Imaging and color measurement · Reports
Recovering and characterizing an archived spectroradiometer set
The experiment this would ideally have been
A controlled repeatability study fixes and records geometry, integration time, and instrument configuration; monitors the source independently or measures a stable reference channel; and interleaves repeat readings to expose warm-up and time-order effects. Those controls make competing source, setup, and instrument explanations testable instead of assuming that one caused the difference.
What survived is the measurement data itself: spectra, recorded XYZ, radiance, and acquisition fields, in files whose names counted acquisitions rather than describing scenes, several stored more than once under different names. The setup records did not survive.
That gap is the whole reason this report stops where it does. The stored differences can be quantified; the archive does not retain what is needed to assign them a physical cause.
Admitting the archive
Two operations, deliberately kept apart:
Content identity. Exact byte comparison separated 89 distinct retained readings from 45 byte-identical aliases. An alias is the same measurement under a second name; it is retained as provenance and analyzed once, because counting it twice would inflate a group’s apparent repeat count.
Grouping. Which distinct readings belong to the same target comes from the retained grouping record. It is not inferred from spectral similarity — that would assume the conclusion, since two readings of one stable source and two readings of two similar sources look alike. The record yields 40 groups: 37 with repeated readings, 3 singletons.
A ledger mismatch, an ambiguous source, a non-uniform grid, a shape mismatch, or a non-finite derived value is a rejection rather than a partial group.
Method
Each group is characterized on three axes, defined in the method companion: the equal-weight spectral integral for level, the relative L2 residual between self-normalized spectra for shape, and the largest pairwise Δu′v′ from recorded XYZ for chromaticity.
Repeat statistics use the sample standard deviation with n − 1. Singletons
report absent variation fields rather than zeros: one measurement establishes a
level and a shape but not a spread, and a zero would read as perfect agreement.
Results
Across the 37 repeated groups:
| Axis | Median | Maximum |
|---|---|---|
| Spectral-integral coefficient of variation | 7.167679309662159% | 41.647103399837285% |
| Maximum normalized-shape relative L2 residual | 0.5177845902558952% | 1.075914397243751% |
| Maximum pairwise Δu′v′ from recorded XYZ | 0.0007029769933166811 | 0.002851948638865613 |
ramp_patch_05 carries the level maximum. ramp_patch_01 carries both the
shape maximum and the chromaticity maximum. The level maximum therefore occurs
in a different group from the other two, and the three values do not describe
one measurement condition.
Same-record XYZ closure
Integrating each retained spectrum against a public CIE observer with equal
sample weights reproduces the XYZ recorded in the same file under one
archive-derived proportional scale of 683.0167582353332, with maximum absolute
relative residual 1.5543122344752192e-13% and RMS relative residual
5.004889505182855e-14%.
This is closure between two fields of the same file. It establishes numerical consistency between the spectral and colorimetric records under one fitted scale. It is not an instrument-accuracy test, and the fitted scale is derived from the archive — the analysis does not identify it as a standard luminous-efficacy constant or infer undocumented instrument-software behavior from its magnitude.
Independent read
A MATLAB R2026a implementation read all 89 retained readings: 89 source-file
identity comparisons, 178 exact binary64 vector-hash comparisons — two per
reading — and 623 numeric-field comparisons at 1e-12
absolute-or-relative tolerance. The largest absolute difference was
4.547473508864641e-12 K, on recorded CCT.
That establishes agreement between the two readers on the retained vectors and numeric fields. It does not independently reproduce the grouping record or the derived group statistics, and it is not a test of instrument accuracy.
Limitations
The retained records do not establish whether the observed differences represent physical change, acquisition variation, or measurement uncertainty. Source output, geometry, acquisition settings, re-aiming, and instrument behavior are not separable, so the result is labeled within-group observed variation — not source drift, instrument noise, or repeatability.
Recorded CCT and Duv pass through as metadata. The source files do not identify the locus and distance conventions needed to recompute them unambiguously, so they are carried rather than recalculated.
Three singletons remain in the output with empty variation fields.
What would resolve it
A repeat acquisition should record the conditions this one lost: documented geometry, integration time, and instrument configuration; an independently monitored source or stable reference channel; and interleaved repeats. Those controls would make source, setup, and instrument explanations testable. Until then, 41.6% is a quantified within-group difference, not an attributed drift or noise mechanism.
Source file: reports/spectroradiometer-recovery.md