Every page here applies one commitment: read the geometry of a structure before assigning it meaning. No model in the loop, no semantic judgment. The same habit of measurement has been pointed at a Caravaggio, a wafer map, a seismic array, a tumour, a robot's success judge and a stock market.
This is the whole body of work, grouped by the field it was pointed at. Measured holds the studies, papers and lab books, with corrections and retractions left in. Made holds the studio tools, instruments and games built from the same measurements.
16Fields
–Papers
–Lab books
–Critiques
–Pages in all
The field map
Each tile is a field the instruments were pointed at. Choose one to jump to its work.
Studies, papers and lab books. Each field has its own instrument, but every instrument makes the same commitment: deterministic, interpretable coordinates, read before meaning is assigned.
A pre-semantic instrument aimed at works whose meaning is already argued over. Each study has a lab book, the working record with every correction kept, and a critique, the essay it earned. The full corpus, with notes on what each study taught the instrument, is on the studies page.
The sixteen studies
Study 01 · paintingMatisse, La Leçon de pianoColour and structure decoupled, made into instability.
The fields the reader computes (luminance gradient, OKLab chroma, cone-opponent channels, committed and soft edges, structural mass) and how regions are read as one system.
Does the structure reader add predictive information beyond simple pixel, regional and frequency features? Six pre-registered tests with planted controls. The decisive one, 859 steered images in 106 groups, found +0.003: structure added nothing there.
A reader of pictorial space: one picture in, readings of its planes, groups and field, and where its own local readings refuse to agree. Its pre-registered claim, that the relief mask reproduces beyond a trivial mask, did not hold, and the page says why.
02
Generative image models · Visual Thinking Lens
Image models repeat the same geometry.
The lens measures the compositional prior that image generators apply to every output, whatever the prompt asks for. It is the field this work began in.
A label-free count of compositional vocabulary. AI-default portraits use a narrower vocabulary of surface and colour than pre-modern oil portraits, and an equally wide vocabulary of placement.
The operating windows of one image generator, measured object-level and mass-level: 864 Z-Image Turbo images, two frozen tests and three exploratory dives.
Standard image metrics read what a picture contains, not its visual vocabulary. CLIP and the kernel side by side on two subjects, each made three ways: CLIP is indifferent; the kernel separates them.
Compare two complete sets of generated images on 17 pixel-traceable axes, with effect sizes and bootstrap intervals. Images never leave your device; no model judges another.
A pre-registered test on 1,080 images: the placement measures transfer to two held-out engines (46 of 48 directions), but the same prompt words make different compositions on different engines.
Compression rejects byte-level sameness across the set, while the lens finds many near relationships in spatial structure: different substrates, kindred composition.
Early evidence from a closed-loop camera experiment in Blender: of 24 deterministic paths, the searched one met the release brief at 24.8% lower loss than the best ordinary moving control. Not yet tested against random search.
One authored Blender shot revised four ways through the scene itself (relight, blue, gold, cleared), set beside frame-by-frame FLUX image editing and a masked baseline. The scene keeps what the frames lose.
Same controls, same available poses, different selection: picking frame by frame passes 17 of 24 mechanical checks; choosing the sequence as a whole passes 24 of 24. One development shot; artist review still open.
Across 495 rendered frames, the camera moves 25 of 38 measured coordinates, mostly through geometry rather than light; focal length is the strongest control. Two pre-registered conclusions were overturned by later checks, and both corrections are shown.
04
Language models · Linguistic Kernel
Text has shape.
The same move applied to text: reduce an output to a small set of structural coordinates and measure it, without asking another model what it means.
What a deterministic text reading can tell apart: prompt against run, engines under batch load, answers cut off mid-formula, and which instruction wins when instructions collide.
Clause density in a finished answer tracks how unevenly GPT-4o's own token-level uncertainty moved, with no access to the model: r = 0.36 after controlling for length and prompt family.
1,200 Gemini, GPT and Claude responses under twenty-prompt batch load: how each engine allocates output, and the rendering-intensive derivations that cut off mid-formula.
93.4% SEM balanced accuracy with no learned parameters. The same coordinates transfer across SEM, wafer maps and aerial lithography without retraining.
A slowness-ratio diagnostic across three arrays, including an artifact peak 46° from the expected Tōhoku P arrival.
08
Precipitation observers · Structural collocation
Disagreement as a measurement.
Radar, infrared, microwave and rain gauges read the same rain field. Their disagreement about its geometry estimates each one's error, and shows when two of them share one, without ground truth.
Triple collocation on rain-field geometry over CONUS: the observer ranking depends on the triplet, and replacing either satellite with gauges reverses it.
A domain-qualification test that closed negative, and the collocation thread it opened, replicated on an unseen month with all six pre-registered predictions met.
A coarsening step that did not conserve mass gave independent observers correlated errors, and a finding built on it was withdrawn. The mechanism, the checks that would have caught it, and a checklist.
A two-view geometric check caught 10 flat printed decoys that a frontier embodied VLM certified as real 3D objects. 19 object types, real phone pixels.
The backward trail: what it started with, what was wrong with it, what was built, and where it honestly loses.
11
Human movement · ZTOYBOX Biomechanical
A negative result, kept on the record.
Can organisational measurements reveal something about the body that produced them, beyond the complete conventional read? The computational phase closed without earning that claim, and says so.
552 trials from 28 participants and 29 linked questions, each attacked by an adversary. The physical claim is not earned; what remains needs a prospective physical experiment.
12
Financial markets · Ω, multi-family
Concentration beyond amount.
A transfer test: does the seismic idea of independent structural reads carry over to the cross-section of stock returns?
Every number is read programmatically from the saved results; each finding is labelled by the strength of its evidence. Version 2 lists what changed since the first draft.
The study record: a generator whose every output pixel is an exact, owner-labelled blend of training pixels, after Kamb & Ganguli's patch-mosaic result.
How far can steering move a contributor's expected credit while every receipt stays honest? A per-contributor bound passed a frozen test in 95% of pairs; a simulated $10,000 settlement moved $266.89, and the largest single loss stayed inside its stated exposure.
Why this pixel? Rebuild a damaged face from source portraits and explicit rules, then remove any one and see what changes. An exact-duplicate test shows that a large matching weight does not mean a source was necessary.
Six experiments in what "similar" means: move, shuffle and recolor an image, then see which scores notice. A second lab trains a small diffusion model in the browser and traces outputs back to images and patches.
What "credit" can mean, formally and in plain language, across transformer and diffusion stacks, with small-model controls. Toy quantities, not production influence or legal allocations.
Eleven linked chapters over sixteen synthetic images: watch a toy denoiser draw on an indexed source memory, with exact, estimated and logged quantities kept apart.
An exact, owner-labelled generator and an instrument for taking a source out. The October run tested whether a collection's creditability can be predicted from its make-up: neither predictor could, and under pixel error exact credit came out nearly uniform.
A picture from an exact generator trained on six public-domain draughtsmen and a pool of museum paintings. Remove an artist, see what changes, and compare who gets paid under exact Shapley, usage and leave-one-out.
A mathematical response proposing an exact observation law for a learned attention block, with two propositions, proofs, numerical checks and a testable programme. Draft for technical review.
II · Made
The measurements, turned into things to use.
Tools for painters and draughtsmen, instruments for the ear, and games that make an idea walkable. Each is built on a measurement from Part I.
14
Drawing & painting · Studio tools
A machine that can make. And reconsider.
Making systems where every move is recorded and a critic argues back, plus pocket looking devices for the gallery and the easel.
A composition editor where decisions persist: change emphasis, frame, arrangement or spacing, drag to propose a position, then see the conflicts and choose how to resolve them.
Before, after, and a declaration of what was allowed to change: a reproducible receipt of what an SVG edit actually touched. The final code flagged all 4 edits reviewers called material and none of the other 20.
Move a painting's color center while locking hue-family chroma, lightness and a protected region. 22 of 42 requests solved under the full contract; the other 20 refused as infeasible, with a check for each refusal.
Recolor one layer and move the fewest others to keep the composition's color center: 9.6% fewer collateral layer changes than a strong baseline, for 2.9% more movement, across 48 Photoshop requests.
Your pile is almost right. Say what it needs by eye and see what to add, how much, and what else it will do; it learns your tubes, scales the batch, and shows the geometry underneath.
Squint at a reference, cut it into a few value masses, and get one pile per mass from a short shared palette, each checked so it stays in its value; with accents kept and a sheet to print for the easel.
Eight paints and a world between them. Find a starting mixture, see which way it leans, and make it your own; studies and reference images stay in your browser.
Tap a colour in a painting or your own photo and get an estimated oil-paint recipe from the paints you actually have. Adjust it, push it lighter, warmer or quieter, and save the mixes worth keeping.
A working notebook for painters: find a way into an image through line, mass, value and colour studies, keep a direction, and follow the painting through its sessions.
A town with no ground, and a Mill that wants it. A turn-based town builder drawn in pen and ink, where anything that loses its path to an anchor falls.
Peek at real combinations of six artists, split 100 coins, then see the exact answer. In every round and under every measure, paying by usage or by leave-one-out scores worse than splitting evenly.
A shadow mystery in twelve rooms. Choose where to shine the light on a hidden sculpture, rule out the forms that don't fit, and learn when light alone cannot decide.
A copper loop that makes sense at every point and climbs all the way around, yet arrives where it started. The depth it gains is what the picture cannot hold.
Your camera image breaks into a grid of moving cells. Each cell measures its own motion and stretches its piece of you along it; the more you move, the further the mirror comes apart.
