Methods and limits

GSOP scores every land tile on Earth by how closely its terrain matches a lunar or Martian target, then ranks the tiles.
It is a screening method. It says where to look, not where to go.
The Moon score is effectively terrain only. Mars is a low-confidence secondary result.
The method failed its own pre-registered negative-control test. The Failure-First Ledger lists every failure first, and so does this page.
The ranking has not been checked against independent data.

The question

The challenge asks for Earth places that could serve as analogs for candidate Moon-base locations or Mars landing and exploration sites. It also asks what makes a place analogous, where the analogy breaks down, and how studying it could help.
GSOP answers one narrow part: which Earth places have terrain like a lunar or Martian target, using open data. It does not say what to test at a place.
An analog is useful only for the properties that match. The properties that do not match are on the ledger and in the limits below.

Limits

What the score cannot see

The Moon score is effectively terrain only. The illumination family was given a weight of 0.2513, but the highest lighting match on any Earth tile is 0.0009, so it changes nothing.
Gravity, air pressure, dust, radiation and regolith are not scored. Neither are temperature, rock composition and thermal inertia for the Moon, or water, dust and temperature for Mars.
An Earth place can match a target's terrain and still differ in everything else that matters for a mission. The ledger lists those differences, most of them provisional.

Data limits

Only tiles that passed the land-cover and data gates have a score. Water and ice, vegetated tiles, built or farmed tiles and tiles without data are left out, so large regions never appear.
Terrain data exist for a share of 0.6444 of land tiles. Tiles without terrain data are never scored. The coverage report names tiles beyond the reach of the main elevation dataset as part of the gap but does not explain all of it.
Earth slope is not corrected for the map projection, so it reads too steep toward the equator and too gentle toward the poles. A latitude band check was run to see whether the ranking holds.
Terrain data change source at the northern edge of the main elevation dataset, and roughness steps up across that seam.
The coordinates of 11 of the sites v1 used are not confirmed from a primary source.

Result limits

Two negative controls scored high on the Moon run: the Namib sand sea at percentile 90.2 and the Rub' al Khali at percentile 75.6, against a pre-registered maximum of percentile 25.0.
On the Mars run the median percentile of the primary sites was 61.3, below the pre-registered minimum of percentile 90.0. The Moon primary-site criterion could not be tested.
The score bands come from varying the weights and not the Earth data, so they understate the real uncertainty. The uncertainty section below prints the full statement.
Everything from v1 is provisional.

Site and region limits

Reference sites come from a teammate-supplied list. Their text is teammate-written and unreviewed, and no source has been checked yet.
Regions and tiles are candidates not in our reference list. They are unreviewed.
Region names are labels from Natural Earth, which is generalised: the desert, plateau or range the centre lies in, one whose outline is within 200 km, or else the nearest populated place. A name is not a claim about the ground.

How the screening works

Tiles. Land is split into equal-area tiles of 36 square kilometres each.
Gates. Tiles that are water or ice, vegetated, built or farmed, or that lack data are removed.
Targets. The Moon targets are Connecting Ridge C1 and the Shackleton rim at the south pole. The Mars target is Jezero.
Features. The Moon score uses slope, illumination, roughness and relief. The Mars score uses sunlight, slope, relief and air density.
Score. Each tile gets a similarity to the target, where higher means more similar. Three purpose profiles weight the features differently: base, landing and mobility, and power and night.
Monte Carlo. 2000 runs vary the weights. For the Moon they also vary the slope and relief measurement error.
Flag. A tile is flagged when it is in the top 5% of scores in at least 90% of runs. This measures stability to the weights, not whether the ranking is right.
Sites. A site gets the score of the best-scoring scored tile within 20 km. Its percentile is the share of scored tiles in that view with a lower score. This is a lookup of stored values, not a test.
Regions. Neighbouring flagged tiles are joined. Groups under 100 tiles and groups within 100 km of a reference site are dropped. Each target lists its top 10 regions by median score.
This site. Pages read stored results. Nothing here recomputes a score, and every number traces to a source file and hash.

Checks fixed before the run

Before results were computed, v1 fixed three criteria: the primary sites must rank high, the negative controls must rank low, and obviously unsuitable ground must be rejected by the gates.
Pre-registration hashes are on the ledger. A hash identifies the plan. It does not mean the plan was met.
Result: the controls failed, the Moon primary criterion could not be tested, the Mars primary criterion failed, and the gate tests passed. See the ledger for the numbers and what each means for the analogy.

What each score uses

Features used and not used, by body
Body	Used in the score	Not scored
Moon	Slope, illumination (no effect), roughness, relief	Water or ice, composition, temperature, thermal inertia
Mars (low confidence)	Sunlight, slope, relief, air density	Roughness, composition, water, dust, temperature, thermal inertia

The v2 limits

v2 run not completed

The reference thermal layers give context for the Moon and Mars targets. This is where each layer cannot match Earth, as written with the data.

Reference thermal layers: where each one cannot match Earth

Limit that applies to every thermal layer

Earth has no surface near the Moon's polar cold-trap temperatures. In this file prp_temp_max at shackleton_rim has a minimum of 76.75 K and prp_temp_avg a minimum of 49.584 K (primary radius, 5 km). Earth's surface cannot reach those values, so an absolute match is impossible. Only the within-body rank can be compared.
The Moon has no atmosphere, so its surface temperature follows sunlight, slope and shadow. Earth's surface temperature also follows air, humidity, cloud, vegetation and water. Two places with the same rank can be warm or cold for different reasons.

Per layer

prp_temp_avg (Moon, PRP `temp_avg`)

What makes it analogous: annual mean temperature is a long-term thermal budget, the same kind of quantity as the Earth annual mean of (day+night)/2.
Where it breaks down:
- PRP temp_avg is a model value at 2 cm depth (product label), not a measured skin temperature. Earth (day+night)/2 from land-surface-temperature products is a skin temperature at a satellite overpass time. The depth and the definition differ.
- The PRP is a thermal model fit on a triangle mesh of about 500 m edges (about 613 triangles inside 5 km at connecting_ridge_c1, row prp_triangles, n). Sub-mesh variation is not represented.
- Earth skin temperature has cloud-gap and overpass-time bias; the Moon value has none of these.

prp_temp_max (Moon, PRP `temp_max`)

What makes it analogous: peak heating on a given surface, matched to the Earth annual maximum of monthly day LST.
Where it breaks down:
- PRP temp_max is the modelled annual maximum at the surface. The Earth partner is a maximum of monthly satellite day values, which samples a few clear-sky overpasses and can miss the true peak.
- On the Moon the maximum depends on grazing sun and horizon shadowing at the poles. On Earth it depends on latitude, albedo, moisture and cloud. The two are not driven by the same terms.

prp_cold_trap (Moon, share of triangles with `temp_max` below 110 K)

What makes it analogous: it measures how much of a site stays cold enough to hold ice, which links to the water and ice layer of the analog argument.
Where it breaks down:
- 110 K is the pre-registered threshold (runs/v2/PREREG_B.md). What the sources say, checked on 2026-11-15:
  - Hayne et al., "Micro cold traps on the Moon" (arXiv:2005.05369, HTML v1, read to about 60% of its length; the PDF did not render): Section 1 gives "<110 K" as cold enough for ice to last longer than the age of the solar system. Section 2, Eq. 1 uses the maximum temperature of permanent shadows below 110 K. The Figure 3 caption calls 110 K the adopted sublimation temperature for water ice. Section 3 says Williams et al. (2019) used a Diviner threshold of 110 K. Appendix C was not read.
  - Diviner team post (diviner.ucla.edu, "New Diviner Seasonal Polar Data Products", 2020-07-20): surfaces below 110 K can cold-trap water over a billion years; the map caption refers to maximum summer and winter temperatures. It names Williams et al. 2019, JGR 124:2505-2521.
  - Paige et al. 2010, Science 330:479-482: not opened (publisher and USGS pages returned HTTP 403). Whether it uses 110 K is not known from this run.
  - Williams et al. 2019: not opened.
  - PRP label dlre_prp_south.xml and the PRP collection label: no statement about 110 K or cold traps. The collection label cites Paige et al. 2010 as its reference.
  - Fit to PRP: the sources apply 110 K to a maximum temperature of shadowed terrain. PRP temp_max is a modelled annual maximum at the surface. The definitions are close but not identical.
  - No Earth surface sits below 110 K at its annual maximum, so the Earth partner for this layer is "not covered". It is a Moon-only context number, not a scored pair.
- Cold traps smaller than the 500 m mesh are not resolved, so the fraction is a lower bound at this scale.
- shackleton_rim shows 88 of 612 triangles below the threshold at 5 km; connecting_ridge_c1 shows 0 of 613. Both statements are specific to the 5 km radius and move with radius (rows with suffix _R2p5km and _R10km). The two sensitivity windows at 10 km overlap each other.

prp_ice_depth (Moon, PRP `ice_depth`, reference context only, not scored)

What makes it analogous: it answers the same question as Earth permafrost depth (how far below the surface ice is stable).
Where it breaks down:
- The Moon value is the depth where ice would sublimate at 1 kg m-2 per billion years (label). Earth permafrost depth is set by ground temperature and heat flow over millennia, with liquid water and salts that the Moon does not have.
- -999 means deeper than 2.8738 m. In this file connecting_ridge_c1 has 421 of 613 triangles at -999 and shackleton_rim has 280 of 612. These are counted separately and never averaged.
- The file also holds -1.0 in 100,319 of 2,880,000 records, which the label does not define. Supplemental rows report it separately (see runs/v2/PREREG_B_DEVIATIONS.md, item 11). Cross-count at shackleton_rim, 5 km radius (rows prp_ice_depth_x_cold_trap): 87 triangles hold -1, 88 have temp_max below 110 K, 87 are in both, 0 hold -1 without being below 110 K, 1 is below 110 K without -1. At 10 km: 563 hold -1, 580 are below 110 K, 563 are in both, 17 are below 110 K without -1. At 2.5 km neither group has a triangle. These are counts only; the label and the PRP collection label do not say what -1 means, so it is not interpreted here.
- The product shows where ice could be stable. It does not show that ice is present.

rcp_diurnal_range and pcp_diurnal_range (Moon): not covered

No v1 reference site is non-polar, so the RCP tile (Lacus Mortis only) was not used.
The PCP polar diurnal files are about 250 MB per local-time file and 96 per season. They support byte-range reads (HTTP 206 in the step 2 probe) but were not downloaded, because the pre-registration fixed diurnal range as "not covered" for both polar sites.
Consequence: the Moon diurnal range to Earth day-night range pair has no Moon values in this run.

tes_ti_day and tes_ti_night (Mars, thermal inertia)

What makes it analogous: thermal inertia is a property of the near-surface material (grain size, rock, cementation). Earth day-night temperature range responds to the same property, inverted: high inertia gives a small range.
Where it breaks down:
- The Earth partner is an inverted proxy. Earth day-night range is also driven by air humidity, cloud, vegetation and soil moisture. A damp soil and a high-inertia rock can have the same range.
- The TES maps are 20 pix/deg, 2.96 km per pixel (label). About 38 pixels fall inside the 10 km radius at jezero. A 3 km pixel averages terrain that an Earth 1 km pixel would separate.
- The labels give no scale factor and no fill value; the whole-map range is 0 to 4999 (deviations file, items 5 and 9). The values at jezero are well inside that range.
- Interpolated pixels are flagged by the mask files. At jezero the interpolated share inside the radius is 0 for both maps (row frac_interpolated).
- Mars atmosphere is thin and dusty, so the atmosphere does not damp the surface range the way Earth air does. The day-night range on Earth would partly measure the air, which the Mars thermal inertia does not.

mars surface temperature: not covered

No surface-temperature product was pulled for Mars. The TES TIMAP bundle holds thermal inertia only.

arcadia (Mars site): not covered

The v1 target file has no centre for Arcadia Planitia (docs/DECISIONS.md D10). No coordinates were guessed, so no layer is extracted for it.

lunar terrain: not covered here

LOLA terrain was skipped on purpose; the v1 run already carries Moon terrain.

Reading the CSV

stat = not_covered rows have an empty value. The reason for each is in this file.
Sensitivity rows carry the suffix _R2p5km or _R10km on stat.
The sha256 column holds the checksum of the source file the number came from. Mask-file checksums are in data/reference/thermal_refs_files.csv.

Site evidence limits: not available yet. The evidence run has not produced them.

Data resource table: not available yet.

Uncertainty

The Monte Carlo varies the weights and, for the Moon only, the slope and relief measurement error, and applies no systematic Earth bias to any of the 15 Earth columns, so the reported score bands and robust-tile counts understate the real uncertainty.

Data and reproducibility

One build command turns the stored results into the data files this site reads. A second run gives identical files.
Every number on a page has a row in a claims file that names its source file and hash. A check on every change fails on banned wording, an unsourced number or an altered uncertainty sentence.
The open data behind the scores are on the data credits page.
Anything v1 did not store reads "not in v1 outputs". Nothing is estimated.

Glossary

Tile: One equal-area cell of the Earth grid. Each gets at most one score per target.
Gate: A rule that removes a tile before scoring: water or ice, vegetation, built or farmed land, or missing data.
Percentile: The share of scored tiles in a view with a lower score than this one.
Flag: A tile that stays in the top group across most Monte Carlo runs. It measures stability, not correctness.
Negative control: A place that should score low. A high score means the ranking cannot tell it from a good match.
Pre-registration: A written plan of the criteria, fixed before results were computed.
Reference site: A place named as an analog in our list. It is not a scored result.
Candidate: A tile, region or suggested place that is not in our reference list. Unreviewed.
Provisional: Not yet checked by us. Most ledger entries are provisional.