Structure has geometry. Measure it.
Wherever a system generates structure (an image model arranging visual mass, a language model producing text, tissue organizing under a microscope), that structure has geometry, and geometry can be measured without knowing what the output means. Parallax Metrology builds deterministic structural instruments: no model inside the measurement, same input, same output, every coordinate traceable to countable evidence. Where a benchmark result comes from a fitted readout on those coordinates (a classifier or regressor), the repository says which.
| visual-thinking-lens | Where visual mass sits in an image, how it spreads, and whether a generator ever leaves its default. Python package, CLI, notebooks, agent skills. Try it in your browser → · or play Where's the weight? |
| linguistic-kernel | The structure of text, especially model output: where an answer puts its weight, how it breathes, how it holds together, and when it's been cut off. Eight countable axes, no judge model. Play: Guess the Engine → |
| inspection-coordinates | Fifteen physics-grounded coordinates for semiconductor inspection images (SEM, wafer maps, aerial). No training; matches the best classical descriptors on public benchmarks, reproducible in minutes. Process-monitoring pilots welcome. |
| spatial-kernel | Is it really there? A deterministic two-view check of robot task outcomes on ordinary phone photos. A frontier robot VLM passed 10 of 18 printed decoys as real objects; the geometry caught all 10. Play: Real or printed? → |
| structural-topology-kernel | Fifty training-free structural coordinates for microstructure images, with a matminer featurizer. 0.832 AUC on steel micrographs grouped by specimen (HOG 0.728); the classification, bainite and permeability results reproduce from shipped feature tables. |
| parallax-core42 | Forty-two named, deterministic measurements of an H&E histology patch, built to find the batch you didn't know you had: which named axes split a cohort by site, scanner or production route. On TCGA colorectal slides it separates three production routes (balanced accuracy 0.977 with a named LDA readout; the mis-scaled derivative slides stand apart without labels). Complements colour features, not a better classifier than plain texture; see 'How it compares'. |
| notan | The big light and dark shapes of a picture, read the way a painter squints, and which parts of it are built from value and which from colour. Texture-robust; nine test images with known answers ship with it. |
| notan-route | The order in which a picture's Notan masses lead from one to the next, and the transition field behind it: which masses lead to which, and the basins they drain into. A property of the mass map, with no claim about eyes; ten known-answer gates. |
More instruments (pathology and others) will open here one at a time, each built the same way: faithful to its research notebooks, tested against them, and documented with what it can't see.
- How image measures lie quietly (code): nineteen ways a hand-built image measure returns a plausible, wrong number without raising an error, each with a case, a live demonstration and the check that catches it.
- he-checks: quiet failures in H&E pathology pipelines and datasets. A Macenko stain normaliser whose haematoxylin and eosin vectors collapse into one direction through an SVD sign convention (0.17° apart in our own TCGA work, about 14° once fixed), 25 TCGA colorectal slides whose stated pixel spacing is about 2× too fine (with their enlarged tiles in the Kather TCGA tile archive), and purple "blank glass" in every background tile of the CRC-VAL-HE-7K benchmark. Checks, a sign-safe estimator and the slide list.
- structural-collocation: a study. Triple collocation on the geometry of the rain field (radar, GOES-16 infrared, IMERG microwave, gauges): which observer looks best depends on which three are compared, and shared error can be detected without ground truth.
- paint-mixing: a tool for painters. Kubelka–Munk mixing for the questions asked at the palette: what to add to a pile that is almost right, whether your tubes can reach a colour, and one shared palette for a value plan.
- shan-shui-inf: a play piece. A fork of Lingdong Huang's {Shan, Shui}* infinite landscape scroll, drawn in learned hands (Shen Zhou, Shitao, Buson, Rembrandt, Van Gogh and more). Same seed, same landscape, different hand. Scroll it →
- Measure something and share the fingerprint (images, text, robot scenes).
- Found a reading that's obviously wrong? That's the most useful report there is.
- Issues labeled
portandgood first issueare open doors.