How to read a reading
A structural report is forty-odd numbers. On their own they are inert. This page takes one work — Matisse's La Leçon de piano (1916) — down the four layers in order, and names each measurement where it lands on the canvas. Read this once and the thirteen studies become legible.
The discipline starts with the eye. Look at the painting and say what you see, in words, before the first command. The numbers then test that reading instead of seeding one. What follows is the reading after the looking, not a substitute for it.
Layer one
Everything begins with nine per-pixel fields computed over the whole surface. Four are primary — the ones regions are read against — and the rest are substrates for specific measures. None of them knows what a piano is. They know where the surface changes, how sharply, in what channel, and whether the change survives blurring.
The mass field is the one to watch. It is a weighted blend of the other
three, lightly smoothed, and it is what the region pass carves up. Read against the
reproduction, it already tells you the painting concentrates its weight in a single
left-of-centre body and leaves most of the surface quiet.
Layer two
A single call, read_image(), runs one region pass — a watershed over the
density field — so that the kernel coordinates, the relation system, the colour and tonal
measures, and the state classifier all describe the same regions. Here is what it
said about this painting, grouped by what each answer is about.
Five construction families (edge, tone, colour, perturbation, mass) are each asked where the weight is. Ω measures whether they point to the same place.
Up to twenty watershed islands, each with a role (anchor or satellite), the corridors between them, and how stable the whole arrangement is under a crop. Counts are capped, so entropy — not the count — is the comparable quantity (finding I-17).
State: counterweighted — "a region reads as a load-bearing balance term more than a construction outlier." The painting is held in tension between an off-centre main body and a smaller counterweight, across a surface that is mostly quiet. That is a structural fact, not a judgement: the classifier never grades.
Layer three
The spine reads the load-bearing structure. Some things it is built to pass over — small marks with leverage out of proportion to their mass. The accent lens is a separate instrument for exactly those: it works at 900 px, returns the top marks by leverage, and reports each one's kind and its role in the delay-grammar of the eye.
Other lenses answer other questions, each pinned to its own working scale so its calibrated constants hold:
| Lens | Question | Invoke |
|---|---|---|
| Stance | Does the centre of gravity keep changing heading as detail dissolves across scales (torque), or hold one heading (stand)? | isr stance |
| Layered delta | What kind of "same" are two images — same pixels, same tone, or same spatial skeleton built from different pixels? | isr delta |
| Compare | Across 2–20 images in one coordinate space, where does this set genuinely differ? | isr compare |
| Position | Absolute displacement from a fixed geometric origin on 18 axes — cross-domain coordinates with no reference corpus. | isr position |
Layer four
When a study needs a move the spine does not have — an emergence index, a substrate audit, a null control — it is written once, generalised so it takes any image, and shelved. The next study greps the shelf before inventing anything. The shelf answers with less authority than the spine, so it is checked, not deferred to.
Above the shelf sits the candidate register: the machinery that decides what is ever allowed to become a permanent spine metric. A measure must clear a promotion bar across the whole corpus, including a null-control check, before it graduates. Most do not — and the register records the declines as carefully as the promotions. That is the subject of the method page.