LUT Inspector: Interrogate the Cube

Three probes tell you most of the story. The endpoints: where do (0,0,0) and (1,1,1) land? Lifted blacks say film emulation or a legal-range remap; crushed or clipped ends say a display transform or a damaged file. The neutral axis: run grays through and watch the three channels, a technical conversion keeps them locked together; separation is tint, deliberate (a look, a white-point move) or not (a bad LUT).

The mid-gray slope: the effective gamma at 18% gray separates "seasoning" (slope near 1) from "conversion" (slope far from 1, log decodes run steep, inverse transforms run flat). Those three probes are exactly the first three tiles in the metrics grid, and they are the same questions you would ask a display with a probe. Same discipline, different patient.

The classifier is honest about being a heuristic: it measures displacement (how far colors move on average), neutral tint, saturation ratio, hue rotation, endpoint behavior, and mid-gray slope, then matches the signature. Near-zero displacement reads as identity; tiny, neutral-safe moves read as calibration-scale correction; a strong slope change with locked neutrals reads as technical conversion; lifted blacks with a soft shoulder and compressed saturation read as print emulation; everything else with deliberate tint or saturation styling reads as a creative look.

Signatures can be forged (a look built on top of a conversion shows both) so the verdict lists its evidence instead of hiding it. Treat it the way you treat any single measurement: a strong prior, confirmed on the picture and the curves, never a substitute for knowing the LUT's intended input. When the verdict and the vendor's documentation disagree, believe the measurement, then ask why.

Synthetic stimuli find defects; pictures find intent. The built-in test frame packs a gray ramp (banding, tint, clipping), color bars (primary handling), skin patches (the tones clients actually complain about), and a sky gradient (where posterization hides). A LUT can pass all of it and still be the wrong look for a face, which is why the wipe exists: load a frame from the actual show, not a pretty stock photo.

One habit worth stealing from calibration: judge at the wipe line. The eye adapts within seconds to a full-frame grade and starts calling it neutral; a hard split denies it that adaptation, the same reason probe patches beat by-eye monitor matching. And remember what a chromaticity-only view would hide: a LUT's damage is often tonal, not chromatic; the ramp rows catch what the vectorscope forgives.

Clipping fraction: the share of the input volume the LUT pins to 0 or 1. Some is legitimate (display transforms clip out-of-gamut input); a lot means data destruction, anything downstream can never recover it. Neutral monotonicity: gray in, gray out should never get darker as input gets brighter; a reversal is a broken or truncated file and shows as solarized bands.

Resample safety: the inspector rebuilds the LUT at 17³ and measures what breaks. A smooth look survives; a steep conversion loses shadows. That number is practical, not academic, it is exactly what happens when a 33³ file is loaded into a device with a 17³ slot, which real hardware does silently. If the 17³ error is visible, that LUT must never ride in small-slot gear.

Classification determines placement. Technical conversion: belongs at defined chain positions (camera out, monitor in, deliverable render) and is never adjusted to taste. Creative look: rides between conversions, travels with the project, and must be paired with its intended input space. Calibration: lives in the last device before the glass, expires with panel drift, and never touches recorded media.

The physical version of this discipline (which box, which slot, which order, who is allowed to change it) is the next module: LUT Hardware and Workflows, from dedicated processors and DIT carts to live-event switchers applying cubes to LED walls in real time. If it isn't measured, it isn't calibrated.