

Where the labels come from

Why you can trust them

The taxonomy validates itself

Signature mood

The register

Per channel

Materials

An absence worth noticing

How we differ

House style, and the flex

Closest to our own eye

Furthest

Most creative range

The gap

The gap, drawn to scale

Order is time, not hue

The pale tail is what the filter cost

One outline on the right

What the published work cannot reach

The measured null: the directional asymmetry does not exist

Why the null is matched

Seven ghosts

Horizontal position measures nothing

The gap with the pictures in it

Sampled along the ranking, not from its end

Position is rank, spacing is nothing

Why 95% here and 0% on the translation field

The Colour of the Year test

The finding

Lineage — what we pinned, against what came out

What changed

What pull means

The spread

The archive's shape

Rings

Palette sheet, mosaic, tree, kinship

The refusal, the spectrum, and a null that didn't hold

The corridor, walked several ways

Eight corridors

The number to read

The archive, other cuts

Palettes, properly

The loners

Clusters nobody named

One idea, or a pile?

The most typical image, and the least

The average is a white room with light coming in

Materials

The hundred who appear once

What colour is a client?

A correction: the loners were mostly text, and half still are

What the filter caught, and what it did not

The whole archive

The archive at B0

The one graph worth drawing

Mehretu — 21 drawings from one archive

After Mehretu