Peter YuillThe Crucible Year

The Oracle

Oracle · deep archive

Oracle Constellation

Every rated painting of the year and every pattern the Oracle found in the records, drawn from vae.db. Facts and correlations only: what the data shows, not what it means.

V23, studio photograph, five reads V23 lite, phone photo, one read. Same scale, never averaged with V23 Peter's own rating, where given

Right eye, cyan: reads the data. Measurement, structure, the engine's ratings Left eye, red: reads the room. Peter's own eye, and where it parts from the engine's

FIG 01

The year

Every rated work is one ray, in the order it was photographed, round Bangkok's three seasons from the first batch in December 2025 to the latest. The longer the ray, the higher the rating; solid rays are V23, hollow-tipped ones V23 lite; the inner tracks are the five axes and gaps split the batches. Ring by ring lays the same works out as tiles, one ring per quality, to follow one quality round the year.

FIG 02

The rising floor

Each quality V23 reads, as a channel through the year. Every work is a mark at its rung; the banks are the floor and the ceiling, where the lowest and highest tenth of the works sat at the time. A weir marks the day a quality stepped up for good, and the notes name what the Oracle found held. Read the lower bank: the weakest works improved more than the strongest.

FIG 03

Pace and surface

The pace of making, as a pulse: for every day of the year, how many works had been dated in the fortnight before it. The shaded storms are the bursts, the busiest quarter of the year. Below the line the batches answer: the Material and surface rung each was given, and how often the reads found a water stain. Batch against batch, this is the strongest pattern the Oracle has found in the practice itself.

FIG 04

Flaws

Every flaw V23's five reads named, drawn as it behaves on paper. The sheet shows where the reads placed them, every kind together; below, each kind with its own sheet and, along the year, the share of each batch's reads that named it. V23 lite names no flaws, so the early months are absent. Hover the sheet for a zone, or a batch for its share.

FIG 05

What goes with what

Every finding is a strand from a fact on the rim to the score it goes with. Cyan goes with higher, ember with lower. The bright strands held up; the haze behind them are leads, still to be tested on new work. A fact read over several windows (3, 7 and 14 days) is one point on the rim. Hover or tap the rim to follow one fact or one score.

FIG 06

Five reads

V23 reads every work five times and takes the middle. Each work here is a column of five marks per quality, one per read, in the order the works were photographed: a column that sits as one block was a clear call, a split column a close one. Above, how few changed reads would have moved the rating a full point. Below, the rungs the measurement caps took away, and the run whose batch rated furthest above Peter.

FIG 07

Two engines

V23 reads the studio photographs; V23 lite reads the older phone photos. Thirty works were read by both. On the left, each of those works as a strand from its V23 rating to its V23 lite rating. On the right, the factors the Oracle tested against both ratings that went somewhere, each a needle: up for higher, down for lower, its two halves the two engines. Where the needle runs straight the engines agree.

FIG 08

Against the pattern

From everything measured about a work, the Oracle predicts each score V23 gave it. These are the works that broke the prediction furthest: the hollow mark is where the profile said the score would be, the solid one where V23 put it. A fall is a work that measures well and was read low; a leap, the reverse. Hover or tap a work to see it.

FIG 09

Each paper, its own year

One constellation per paper, grouped by how the paper is sized. Each ray is a painting on it, at its date on the year's clock (December at the top, running clockwise) and long for a high rating: solid rays V23, hollow tips V23 lite, a red dot Peter's own rating, the base dot his early triage. Inside, the five qualities V23 read on that paper against every paper (dashed), and a ring of the inks it took. Beside each, the spread of its ratings and what the Oracle found about the paper itself. Papers were used in different months, so read these as portraits, not verdicts. Hover or tap a ray to see the painting.

FIG 10

Beliefs under test

Each lane is one belief about the work, tested against the whole archive. The dot is what the data says now, the line its uncertainty, and the faint trail behind it where the same test stood in each of the Oracle's earlier runs. Right of the band the data agrees, left it disagrees; a "no effect" belief holds while its dot stays in the band.

FIG 11

The words

How often Peter's own writing reaches for each idea, per thousand words, from the first records of the practice to the Crucible year. Each phase is a layer laid down on the last, oldest at the bottom; each idea a vein whose width is how often it was written. The marked shifts held up: a rise or fall too large for chance across the sections written in those periods.

FIG 12

The sky over the table

The year drawn against the real sky. The rim carries the twelve signs; inside it runs the band of sky the Sun crosses, with its actual stars, constellations and Milky Way, which have slipped almost a sign from the signs named after them. Every painting sits where the Sun stood on the day it was photographed, its ray long for a high rating, so the year runs from Capricorn round to Libra and the dark arc is the part still to come. Inside, the planets' paths across the same days, each drifting outward as the year goes on so that where a planet doubles back (retrograde) it draws a loop; the Moon's face on every photo day; and at the centre the Moon itself, from real imagery, as it stood that day. Hover or tap a painting to see the sky on its day; hover a planet for what the Oracle found about it.

FIG 13

The sky, and a borrowed sky

Every sky pattern the Oracle recorded, drawn as stars above the horizon, and below it, like a reflection, the same tests run on a borrowed sky: the real positions with their dates moved half the record away, so they belong to no painting. Each mark is one pattern, placed by its strength; the large stars passed as signals, and each is named. The dashed line on each side is the other side's outline, flipped across: if the planets had a hand in the work, the sky would outgrow its reflection. The two shapes almost coincide. Hover or tap a mark to find its twin.

FIG 14

The Oracle's own year

The Oracle has run several times as the archive grew. Each ring is one run, the first at the centre; each finding a thread out from the run it was first tested, lit where it held as a signal, dim as a lead, dark where it fell to noise. The outer dark is what the Oracle could not test at all.

The Oracle

What the archive notices.

Every painting is dated, its materials recorded and its surface measured. The Oracle reads all of it at once and looks for patterns that hold across the whole year. 354 have held up so far: 229 from the run drawn here, and 125 from a second, separate search of the same records.

Everything on one pageThe full Oracle ConstellationAll fourteen figures together, with every work's ratings.

Every test of the latest run, one dot each 229 held as signals1,484 leads12,427 fell away

What the Oracle is.

Plain statistics in Python, with no AI and no opinions. It puts everything recorded about the paintings and the days they were made into one table, tests every possible connection, and keeps only what survives strict checks against chance and against time. It does not know what a good painting is; it only reports what holds. Latest run: 2026.10.26.

ONE RUN

From records to findings.

Every number below is from the latest run. A finding has to pass every stage; most do not.

1 COLLECT11 sources

The skymoon, planet signs, retrogrades and aspects, tested like any other source119

Placebothe same sky with its dates shifted half a year: the chance check119

V23the reference engine's ratings, rungs and measurements32

The registrydate, size, paper, ink, batch of every painting31

V22 geometrymasses, voids, edges, gestures from the 16-bit masters30

V23 litethe early works, read from phone photographs14

Weekly logPeter's own record of each week6

Weatherthe city's daily weather5

Sequencewhere each painting falls in the order of the year4

Whole yearone rating per painting across all of them, system named2

Stepshow far Peter walked each day1

2 ONE TABLE402 paintings

One row per painting, 361 columns. The 316 days of the year join each painting through the 3, 7 and 14 days before it was made.

3 TEST14,140 tests

Associations every predictor against every outcome, raw, within one paper, within one batch
Time trends trend, change point, floor and ceiling across the year
Batches paintings photographed on the same day compared as a group
Prediction can anything forecast a rating at all, tested on held-back works
Outliers paintings far from what their circumstances predict
Streams where the two rating systems overlap, kept apart

4 GRADE229 signals

229Signal: survives correction for the 14,140 tests run (false discovery rate under 5%) and the effect is real in size (rank correlation 0.20 or more)

1,484Lead: suggestive, worth watching, not claimed

12,427Everything else: noise, discarded

THE PLACEBO0 vs 16.0

The Oracle assumes nothing about any source, the sky included: the moon and planets are tested like the weather. To check itself it also runs a placebo, the same 119 sky positions with their dates shifted half a year, so they are linked to nothing. By chance about 16.0 of its tests should pass as signals; 0 did, inside the margin the Oracle allows. Had more passed, the signal threshold would have been halved for the whole run.

TIME HELD CONSTANT30 days

A pattern only counts if it still holds between paintings dated within the same 30 days, keeping at least half its strength. Otherwise it is just the year changing.

BATCHES ARE ONE UNIT49 photo days

Paintings photographed the same day share their day. Anything shared by a batch is tested batch against batch, so one prolific day cannot count as twenty pieces of evidence.

RATINGS ARE OUTCOMESnever inputs

An engine's reading can be explained, never used to explain. Only measured facts and recorded circumstances may predict.

The Oracle runs again whenever the archive grows. A finding that fades on new paintings is marked fading; one that holds on paintings registered after it was found is marked confirmed. How each painting was measured is on the Method page.

What it found.

The clearest pattern first, then how each quality moved, then every pattern at once.

How sure it is.

How close each of the engine's calls was, whether its two versions agree, and how every finding has held from one run of the Oracle to the next.

What it was asked.

Statements about the work, tested against the whole archive, each with its verdict and where it stood in every earlier run.

The sky.

The Oracle rules nothing out in advance, so the moon and planets are tested like the weather. The year as the sky saw it is drawn in chapter 1; here is what came of the test.

Everything on one pageThe full Oracle ConstellationEvery figure on this page and the rest, fourteen in all, together on one page in the Archive, with every work's ratings and how to read them.Open it →