Which Earth is Mars — and what it proves.

A different reference gives a different answer — which is the point

The Moon is the other half of the challenge, so it was run on real NASA Goddard LOLA 5 m/pixel terrain: all six published south-polar candidate regions, screened against the same Earth roster on an exact common grid — 500, 750 and 1000 m cells covering 9 km on every unit, on both bodies. Every site is rough terrain. Across 18 of 18 site × support cells the sole non-dominated Earth candidate was the same: the Nevada Basin & Range. The two smooth non-volcanic terrains that answered for Martian volcanic plains were the worst matches here.

Three things stop that reading as a finding. The candidate is not always close: at Shackleton rim even the survivor is 70% away on its worst axis, so it is the least-dissimilar of a weak shortlist — the exact failure mode this project withdrew a claim over on the Mars side. The three declared axes are not three axes: mean slope and 3×3 relief are the same gradient measured two ways — relief = 2 d·tan(slope)·ρ, with ρ provably confined to [1, √2] and depending only on the gradient's direction — checked against the engine's own functions to 4.4e-16 m and confirmed on the panel at effective rank 1 of 3. And an earlier version of this comparison did not put both bodies on the same ground at all; that is corrected and rerun, the ordering held, and the distances moved.

So the conclusion was then attacked rather than defended: 30 counted perturbations — window origin, support jitter, footprints from 3 to 10.5 km, both alternative resampling kernels, edge and extreme-cell masking, DEM noise, tilt and smoothing, descriptor dropout, candidate dropout, one added candidate — declared in a config committed before the run, with the pass rule fixed in advance. It survived 28 of the 30 perturbations: only 18 of 504 site × support cells moved, a cell-level stable fraction of 0.964, verdict ROBUST. The two that reverse are the falsifiable part and are named: a 3 km window changes the survivor, and adding an eighth candidate makes Nevada a co-winner rather than the sole one.

Then the two questions that a ranking always gets mistaken for were answered separately, as first-class outputs. Is any candidate inside the reference's own morphological variability at this scale? — not established, at every cell, because the yardstick fails: where both descriptor bands can be calibrated — 6 of the 18 cells, the six at 500 m — the reference's own self-distance band comes out so wide that it admits at least 4 of 5 other lunar sites, and a band that says adequate to terrain from demonstrably different places has established nothing. At the other 12 there is no relief band at all — the held-out sub-windows do not supply one at 750 or 1000 m — so the verdict there rests on the slope band alone, which also admits at least four of the five. The verdict is the same either way, and it rests on a weaker test at those twelve; the second band is missing rather than wide, and that is stated rather than folded into the sentence above.

Is the evidence enough? — marginal, with the reasons named. And adding those sibling sites as controls produced the least comfortable number on this page: at 18 of 24 site × descriptor statistics the unit nearest a lunar reference is another lunar site, not any of the seven Earth candidates. The ranking is definite at all 18 cells and it is an ordering within the Earth roster — it is not evidence that the roster contains the Moon.

It stays exploratory, not confirmatory: six nearby sites are not six independent tests, and the error-ensemble check is 6 of the 100 published realisations. No base suitability of any kind is claimed — morphology only.