V23 measures studio photographs by code, takes five independent model readings on five fixed ladders, checks them against rules, and computes a rating by plain arithmetic. Early phone photographs are read once by V23 lite, a separate instrument. V23 is the twenty-third full rebuild since the year began. This page follows one painting, T_359, through every stage, with the engine's own images and numbers.
Code does everything that can be measured. A model judges only what measurement cannot reach, and its judgement is constrained on both sides by code. Cyan marks what is measured; amber what is written.
VAE V23 (Visual Analysis Engine), version 23.8, the reference engine since 6 October 2026. One command rates one painting.
Each painting is photographed flat in the studio and kept as a 16-bit TIFF master. From it the archive makes a family of views, each built to make one kind of truth easier to see. These are the real files for T_359.
The earlier engine, V22, looked at five of these views and measured the 16-bit master directly. The current engine needs only the colour view and its tone stretch: that is cheaper, and the validation below shows it is no less accurate.
Before any model looks, Python measures the sheet. The photograph is converted to brightness, scaled so the paper reads as pure white, and every pixel is classed by fixed thresholds. The same code, the same numbers, every time.
Left: T_359. Right: the same sheet as the code classes it. Cyan is black ink (below 0.30 brightness), deep teal translucent grey (0.30 to 0.75), stone the palest ink, near black is bare paper (above 0.88).
Every inked pixel's brightness, counted into 64 bins and smoothed. A family of tone is a peak holding at least 4% of the ink. T_359 has 3: a black, a mid grey and a pale wash. A sheet with two or fewer cannot be rated above the bottom rungs for tonal depth or layering, whatever the model thinks.
How much of the sheet carries ink.
Of the ink, the share that is solid black.
Of the ink, the share that is layered grey.
Distinct families of tone, counted as above.
Broad gradation inside the solid black. Under 0.045 reads as flat, dead black; T_359 sits just above it.
Of the right edge, where ink runs off the sheet (top 17%, bottom 73%, left 48%). Recorded for the written reading; never shown to the model.
These five facts go to the model labelled as true. Each was tested against Peter's own ratings before it was kept: the count of tonal steps tracks his eye most closely (rank correlation 0.55 on the 13 development works). Two candidate measures were dropped because they did not: one read JPEG noise as texture, the other fired on good sheets.
V22 measured the 16-bit master itself, and its records remain in the archive for 232 works: 557 ink masses traced to their outlines, 1,194 enclosed voids, 4,215 edge segments classed. Here, on T_359: the outline of its one principal mass (cyan), the 5 voids it encloses (amber circles of their true area, placed in their grid cells), and the path of its main gesture from start to end (red, 13 turns). Of 17 edge segments, 47% are crisp, 41% soft or bloomed, 12% broken.
A vision model (Claude Opus 5.5) is shown two images and the measured facts, and places the painting on five ladders. It never gives a rating, and never sees one. It does this five times, independently, and every read is checked by code.
Left, the target board: T_359 in colour above its tone stretch. Right, the reference board: three paintings Peter rated, the same three for nearly every painting the engine reads (R1 T_140, a weak sheet; R2 T_206, a middling one; R3 T_334, one of the strongest). The model judges against these three sheets. That is how the scale stays anchored to this body of work.
Every rung has a name and a written definition. Rung 2 is what the median sheet in this archive does; rungs 3 and 4 need evidence that is rare here. Close calls resolve down. Highlighted: where T_359 landed. Form, Tonal depth, Layering and Composition are the critical axes; Material records how ink and paper behaved.
Each read is a separate call with no memory of the others: a first impression in one read cannot bend the next. On T_359 the five agreed on every axis. Across all 252 paintings, the five reads put an axis on the same rung 79% of the time, were one rung apart on the rest, and never further apart than that.
Radial arc bands follow brush travel with dry-brush ends; the big dark block right of centre and the tall right bar read bulky and slab-like.
WEAKEST Large dark slab-like rectangle right of centre
STRONGEST Curved arc bands around the left radial hub
Three tonal plateaus; dark interior modulation 0.047, barely above flat. Fan greys are stepped and grained, but the main dark masses stay fairly uniform inside.
WEAKEST Dark rectangle right-centre, nearly flat interior
STRONGEST Stepped translucent greys in upper-right fan
Many translucent passes fan out from two hubs, each blade still readable where it overlaps; the sequence builds depth from pale outer blades to dark inner masses.
WEAKEST Dense dark overlap near the left hub
STRONGEST Overlapping translucent fan blades, upper right
Two radial centres drive a clear left-to-right sweep. Ink covers 80% of the sheet and the right half is jammed with dark blocks, so the empty paper is mostly leftover corners.
WEAKEST Crowded, heavy right half with little breathing space
STRONGEST Rotating sweep from left hub toward upper right
Ink is controlled and the dry-brush edges are clean. There is a small round spot at the right edge and faint specks on the left paper; nothing damages a principal mass.
WEAKEST Small dark drop at the right edge
STRONGEST Dry-brush streaked blade ends, upper left
Each read must name every flaw and where it is, in a ninth of the sheet. Across five reads of T_359: 2 distinct places, named 10 times in all, 0 of them major. Five reads agreeing on a flaw is stronger evidence of one flaw, not five flaws. A major flaw holds Material at rung 1. Across the archive: 1,205 distinct places from 3,788 mentions.
The model's answers pass through code that it cannot argue with. Measured facts outrank its reading. Then the number is computed by fixed arithmetic.
Tonal depth held at rung 1.
T_359: 3 steps, rule not appliedMaterial held at rung 1. Mild waves in the unmounted paper are never a flaw.
T_359: no major flaw namedThe rating cannot exceed 4.
T_359: lowest axis 2.0Any read that breaks the strict reply format is discarded.
T_359: 5 of 5 validThe scale was set on 24 paintings chosen at random from the archive, before Peter rated them, so that the median sheet scores 5, the tenth percentile 3 and the ninetieth 7. No constant was fitted to Peter's ratings. A 10 is never awarded: the top of the scale is left open.
The whole run for T_359 took 11,169 tokens and about a minute and a half. The previous engine spent about 90,000 tokens per painting and agreed with Peter less well. Every run is stored with its prompt fingerprint, its five replies, the reasons for each rung and every flaw, so any number on this site can be traced back to what produced it.
An engine that agrees with itself is not enough. It has to agree with the eye that made the work, on paintings it was never tuned on, and it has to say the same thing twice.
Left: Ten works across the engine's range were set aside, scored by the engine (version 23.6) and sealed. Peter then rated each by eye, wrote his reasons, and registered both (2026.09.30) before the seal was opened. Rank agreement 0.95; average miss 0.76 of a point; average lean −0.13. No sheet Peter rated 4 or below scored above 5. Shaded: within one point. Read this as promising on a small test: ten paintings, one rater, on engine version 23.6 (since then, 23.7 began reading colour ink as deliberate material, which touches one of the ten, T_310, and 23.8 changed only what is recorded), and the average miss of 0.76 is just outside the 0.75 bar set beforehand. This measures agreement with Peter; repeatability, below, is a separate question. Right: The 13 development paintings, V22 (hollow) against Peter: average miss 1.39, a lean of +1.18 (it flattered), and it rated dark, clean, lifeless sheets 6.8 to 8.1 that Peter rates 2 to 5. On the same 13, V23 misses by 0.92 on average with rank agreement 0.86. Each mark sits at Peter's rating across and the engine's up; the line under it is the miss.
Two complete passes over the 13 development works, rank against rank. Average change 0.22 of a point; 6 of 13 identical. A difference under half a point between two works is noise.
Paintings Peter rated, used to choose between designs. Every design was scored on them and is logged, including the ones that failed.
Never seen during development. Peter's ratings registered before the engine's were revealed.
Versions 23.0 to 23.8, each a single change, kept or rejected on the numbers.
Average miss against Peter on the 13 development works for each design in turn (the path; lower is better), with rank agreement beneath each. The honest caveat, recorded at the time: the prompt was revised once after seeing these results, so they are development figures. The held-out ten above are the real test.
V23 is a count: the twenty-third full rebuild of the engine. It was never designed in one sitting. It has been under construction for as long as the paintings have, from the first reader in the opening weeks of the year to the version running now, and most rebuilds went through many variants of their own before they were replaced.
Each cell is one full rebuild of the engine, in order. A rebuild began from what the last one got wrong; inside each, variants were tried and discarded, sometimes by the dozen. Where the record keeps their names they show the depth: version 9 ran to 9.2.4, version 16 to its eighth lettered variant, version 18 through four variants on two models, version 23 to 23.8. More than forty named versions and variants survive in the record, each with what failed. Hover or tap a cell.
These are the turns that mattered.
Descriptions drifted generous: the model admired instead of judging.
A firewall between the judgement and the description, so praise cannot raise a number.
Fourteen gates proved too many; five sequenced tests replaced them.
Three models compared on the same works. The engine measures the floor reliably and cannot honestly reach the ceiling.
Twenty-plus versions on three models, logged failure by failure. The basis of the FINDINGS report.
An outside critique of V20's fourteen categories rebuilt the frame: material, motion, structure, perception.
Python measures the 16-bit master before any model looks; 14 axes, code ceilings.
One eighth of V22's cost, and agreement with Peter's held-out ratings at 0.95.
Colour ink read as deliberate (23.7); every read records its strongest part and each flaw's place (23.8).
The tonal-step rule had held down dense, many-layered sheets; it now limits Tonal depth only, chosen on eight paintings Peter rated blind (average miss 0.55). Mildly wavy unmounted paper is never a flaw.
Seven reports sit behind the ladders, the rules and the way the model is asked. Some were commissioned syntheses of the literature on ink, brush and abstraction; one was written from the engine's own failures. All are kept in full in the archive.
Why the engine is built the way it is. A model's first impression bends everything after it, including its own self-checks, so the engine takes five separate reads instead of asking one read to doubt itself. Three-level scales collapse competent work into the top level, so the ladders have five rungs. The model never sees a score, so it cannot aim at one.
Rated each category for validity, merged the ones that measured the same thing, and grouped what remained into four domains: Material and rheology, Kinematic and temporal, Spatio-structural, Perceptual and contingent. V21 was built on it.
How a mark is read: brush loading, speed, pressure, dry-brush (feibai) against exhaustion, and a seven-level hierarchy of execution. The vocabulary for reading a single mark.
Edges read relative to the paper's absorbency; a five-level evidence scale (observed, technically inferred, process-inferred, interpretive, speculative); criteria for overwork and for the stopping point.
Composition without subject matter: balance, tension, figure and ground, the role of the void. The vocabulary for reading a whole sheet.
Arnheim's perceptual dynamics and other ways of naming what makes an abstract sheet hold. A second, separately written answer, kept as a check on the first.
What each paper and ink in the archive physically does, written for the engine with every rating removed, so the model learns materials without learning preferences.
Stated plainly, because an instrument is only as credible as its declared limits.
It reads photographs, not paintings. Scale, sheen and the physical presence of a two-metre sheet do not reach the camera.
It measures the floor better than the ceiling. It reliably separates a failed sheet from a sound one. What makes the very best work exceptional is, by design, left open: no painting is ever given a 10.
It disagrees with Peter on some works, consistently. A few strong sheets it reads as blunt or flat inside, pass after pass. That is a real disagreement, and it stays in the record.
The earliest works are read differently. The 143 works with no high-resolution master exist only as older phone photographs. They were read once each by a lighter version of the same engine (V23 lite). Its numbers are kept as their own series and never averaged with these.
Peter's own ratings never set the number. He rates privately, before he sees the engine, so any gap is honest. The engine is the reference; his eye is how the engine was tested.