# What the scores mean

The recipe catalog reports measurements of the **actual table shipped in each profile**, decoded from its XMP. RGBTable evaluation uses native tetrahedral interpolation; a trilinear proxy is not used.

## Fitting error

`fit.mean`, `median` and `p95` are CIEDE2000 between predicted and NX Studio chart samples. Both are encoded Adobe RGB (1998), converted to D65 Lab; kL=kC=kH=1. Samples with any input or target channel below 0.0005 or above 0.999 are excluded. `sample_count` records the number retained.

The primary score is a **current-table chart evaluation**, not independent unseen validation and not a full-photograph score. Where a YC7 fit retained a separate chart test split, `current_fit_holdout` contains that historical held-out result. Random chart holdout is not equivalent to a new real-world scene.

Three recipes (LeicaSTD, At Dusk and Blue Hour) use the verified detail-off NX chart. The other historical chart exports did not consistently disable detail settings. This is recorded per recipe; treat scores from different measurement policies cautiously. Patch interiors reduce, but do not prove elimination of, local-rendering effects.

CIEDE2000 describes local color differences. It does not model all spatial perception, visual attention or whether a difference looks pleasing. A small neutral tint across a large sky can matter more than a higher local ΔE in texture. Do not choose profiles solely by one average.

## Banding risk

`face_jump_p99`, `face_jump_max` and `face_jump_rms` are the 99th percentile, maximum and RMS of the Frobenius norm of the output-RGB Jacobian change at **every unique internal tetrahedral face and neighboring-cell face**. RGB coordinates are normalized 0–1; slopes are per unit input RGB. A 32³ table has 351,726 evaluated faces.

These are structural scores, not a count of visible bands and not ΔE. Tetrahedral interpolation is continuous, but its derivative is piecewise constant. A large change in slope can reveal contours in smooth skies or bokeh even without a discontinuity in RGB values. Strong legitimate tonal changes can also increase this score; there is no calibrated universal visibility threshold. 8-bit preview quantization, display scaling and JPEG compression introduce additional limitations.

The YC7 fitter directly penalizes mixed changes inside cells and changes across cell faces. More lattice samples alone do not guarantee smoothness. **Every recipe is audited globally**, independently of the four photographed color paths.

## Photo previews

Four sanitized Z8 samples cover bokeh, sky and two landscapes. The default gallery applies the decoded table to area-averaged 16-bit Camera Raw Adobe Standard exports, then converts Adobe RGB to sRGB for display. They are **numerical previews**, not full Lightroom exports; averaging before a nonlinear transform can alter fine-detail colors. Native-resolution crop previews are calculated before resizing. Source documentation uses lightweight thumbnails; the separate gallery Release ZIP retains the original overview and native crop images. Resizing can conceal or change visible banding, so use the full gallery or RAWs for that evaluation.

Selected previews are also rendered through actual Camera Raw and compared separately. Their validation report records the sampling and errors; it is not generalized to every profile or Amount. There are no NX exports of these four scenes in this release, so the gallery must not be read as an NX fidelity comparison.

All settings and provenance are recorded in `gallery_manifest.json`. Use the RAW sample asset and actual Lightroom for a final visual check, especially at 200% Amount. Owner review labels apply to the earlier YC7 evaluation and are not certification of every new sample.