"And what sticks around will be what's interesting."
Interestingness is the improbability of the configuration. Each metric is a line; a line pressed far out of the middle (a hard press) is improbable by luck — the fingerprint of a choice. So is every line landing dead-center: not luck, a prior. The mushy middle is the probable baseline — junk, quick to call. One hard press is discipline; several coordinated is harmony; many is mastery — a parabola almost nothing reaches. Coherence stays the junk gate (no map → junk whatever it presses). Works on text, signal, and image; deterministic; the field is the frame, and it can be challenged.
the field — prior corpus · one example per line · novelty is distance from this ▾
The engine learns the "expected" from these lines. Change them and the verdict changes — a black square is a cliché against a prior of shapes and a rupture against a prior of paintings. The frame is part of the measurement.
field — prior of simple signals (trend · flat · single cycle · light noise). Novelty is distance from these in shape-space; each model read is held to a √n noise floor so short random series can't fake a fit.
Drop an image · or click independent reads (the Parallax principle); novelty is distance from a field of ordinary images
field — figurative domain model: 42 real figurative images (AI standard + interesting + museum), with the 6 features computed in-browser the same way a dropped image is, so the null is real and the comparison is parity-clean. A press is a departure from the figurative norm. In-browser uses the 6 cheap features; the high-fidelity VTL gate is the batch Python path.