Skip to main content

Earth Grid — evidence

Back to the short Grid page. This appendix is reference material, not an additional sidebar branch. Geographic calculations dated 2026-11-10; no geographic terrain GLBs were generated. The previous L7 proposal is retained as historical/reference inventory, not a selected initial plan. Current area-proportional options use different conventional area assumptions and are not geometric counts. The later synthetic compute probe is not terrain-generation throughput.

Mask policy​

Policy ne-map-units-v1-conservative-s2 is a provisional administrative planning mask, not an approved physical-continent or flight boundary. Use every one of the 298 map units from Natural Earth 5.1.1, at 1:10 million cartographic scale. “10m” means ten million, not ten-metre accuracy. The archive contains the source documentation and WGS84 projection metadata.[1][2]

  • Assign each complete map unit by its CONTINENT field. This deliberately puts all Russia in Europe, all Turkey/Kazakhstan/Indonesia in Asia, and all Egypt (including Sinai) in Africa. It does not split countries at physical continental divides. Source map units separate some overseas territories: French Guiana is South America, Greenland is North America. This classification must be reviewed before it becomes a flight mask.
  • Map source Oceania to the reader label Australia / Oceania. All mapped islands remain; unrepresented islands cannot contribute cells. Western New Guinea stays with Indonesia/Asia rather than being silently reassigned.
  • Explicitly assign the eleven Seven seas (open ocean) map units below. No feature is dropped and no nearest-continent inference is used.
Assigned maskMap units
AfricaSeychelles; Saint Helena; Réunion; Mauritius
AsiaMaldives; British Indian Ocean Territory
EuropeAzores
North AmericaClipperton Island
AntarcticaFrench Southern and Antarctic Lands; Heard Island and McDonald Islands; South Georgia and the Islands

Use source polygon footprints, including inland water represented inside those footprints; preserve explicit polygon holes. Antarctic coast/ice-shelf treatment follows the fixed source, not contemporary ice conditions. Include any cell with a conservative spherical intersection, even if most of the cell is water. No additional offshore buffer is counted. The earlier 25 km buffer is historical, separately uncosted analysis in the technical notes; water is proposed as procedural coverage, not another land GLB layer. This inventory alone is not a flight geofence, collision surface or guarantee of complete real-world land coverage.

The source is public domain according to the publisher-maintainer’s licence file; the immutable snapshot and SHA-256 are pinned in source-lock.json.[9]

The website’s terms/detail endpoints returned HTTP 406 during retrieval; the archive’s included README and maintainer licence were successfully retrieved instead.[1][3][7]

This finding concerns the map units only, not DEM, imagery or future derived terrain rights.

Spherical counting algorithm​

  1. Verify the source ZIP SHA-256 against source-lock.json; do not silently accept newer bytes at the same URL.
  2. Read each shape’s rings, remove duplicate closing/consecutive coordinates, map longitude/latitude to unit-sphere points, normalize the smaller side of each ring and initialize nested S2 polygon loops. Retain holes. Join vertices by S2 geodesic edges—not planar longitude chords. Reject invalid spherical polygons.
  3. The pinned Egypt source has one self-intersecting ring. Apply Shapely make_valid only to that named record, preserve every resulting polygon/ring and require the repaired S2 polygon to validate. Reject repair needs in any other record. No coordinate simplification is applied. Record the repair in the output.
  4. Traverse all six S2 faces, pruning cells whose polygon MayIntersect is false. Collapse a subtree only when Contains confirms the whole cell. Otherwise recurse to L11 and retain intersecting cells. The S2 predicate is conservative: numerical tolerance can retain an almost-disjoint boundary cell; the implementation documents an error under 15 nanometres.[6][8]
  5. Represent the retained L11 cells as merged inclusive integer ranges in fixed-level Hilbert order. Derive L6–L10 by mapping those cells to their ancestors; union overlapping/adjacent ranges. Count range lengths. These are enumerated covering counts, not area quotients.
  6. Union all seven continent sets for global assets. A boundary cell can belong to multiple continent catalogues, but its complete cell geometry/content must be globally identical and stored once. Selected-continent flight restrictions belong to a separate session mask, not a continent-clipped GLB. If future content differs by continent, this deduplication model must change.

S2 provides globally hierarchical cells, including poles and face seams; level does not imply equal area or constant side length.[4] The historical/reference L7 nominal 72 km label is approximately the square root of average cell area, not an exact edge length or terrain sampling resolution.[5] The preserved L11 coverage supplies the ancestors used for the preserved L7 count; publishing L11 assets is historical/future comparison, not part of the initial plan. Neither these averages nor continent areas are used to calculate counts.

Precision: integer counts are reproducible for the pinned polygons, repair, assignments and conservative predicate. Natural Earth generalization, omitted small islands, physical-border choices and future offshore coverage dominate practical uncertainty. Do not present these as survey-accurate final production quantities. No defensible percentage error against an approved high-resolution mask has been measured.

Reproduce​

From the repository root, in an isolated Python environment (the recorded run used Python 3.14):

python3 -m venv ../earth-grid-venv
../earth-grid-venv/bin/pip install -r research/earth-grid/requirements.txt
curl -fL https://naturalearth.s3.amazonaws.com/10m_cultural/ne_10m_admin_0_map_units.zip -o ../ne-grid.zip
../earth-grid-venv/bin/python scripts/earth_grid_count.py ../ne-grid.zip
../earth-grid-venv/bin/python scripts/test_earth_grid_count.py --source ../ne-grid.zip
node scripts/earth-grid-budget.mjs
node scripts/earth-grid-pages.mjs --check
pnpm test

The ZIP hash gate is enforced by the counter. Keep the downloaded ZIP and environment outside committed/publication directories. --out permits a second calculation directory for byte comparison without replacing the checked-in evidence. Requirements pin S2 0.14.0, PyShp 2.3.1, Shapely 2.1.2, NumPy 2.5.4. These are analysis-only tools, not application dependencies.

Repository artifactEvidence
research/earth-grid/source-lock.jsonSource URL/version/ZIP and member hashes; immutable licence URL/hash
research/earth-grid/counts.jsonEvery level, membership deduplication and all 298 assignments/repair flags
research/earth-grid/cover-intervals.jsonReconstructable L11 cell sets; hash recorded in counts
research/earth-grid/budget.jsonUnrounded byte counts and timing sensitivity results
scripts/earth_grid_count.pySpherical polygon coverage algorithm
scripts/test_earth_grid_count.pyGeometry fixtures and separate S2RegionCoverer cross-check
scripts/earth-grid-budget.mjsWorkload, storage and time formulas
scripts/earth-grid-pages.mjsEight short reader pages and linked area, L7-reference and benchmark appendices generated from canonical inputs
tests/earth-grid.test.mjsIndependent interval cardinality, budget and navigation regression checks

Geometry validation compares all 298 source records at L8 against S2’s separate S2RegionCoverer path, plus full L11 coverage for Lithuania, Fiji and Kiribati. Fixtures cover both poles, antimeridian, holes, a tiny island and exact cell edges. A second full counter run must reproduce the same counts and interval bytes. This cross-check uses the same S2 library; it does not independently validate the source geography or certify real-world coverage.

Production accounting​

The historical L7 scenario for one content version has one GLB per unique included L7 spatial cell, one mesh variant: 32,644 GLBs. There are 98,304 full-sphere L7 addresses, not 32,644 whole-Earth cells. No parent GLBs are included. The preserved L6–L11 counts describe the historical 9,777,734-asset alternative, not the initial plan. No L12+, city refinements, buildings, vegetation or extra mesh variants are counted. Additional full variants multiply GLB counts/bytes; a partial refinement requires its own counted cell set. A content change retains the stable cell ID but creates new versioned bytes.

Formulas used by the checked-in budget:

N = global unique cells at L7 = 32644
M = all seven continent memberships at L7 = 32901
GLB bytes = N × assumed bytes per GLB
metadata bytes = N × 512 + M × 32
one release = GLB bytes + metadata bytes
retained + staging = one release × (3 + 1)
capacity allowance = (retained + staging) × 1.20
worker-hours = N × seconds per GLB / 3600
calendar days = N × seconds per GLB / (86400 × workers × efficiency)

The 0.25/1/2 MiB scenarios are assumed encoded GLB file sizes, without assumed HTTP compression savings. They apply to the proposed L7 assets for a transparent sensitivity model; real files may be much larger. They are not achieved compression ratios or statistical confidence bounds. MiB = 2²⁰ bytes, GiB = 2³⁰, TiB = 2⁴⁰. Timing scenarios and byte budgets are independent: the table does not claim a 2 MiB model can be produced in one second.

Metadata means a 512-byte global per-asset allowance plus a 32-byte per-continent membership reference. It is a budget, not a measured serialized schema. Shard manifests for bounded downloads; never send the entire Earth inventory to a client. Object-store minimum billing size, indexes, erasure coding, replication, backup and egress charges need provider-specific costing. Three retained versions plus one full staging copy assumes no byte-level reuse. The 20% allowance is added capacity, not a claim that 20% of provisioned space stays free.

Still uncosted: source DEM/imagery and their local caches; intermediates and encoder scratch; collision/height sidecars; source and output backups; separate CDN copies; processing failed attempts; QA tooling and human review. Therefore the storage table validates output provisioning under stated assumptions, not all storage required to operate the eventual production pipeline. A full capacity commitment requires the source selection and measured scratch/sidecar sizes.

Honest generation-time boundary​

A bounded synthetic GLB benchmark measured preparation, export and format validation only. Real source-to-geographic-GLB generation remains unmeasured. The counter’s elapsed time is not a terrain-generation rate. The retained budget JSON’s 1/10/60-second and eight-worker scenarios are what-if arithmetic, not measurements or delivery estimates; they are intentionally absent from the short reader pages.

Before committing a schedule, authorize a representative pilot spanning coast, mountains, flat terrain, polar cells, face seams and islands. Record the exact CPU/RAM/storage, threads per worker, source resolution, geometry/texture profile and cold/warm cache. Measure acquisition/preprocessing, mesh generation, texture baking, encoding, seam/collision/content validation and publication separately, including failures and retries. Test concurrent workers against shared I/O and RAM. A worker using four encoder threads is not equivalent to one CPU core; eight workers cannot be assumed to scale eightfold. Publish stage distributions and asset-size distributions, then recalculate wall time and scratch capacity.

Human effort is not worker-hours: approving masks/data rights, designing seams/collision/materials, writing the pipeline, reviewing representative quality and resolving failed regions remain unestimated work. Current conclusion: the L7 land proposal is 32,644 unique assets; start with a separately authorized real-source pilot, not a mass-generation promise.

Sources​

[1] https://naturalearth.s3.amazonaws.com/10m_cultural/ne_10m_admin_0_map_units.zip [2] https://www.naturalearthdata.com/downloads/10m-cultural-vectors/10m-admin-0-details [3] https://www.naturalearthdata.com/about/terms-of-use [4] https://s2geometry.io/devguide/s2cell_hierarchy.html [5] https://s2geometry.io/resources/s2cell_statistics.html [6] https://github.com/google/s2geometry/tree/v0.14.0 [7] https://raw.githubusercontent.com/nvkelso/natural-earth-vector/master/LICENSE.md [8] https://raw.githubusercontent.com/google/s2geometry/v0.14.0/src/s2/s2shape_index_region.h [9] https://raw.githubusercontent.com/nvkelso/natural-earth-vector/c55565d81cca92368200e0c72b0fb832cdea5203/LICENSE.md — Natural Earth public-domain licence, immutable snapshot