Method

One statistic for how far the signal moves, one protocol for what that costs. The two decisions below changed the answers more than any modelling choice.

The appearance statistic

RMS pixel distance in uint8 between mean no-contact frames, 12 px inset, plus per-channel medians. Fixed up front so it cannot be tuned after seeing the answer.

The noise floor comes first — a distance between two means is meaningless until you know what two means of the same cell give. Disjoint halves of each cell's 100 repeats: 1.43 total / 0.28 low-frequency.

Decision 1 — across datasets, only the low-frequency part is comparable

Resizing to a common geometry and comparing total RMS is wrong, and was falsified before any result existed: a resampling round trip attenuates a white-noise-grade difference by ~40 % (7.88→4.71) while leaving a low-frequency one invariant to 1e-4. Harmless only if wear were pure illumination drift — it is 48 % low / 41 % high. So rms stays within one dataset; rms_lowfreq goes across. A regression test pins this.

The two axes are physically different: units differ in illumination (low-frequency, global), pads differ in marker layout and texture (high-frequency, local). Consequence: any cross-dataset comparison systematically under-represents the gel axis. gel_axis.json → low_frac

Decision 2 — trends are fitted inside continuous acquisition segments

The campaign ran over four days with breaks up to 88 h, and appearance jumps at exactly those boundaries — a trend fitted across one measures re-mounting. Records are split at acquisition gaps before any slope, and drift is scored against a permutation null, because the segment baseline contains frames that are themselves scored, making Spearman positive even under zero drift.

Scoring protocol

Rank correlation via a shared scorer, self-tested each run (monotone >0.9, shuffled <0.5). Transfer is always fit-on-source / apply-to-target explicitly, paired with a matched-fit-size in-domain control. Encoder comparisons project every feature set to a common width (PCA on the source only) and report a k = 32…256 sweep — without it the comparison ranks output width, not representation quality.

Retracted after independent review

Two adversarial review agents checked every number against the code and the raw data. Four headline claims did not survive. Kept here because a results page that only shows what worked is not evidence.

retracted claim	why it failed
The encoder ranking reflected representation quality	It ranked feature width. The scorer changes behaviour at p ≈ n_fit; holding the feature type fixed and varying only width swings transfer rho +0.04 → +0.91, and the old table was monotone in width.
“7/7 wear segments positive, p ≈ 0.008”	The null is not centred on zero — the baseline contains scored frames. A permutation null leaves 3 of 7 significant.
Gel axis = 16.35	FEATS gel index 5 is markerless, so a third of the pairs were a gel model change. Corrected to 13.03, with the model change reported separately at 27.19.
Shift-predicts-cost at p = 0.0033	12 directed pairs carry only 6 distinct shifts. Mantel permutation gives p=0.083.

Earlier retractions, and ideas rejected with numbers

Sparsh batch order. An earlier plan used sphere/batch_1..6 as a sequential wear axis. All 39 pickles carry mtimes inside one 2-minute download window; nothing on disk makes the index chronological. Usable only as an unordered grouping.

“25 h wear ≈ one gel swap.” An invalid cross-resolution comparison (640×480 vs 240×320, and wear is 41 % high-frequency). Withdrawn.

Background colour from random frames. Contaminated by contact: within one directory the descriptor moves 3.35 RMS from lowest- to highest-force frames, the same size as the “cluster separation” it produced. Replaced by the dataset's own 656 genuine no-contact frames, which cut the within-directory spread to 0.017.

Damage severity. Two criteria (frequency content and spatial concentration) disagree across the two runs, and no dataset here carries a damage annotation. No “damage” conclusion is drawn.

Data sources

source	n	availability	role
FEATS no-contact grid	3,000	public	5 units x 6 gels x 100 repeats, fully crossed and balanced
FEATS transfer splits	393 / 395	public	the dataset's own sensor+gel shift, force-labelled both sides
Gel durability campaign	464 refs	in-house, unreleased	25.0 h cumulative gripping on one Mini; finding 1 rests on it
cnc_mini_26	14,722	in-house, unreleased	markerless, force in filename
FeelAnyForce	13,589	public	markerless, 6-axis wrench
TacQuad (Mini only)	24,866	public	cross-domain; no force labels, contact windows only

Two of the six are lab-internal collections with no README, licence or release, so findings 1 and 4 cannot be reproduced from public sources as they stand. Every figure drawn from them names its directory.

Terms used on this site

out-of-distribution (OOD)	Input that differs from what the estimator was fitted on — here: a different sensor, a different pad, or a different dataset.
reference frame	An image of the gel with nothing pressing it. Reconstruction works on the difference between a frame and this reference, so the reference carries the sensor's own illumination.
markered / markerless gel	GelSight pads either carry a printed lattice of black dots (used to read shear) or do not. It changes the image far more than swapping one pad for another of the same kind.
rms / rms_lowfreq	Root-mean-square pixel difference between two mean images, in 0–255 units. `rms_lowfreq` keeps only the smooth part — the only version that survives comparing images captured at different resolutions.
noise floor	What the statistic reads when nothing has changed: two averages of the same sensor, same pad. Any effect smaller than this is not measurable.
Spearman rho	Rank correlation, −1…1. It asks whether predictions ORDER the true forces correctly, ignoring scale — so it survives datasets with different force units.
in-domain vs transfer	In-domain: fit and score on the same population. Transfer: fit on one, score on another. The gap between them is the cost of the shift.
permutation null	The score the same procedure produces on shuffled data. Without it, a positive correlation can look meaningful when the procedure would produce one anyway.
Mantel permutation	A significance test for distance-vs-distance comparisons, where observations share populations and so are not independent. It relabels the populations instead of the observations.
LUT vs calibration-free	Two ways to turn an image into surface depth. The LUT (lookup table) is calibrated per sensor from known presses; the calibration-free solve is not, which is why the two degrade differently when the sensor changes.
degenerate transfer	The fitted map sends every target sample to one value, so rank correlation is undefined — total failure, not a missing measurement.