Why trust it
A deterministic tool gives the same number every time, which is necessary but not sufficient — a reproducible number can still be measuring the wrong thing. The rules below are what keep a reading honest. Each was earned by a documented correction, and several of the sharpest results in the corpus began as caught errors.
Before the first command: look at the image and say what you see in words. Write down what you expect each key metric to say, and why. Name which metrics will have nothing to grip on this particular work. Carry no target.
Then the numbers test a hypothesis instead of seeding a narrative. A surprise you predicted against is a finding. A surprise rationalised afterward is a story. The lab book is built first; the critique does not begin until the reading is solid.
The R_spatial value was pre-registered as subadditive in the 0.2–0.7 band before any command ran. It landed inside the window. Because the prediction came first, the result is evidence rather than a number talked into meaning something.
A whole-image scalar has purchase on some works and floats on others. The same metric can grip one image and read only global statistics on the next. There are two ways to tell:
This is a per-work judgment, never a verdict stamped on a metric. The rule (finding I-15) is not "β is distributional" — it is "recognise, per work, when a number measures the tool instead of the work."
To ask whether a number needs the intact composition, destroy the composition in a controlled way and see if the number survives. The corpus uses three families, and their disagreement is itself informative — they bracket a question rather than settle it.
| Null family | What it keeps, what it destroys |
|---|---|
| Phase-scramble | Keeps the histogram and the power spectrum; destroys the arrangement. Permissive — it also destroys figures. |
| Patch-shuffle | Keeps the histogram and local texture; destroys the composition. |
| Figure-position shuffle | Permutes named figure boxes; asks whether the real arrangement beats rearrangements of its own cast. Conservative — it cannot vacate the centre. |
A value that survives both scrambles is distributional. A value that dies on both needs the intact composition — the strongest "real" evidence. R_spatial passes; a colour-in-void ranking fails (it is invariant under scrambling, so it is a histogram fact, not a composition fact).
Single null draws are seed-chaotic — one draw flips on luck. The real signal is that the measured value is cut-stable while the nulls are seed-chaotic, not a single point-drop. This retired an earlier single-draw claim and replaced it with a 100-seed sweep reporting an empirical percentile. Null tightness is also image-class-dependent: generated images scramble roughly 3× wider than paintings, so percentile thresholds do not transfer across domains without a recheck.
Every measurement is of (object × reproduction chain × scale window). A photograph of a painting is not the painting (finding I-10). Dense-area measures survive resolution sweeps; thin-structure measures do not. Every number is reported with the resolution and physical scale it came from, and a claim that has not survived a resolution sweep is labelled a reading aid, not a result. Crops are different measurements.
A measure invented in a study does not become part of the instrument by being interesting. It enters a candidate register and must clear a promotion bar across the whole corpus — including a null-control check — before it is ever allowed into the spine. Most candidates do not clear it, and the register records why they were declined as carefully as it would record a promotion.
| Candidate | Decision | Reason on the record |
|---|---|---|
| C-3 R_spatial | Not promoted, kept as a pattern | A torque-emergence axis. The signal is the range-width, not the point value, and no operational rule yet discriminates cut-robust from cut-sensitive works. Promotion waits on that rule. |
| C-4 address | Not promoted, kept as a pattern | Null-validated and it discriminates, but the raw scalar misleads alone — it needs an n≥100 null sweep attached — and its fine sub-signals hit the pre-semantic ceiling. |
| C-1 anomaly | Resolved externally | The Degas structural anomaly resolved into a documented figure only through outside scholarship. The instrument pointed; it could not name. |
An empty register would mean the promotion bar is not doing its job. The declines are the evidence that it is.
An early read of the Degas returned zero structural accents. It looked like a clean null — a painting with no punctuation. It was not: the accent lens was operating at the wrong scale for that brushwork. Corrected, the same painting returns a full accent field. The episode became findings I-5 and I-7, and the standing rule: when a result surprises — especially a null — look under it before believing it. Apparent failures are not discarded. Several became the sharpest results in the corpus.