Earth Grid — historical multi-level analysis
Everything below is preserved historical L6–L11 / 4.5 km analysis, not the current initial plan. The current Earth Grid compares 87 / 439 / 2,200 area-proportional GLB budgets; Option A / 1 is selected, with a separately computed proposed square cover. The later L7, nominal 72 km, 32,644 unique land GLBs proposal is also historical/reference inventory, not a selected starting plan. No multi-level pyramid, 25 km buffer, streaming preset, memory target or startup pass below is approved or carried forward automatically. The old 9,777,734-GLB scenario was over-sized for the initial inventory. Later refinement requires separate evidence.
Back to Grid. This appendix retains the detailed streaming/hardware proposal and its source citations. The counting evidence now supplies a reproducible administrative planning inventory; it does not approve the physical flight boundary. The optional 25 km offshore buffer below is not included in those counts or storage totals. Only one mesh variant per spatial level is budgeted initially. Geographic-count tests are not flight/terrain-performance tests.
Recommendation: use one stable, planet-wide cell address system, but load and permit flight only inside the selected continent's bounded session area. Start with an S2 cube-sphere quadtree, roughly 4.5 km nominal near cells, coarser parents, and one independently addressable GLB per cell per LOD. Prefer broad, modest terrain fidelity over isolated high-detail islands. Improve content without changing cell identities.
Status — proposed engineering strategy, researched 2026-11-10. This explanation is for Architecture, UI/UX and Delivery to evaluate the next design and measurement work. The user authorized continent-limited flight and strategy publication, not continent generation, runtime implementation, benchmark acceptance or ADR approval. No Earth tiles or models are generated by this change. Existing Aethera assets and Earth → Europe → Lithuania → Cities → Vilnius remain unchanged.
One continent is the flight area, not the whole globe
Earth's globe remains a continent-selection surface. COBE is not the terrain renderer. Global addresses enable stable storage, adjacent cells and later reuse; they do not authorize continuous cross-globe travel, automatic continent switching or loading the whole planet.
Proposed session contract:
- Select one versioned continent mask before launch. Its manifest lists included cells, exclusions, allowed spawn points, land/water classification and boundary policy. A shared cell may appear in more than one continent catalogue; it still has one identity and content hash.
- Treat Europe/Asia and other editorial borders as explicit polygons, not coastlines or latitude bands. Review disputed boundaries, islands and Oceania membership before producing content. Selector dots are not boundary data.
- Include a proposed 25 km offshore buffer around included coasts, clipped at the selected session boundary. This is bounded local coastal flying, not ocean crossings to another continent. Render low-detail water beyond the permitted area for a horizon, without enabling flight there. Open ocean needs no land GLB; its manifest entry must explicitly say water/empty, not look like a failed download.
- Give a boundary warning at the larger of 10 km or 20 seconds of predicted ground travel. Before leaving supported coverage, offer turn-back or pause/reposition inside the area; do not silently teleport, wrap coordinates or bounce the aircraft. Exact interaction and boundary safety margin require design approval and a measured turn envelope. A 25 km buffer is not proven sufficient for every aircraft manoeuvre.
- Pin a session to one manifest/content release. Changing continent ends/restarts setup. Missing, offline or rejected tiles keep known coarse coverage; if safe terrain/collision coverage runs out, pause with recovery choices rather than fly into a void.
These are proposed policies, not present application behaviour.
What the existing flight code actually establishes
Read-only inspection used application commit 16144cc74cac4008e72dcb4aa4b4cad14d74df2a on 2026-11-10. This is source evidence, not a runtime benchmark or deployment assertion.
| Observed source contract | Consequence for this strategy |
|---|---|
| Profiles specify trainer 310, airliner 510, fighter 1320; production physics converts profile speed from knots to metres/second.[15][16] | Size the first analytical envelope for the fighter, not a slow trainer demonstration. |
| The force evaluator limits thrust relative to drag at maximum airspeed; it does not clamp total world-space velocity.[17] | These are nominal airspeed limits, not a proven maximum ground speed in wind, descent or unusual states. Add a 900 m/s stress case and measure the real envelope before approval. |
| FlightCanvas imports React Three Fiber; render policy clamps DPR to 1–1.5.[18][19] | Remain on the existing Three.js/Fiber path. Proposed DPR 1 baseline below is a future quality preset, not a claim that today's code implements it. |
| World-space camera defaults are 60,000 m, but package construction overrides the far plane to 71,000 m for Training Airfield/Mountain Valley and 100,000 m for Coastal Range.[20] | A distant/coarse coverage budget is necessary; a small detailed patch alone cannot fill the visible horizon. These existing planar/package contracts need a separate georeferencing design. |
| Mountain Valley manifest allows 6 downloads, 100 spatial chunks, 256,000,000 decoded/GPU bytes.[21] | Existing bounded residency is useful precedent, not an Earth allocation. |
| Training Airfield manifest allows 45 total resident GLBs, 1,610,612,736 estimated GPU bytes, KTX2 ETC1S.[22] | Existing packages have different envelopes. Do not inherit the larger budget as safe on an 8 GB shared-memory laptop. |
The actual Earth ground-speed cap, altitude/visibility envelope, camera FOV, collision residency, origin-rebasing behaviour and browser/device timings remain unmeasured for this design. A source constant is not proof that flight meets it under every condition.
Why this initial grid, rather than equal degree squares
| Candidate | Advantages | Costs and edge cases | Initial disposition |
|---|---|---|---|
| Geographic latitude/longitude quadtree | Familiar data indexing; straightforward bounds and raster ingestion | A degree of longitude does not represent constant ground distance; cells narrow toward poles. Longitude wraps at the antimeridian. Web Mercator-style layouts exclude polar coverage and cannot be an Antarctica solution. | Useful source indexing, not the chosen flight-cell identity. |
| Cube-sphere quadtree, concretely S2 | Six faces, hierarchical four-child cells, globally stable IDs and defined neighbour operations, including polar regions.[3] | Cells are not equal-area or equal-kilometre squares. Face seams and orientation require explicit neighbour tests; S2's spherical indexing is not itself an ellipsoidal height datum.[3][4] | Initial candidate. Reuse a specified mature indexing convention, rather than inventing a second cube projection. |
| Metric regional grid, e.g. 4 km squares per projected zone | Predictable local dimensions; convenient mesh/raster tooling | One projection cannot retain uniform metric behaviour over every continent. Multiple zones/polar projections require zone IDs, overlap ownership and seam transforms. A single continent-scale tangent plane is not the globe. | Strong alternative for a limited regional pilot; rejected as the first global identity because zone stitching becomes a second addressing problem. |
The first and third rows are engineering tradeoff analysis, not a measured performance ranking. S2 is selected provisionally, not as an accepted dependency or framework migration.
Concrete cell and hierarchy proposal
S2's published statistics give L11 average area 20.27 km², with approximate area range 12.18–25.51 km². Example L11 edges are 3–5 km; those examples are not global edge bounds.[4] The nominal dimension below is the square root of average area, not an exact side length.
| Address level | Nominal dimension | Proposed role |
|---|---|---|
| L6 | 144 km | Coarse startup/horizon coverage, with intervening ancestors as necessary |
| L8 | 36 km | Distant terrain |
| L10 | 9 km | Mid-distance and low-bandwidth fallback |
| L11 | 4.5 km | Initial near-flight cells |
| L12 and finer | Not initially generated | Later airports/cities only after evidence justifies refinement |
Publish renderable coverage at every intermediate level L6–L11; the table highlights milestone levels, not permission to skip L7/L9 payloads. Each replacement then has four immediate children, so a four-child swap reservation is meaningful. Before generation, compute each selected cell's actual corners, surface area, minimum width, neighbour orientation and geodesic bounds. Use 2 km effective forward advance as the conservative planning scenario below, not as a claimed S2 minimum. Corner grazing can cross a cell almost instantly: real scheduling must cover the swept corridor and count intersections, not assume a minimum crossing time. Test cube vertices, polar cells and antimeridian neighbours before accepting the grid.
One GLB per cell per LOD, not a continent-sized file
Proposed identity: earth/s2-v1/<cell-token>. Store tokens as strings, not JavaScript numeric IDs. A content URI adds an immutable release/hash and an LOD variant, for example earth/<release>/<cell-token>/lod-0.glb. s2-v1 fixes the addressing convention; changing terrain data, mesh density or texture does not renumber cells. Keep grid version, source-data version, continent-mask version and content hash separate.
Each nonempty renderable cell has one GLB for each published LOD variant, containing that cell's terrain mesh/materials, not many independently fetched fragments and not every LOD inside one huge GLB. A coarser parent is itself a cell with its own GLB. Spatial level (L6–L11) and within-cell mesh LOD (for example lod-0/lod-1) are separate fields. Initial delivery may publish only one mesh variant per spatial level; manifest selection must never request nonexistent variants.
The proposed manifest needs: ID/parent/children/neighbours; geodetic bounds and ECEF origin; local frame and height datum; geometric error; compressed bytes, decoded CPU bytes and estimated GPU bytes; triangles/primitives/materials; content hash; land/water/empty/unsupported state; source and licence references; available LODs; collision/height-sampling reference. Preserve a cheap height representation independently of visual LOD so unloading detail never changes collision height. No flight-safe collision dataset is established here.
Proposed loader trust boundary: accept only authenticated, schema-validated manifests from permitted origins; validate cell tokens and release paths, reject unexpected external GLB resource references, and enforce compressed and decoded allocation limits before attachment. A hash from an untrusted manifest does not authenticate content.
Precision and edges are content contracts
glTF specifies metres and a right-handed, Y-up coordinate system; position accessors use floating-point data rather than a geographic coordinate reference system.[1][14] Proposed handling:
- Keep authoritative geodetic/ECEF positions in double precision outside GLBs. Convert geographic latitude/longitude to the S2 spherical index only for addressing; derive actual vertex positions on the chosen WGS84 ellipsoid with an explicitly converted height datum.
- Put each GLB's vertices near its own origin, in metres. Define local +X east, +Y up and −Z north, with a manifest transform to ECEF. Never put Earth-radius ECEF magnitudes directly in float vertex positions.
- Rebase render origin near the aircraft (initial trigger proposal: 2 km), atomically across camera, terrain, aircraft, effects and physics-facing transforms. Keep simulation/world identity unchanged. This is not permission to move only the camera while collision stays in the old frame.
- Sample neighbouring edges from identical source samples and a common elevation datum. Canonical edge keys and ordering must work across cube faces. Derive normals with shared border samples, add texture gutters, and compare height and normal continuity in generated validation.
- Restrict adjacent visible refinement to one level where feasible. Use matched edge tessellation/morphing and bounded skirts as visual insurance; skirts do not repair inconsistent terrain or collision data. Avoid duplicate water surfaces and z-fighting at coasts.
Suggested future tolerances: shared-edge positions within 1 cm after transform and deterministic shared height samples; normal discontinuity assessed with a grazing-light review. These are proposed engineering checks, not current asset results. Source DEM accuracy is a separate limitation and cannot be improved by a stricter seam tolerance.
Keep coverage while replacing detail
Use projected geometric error to prioritize refinement, borrowing the principle, not mandating Cesium or 3D Tiles adoption. Cesium's selection documentation distinguishes loaded tiles from renderable tiles and explicitly addresses detail disappearing during parent/child replacement.[2] 3D Tiles separates bounding volumes, geometric error and content references.[5]
Proposed initial SSE thresholds: refine above 8 pixels, coarsen below 6, then tune on devices. Use drawing-buffer height and camera FOV, not CSS size. Keep the parent visible until all required replacement children have fetched, decoded, prepared textures and become GPU-ready. Commit the coverage change atomically; retain children until a coarser replacement is ready when moving away. Include overlap in memory accounting. The initial all-required-children policy deliberately trades extra loading for easier no-hole correctness; it is not a claim to reproduce Cesium's optimized selection algorithm.
Before spawning, require low-detail coverage for the entire permitted session area plus its visual boundary, or a separately reviewed bounded launch subarea. Proposed startup aggregate: at most 8 MiB of coarse GLBs/manifest plus water, and 32 MiB coarse decoded/GPU allocation. These caps require future content validation; if a continent cannot meet them, revise coarse fidelity or launch-area policy rather than claim complete coverage. At 10 Mbps with 70% usable throughput, 8 MiB takes 9.59 seconds for transfer alone; decode/upload adds time. Show honest loading progress and do not begin flight with incomplete required coverage.
Quantitative flight and loading envelope
Count-based startup finding: the Grid inventory has 703–2,157 L6 cells per continent. At the storage model’s uniform 1 MiB/GLB, even coarse startup alone is 703–2,157 MiB, so the 8 MiB candidate cap above fails. Meeting it requires much smaller coarse payloads or a separately bounded launch area; neither is demonstrated. The global storage sensitivity is not a validated streaming preset.
All numbers in this section are tool-computed planning scenarios, not measurements. Reproduce with node scripts/earth-flight-budget.mjs. Units: MiB = 1,048,576 bytes; Mbps = 1,000,000 bits/second; 1 knot = 1852/3600 m/s.
Use effective forward advance a = 2000 m, corridor width w in cells and compressed payload b = 1 MiB/cell:
- Crossing scenario:
t = a / groundSpeed. - New-cell demand:
cells/second = w × groundSpeed / a. - Sustained transfer:
Mbps = cells/second × b × 8 / 1,000,000. - Lookahead:
distance = groundSpeed × totalReadyLatency + a.
totalReadyLatency includes queue wait, request/RTT, transfer, decode/transcode, GPU preparation and scheduling. Proposed 8 s normal / 20 s degraded allowances are assumptions to measure at p95/p99, not fetched-asset timings. The normal illustrative allocation is 1 s queue + 0.2 s request + 4 s transfer + 1 s decode/transcode + 0.8 s GPU preparation + 1 s scheduling/headroom. Parallel work does not erase these costs.
| Profile scenario | Speed m/s | 2 km crossing s | 8 s lookahead km | 20 s lookahead km |
|---|---|---|---|---|
| Trainer, 310 kt | 159.48 | 12.54 | 3.28 | 5.19 |
| Airliner, 510 kt | 262.37 | 7.62 | 4.10 | 7.25 |
| Fighter, 1320 kt | 679.07 | 2.95 | 7.43 | 15.58 |
| Additional ground-speed stress | 900.00 | 2.22 | 9.20 | 20.00 |
| Corridor scenario | New cells/s | Payload Mbps, before reserve |
|---|---|---|
| Trainer, 5 cells wide | 0.40 | 3.34 |
| Airliner, 5 cells wide | 0.66 | 5.50 |
| Fighter, 3 cells wide | 1.02 | 8.54 |
| Fighter, 5 cells wide | 1.70 | 14.24 |
| Fighter, 7 cells wide | 2.38 | 19.94 |
| 900 m/s, 5 cells wide | 2.25 | 18.87 |
These strip estimates exclude turns, diagonal/cube-face crossings, side views, retries, far terrain and collision sidecars. They are not worst-case bounds. Count actual swept-cell sets in the later prototype. At 25 Mbps × 70% usable = 17.5 Mbps, five-wide fighter demand plus a 20% reserve is 17.09 Mbps: marginal, not comfortable. Seven-wide and the 900 m/s stress exceed that allowance. At 10 Mbps × 70% = 7 Mbps, the full five-wide fighter preset fails analytically.
Fallback: keep the coarse parent, use a 256 KiB payload scenario with 4 km effective advance (L10 candidate), and defer decorations. Five-wide fighter demand then falls to 1.78 Mbps, before reserve. This is a target compression/fidelity combination to prove, not an achieved asset size. More preload cannot fix sustained bandwidth deficiency.
Propose at most 4 network requests, 2 concurrent decode workers, 1 texture/upload preparation step at a time, and 2 ms main-thread preparation per frame. Cap queued/in-flight bytes at 8 MiB and cancel stale direction requests without losing the parent. Keep close collision/coverage first, forward corridor second, turn/side coverage third, distant detail last. A 20 s fighter lookahead spans far beyond a 5×5 near patch; use coarse/mid cells along that lead, not an unbudgeted strip of fully detailed cells. Abrupt reversal must remain safe on coarse coverage. Request limits need load tests; they do not guarantee latency.
Concurrency must sustain the arrival rate. Four requests with a mean 4 s occupancy deliver only 1 cell/s, below the 1.70 cells/s fighter scenario. That scenario needs mean request occupancy below 2.36 s at four-way concurrency (and additional headroom), or a lower-detail payload. The 4 s transfer allocation above is a tail-latency allowance, not an assumed sustainable mean. Two decode workers similarly need mean occupancy below 1.18 s for this demand. Measure queue stability as well as p95/p99 ready latency; prefer fallback before raising concurrency and peak memory. Serialize parent replacement groups so the four-child overlap reservation is not multiplied by simultaneous swaps. Also measure whole-group throughput and slow-child blocking: individual occupancy limits are necessary, not sufficient to sustain serialized replacements.
Budget for ordinary PCs before raising fidelity
No representative worldwide PC census was established. Valve's October 2026 Steam survey is optional and anonymous and describes its customers, not all PC/laptop users.[9] Microsoft specifies a 4 GB Windows 11 minimum, which is an OS requirement, not proof of AeroSim capability.[8] Microsoft's Surface Laptop 3 specifications document 8/16 GB configurations with Intel Iris Plus or integrated Radeon graphics, demonstrating relevant non-gaming device classes but not their population share.[13] Together these justify testing constrained hardware, not claiming a majority is supported.
Proposed tiers:
| Tier | Hardware test class | Render target, not a promise |
|---|---|---|
| Lower baseline | 8 GB total RAM, integrated/shared GPU, ordinary quad-core laptop; include an older Intel UHD/Iris-class machine | 1280×720 drawing buffer, DPR 1, 30 fps; low textures/shadows; 60 fps exploratory only |
| Standard | 16 GB RAM, recent integrated GPU; include AMD and Intel | 1920×1080 drawing buffer, 30 fps required candidate; 60 fps stretch after measurement |
| Additional coverage | 8/16 GB Apple integrated/unified-memory hardware where available | Safari and Chromium paths; do not infer equivalent GPU headroom from RAM alone |
Record exact model, RAM, OS, browser, driver, power mode, thermal state and display resolution. No named device above has passed this Earth workload. Four-GB machines and software rendering are not initially promised; provide a capability explanation rather than crash.
Proposed lower-tier terrain allocation
| Resource | Budget or analytical allocation |
|---|---|
| Compressed near-cell GLB | target ≤1 MiB; hard candidate cap 2 MiB, otherwise split/refine or reduce fidelity |
| Decoded CPU + estimated GPU per near cell | ≤8 MiB, counting both copies |
| Near resident set | 25 cells × 8 MiB = 200 MiB |
| Mid/parent coverage | 64 MiB |
| Distant/startup coverage | 32 MiB |
| Steady terrain subtotal | 296 MiB |
| Conservative transient decode allocation | 4 jobs × 16 MiB = 64 MiB; deliberately above the 2-worker concurrency proposal |
| Replacement overlap | 4 cells × 8 MiB = 32 MiB |
| Pending download buffers | 8 MiB |
| Analytical peak terrain subtotal | 400 MiB |
| Terrain hard planning ceiling | 512 MiB including CPU/GPU allocations and uncounted terrain overhead |
| Whole-tab working-set planning ceiling | 1.5 GiB, to be measured separately including aircraft, physics, UI, workers and renderer |
A 5×5 set is a residency budget, not a guarantee that it covers the camera frustum or every S2 neighbour pattern. Reduce near detail if the selected set exceeds it. Use at most 500,000 terrain triangles and 100 terrain draw calls initially; whole-scene candidate ceilings 750,000 triangles / 150 calls, including additional passes. These are budgets to test, not hardware-derived laws. A 65×65 vertex grid has 8,192 surface triangles before skirts; 25 such cells use 204,800, leaving allowance for parents/edges. Low-altitude/city quality may be insufficient; do not quietly multiply the budgets.
Compressed download bytes are not resident memory. A 1024² RGBA8 texture with a full mip chain is approximately 5.33 MiB before duplicate CPU/transcode storage. This can consume most of an 8 MiB cell allocation on its own. Prefer restrained shared material palettes and smaller textures; count shared allocations once with reference ownership. KTX2/Basis transcoding targets depend on available GPU formats; test fallback expansion.[6] Geometry compression reduces transfer but still needs decoded vertex/index buffers.[7] Count both CPU and GPU copies conservatively on shared-memory GPUs, then measure rather than assume their physical duplication.
The Device Memory specification returns a coarse capability signal, not free VRAM; its current W3C publication is a Discontinued Draft dated 2026-10-20.[10] The Measure Memory proposal is not a W3C Standard, requires cross-origin-isolated capability for its method, and provides implementation-specific estimates.[11] Feature-detect optional telemetry; never require these APIs or infer safe residency from navigator.deviceMemory. Use application allocation accounting, renderer counters, browser tooling and OS process observations together. None alone proves total GPU/process usage.
Data, compression and progressive quality
Proposed first content: coarse relief, consistent water/coasts and a small material palette. No initial photogrammetry, global building layer, dense vegetation or satellite-texture requirement. Increase selected airports/cities later under the same stable IDs, with new immutable content versions and unchanged parent fallback.
Copernicus DEM provides GLO-30/GLO-90 worldwide land coverage at 30/90 m and describes a surface model including vegetation and structures, not bare-earth terrain. EEA-10 has restricted access; availability and licence/access conditions differ by instance.[12] Candidate evaluation must check voids, islands, polar coverage, water editing and vertical datum conversion against its product documentation. Do not treat nominal raster resolution as geometric accuracy, complete ocean bathymetry or flight-safe collision evidence.
Before any content generation, approve a source ledger for each input: publisher, exact release, retrieval date, coverage, resolution, horizontal/vertical datum, licence text and attribution, redistribution/derived-work rights, checksum and processing history. Keep DEM, imagery, land mask and authored mesh rights separate. Existing globe art is not flight-quality geography. The earlier Natural Earth terms-page request returned HTTP 406; the later counting appendix verifies the map-unit source through the downloaded archive and publisher-maintainer licence. This does not establish rights for DEM, imagery or terrain-generation inputs.
Prefer evaluating KTX2/Basis textures and meshopt geometry on the existing loader before selecting the encoding profile. Their glTF extensions define required capability declarations and fallback choices; an unsupported required extension must not be treated as a usable asset.[6][7] Pin decoder versions and review actual decoder licences, notices and provenance separately from source geography. Compression tooling does not grant redistribution rights to input data. Benchmark decoding/transcoding on the lower tier, including an uncompressed fallback within the same memory cap. No decoder migration or new dependency is installed by this strategy.
Evidence required before generation or implementation
All rows below are future gates — not passed tests. Documentation build/browser checks validate this page, not terrain flight.
| Validation | Proposed acceptance evidence |
|---|---|
| Grid/coverage | Geodesic bounds and swept-path cell counts; poles, antimeridian, cube edges/vertices, shared Europe/Asia cells, islands, coast buffer and no cross-continent wrap; empty/water distinct from missing. |
| Content contract | One GLB per published cell/LOD; reproducible source/hash/licence ledger; same ID after content revision; explicit datum/frame; ≤1 cm shared-edge transform tolerance; no collision discontinuity on visual LOD change. |
| Low-tier performance | Real 8 GB integrated GPU at 720p/30; p95 frame time ≤33.3 ms and p99 ≤50 ms during a 30-minute route, no increasing memory after repeated route loops; report frame hitches separately. |
| Standard performance | Real 16 GB integrated GPU at 1080p/30 with the same stability gate; 1080p/60 stretch p95 ≤16.7 ms, p99 ≤25 ms. 720p/60 is a separate comparison, not inherited success. |
| Browser/device matrix | Current Chromium/Edge and Firefox on Windows/Linux as applicable; Safari on macOS; record exact versions and WebGL/texture capabilities. CPU-throttled headless tests do not replace physical integrated-GPU testing. |
| Flight envelope | Trainer/airliner/fighter nominal speeds, 900 m/s stress, maximum supported altitude, dives/tailwind, tight turns, reversal, diagonal travel and high-speed coast/border approach; derive actual ground-speed and turn bounds. |
| Networks | Cold/warm cache; 100/25/10/5 Mbps, 30/100/300 ms RTT, 1% loss, 20 s stall and offline transition; record p95/p99 queue/fetch/decode/upload/ready latency and actual corridor demand. |
| Failure and lifecycle | 404, corrupt hash, failed decode, unsupported texture, GPU context loss, cancelled request, quota denial, stale manifest, hidden/resumed tab, continent change and memory pressure; parent stays until replacement is ready, safe pause before unsupported coverage. |
| Startup and memory | Coarse selected-area completeness before launch; ≤8 MiB startup payload candidate; resident/peak byte inventory, swap overlap, disposal and whole-tab/OS observations; reject expansion beyond budget, do not delete coverage to make a counter green. |
| Visual review | Low-altitude terrain, grazing-light seams, coast water, high-altitude horizon, mountain silhouette and parent/child transitions at desktop resolutions; approve quality separately from numerical performance. |
Next decision: review this proposal with Architecture/UI/UX, choose representative physical devices and a source-data candidate, then authorize a small validation sample separately. Only measured results should settle cell level, fidelity, latency windows, boundary handling and supported-hardware claims. A continent-wide generation run comes after, not as part of this publication.
All sources below were retrieved on 2026-11-10; their dated URL/title ledger is retained in research/earth-flight-strategy/sources.json. Application links require project permissions and are pinned to the inspected commit. Source standards describe formats and algorithms, not validation of the proposed AeroSim budgets.
Sources
[1] https://registry.khronos.org/glTF/specs/2.0/glTF-2.0.pdf — glTF 2.0 specification
[2] https://cesium.com/learn/cesium-native/ref-doc/selection-algorithm-details.html — Cesium selection algorithm
[3] https://s2geometry.io/devguide/s2cell_hierarchy.html — S2 cell hierarchy
[4] https://s2geometry.io/resources/s2cell_statistics.html — S2 cell statistics
[5] https://raw.githubusercontent.com/CesiumGS/3d-tiles/main/specification/README.adoc — 3D Tiles specification source
[6] https://raw.githubusercontent.com/KhronosGroup/glTF/main/extensions/2.0/Khronos/KHR_texture_basisu/README.md — KHR_texture_basisu
[7] https://raw.githubusercontent.com/KhronosGroup/glTF/main/extensions/2.0/Vendor/EXT_meshopt_compression/README.md — EXT_meshopt_compression
[8] https://www.microsoft.com/en-us/windows/windows-11-specifications — Windows 11 requirements
[9] https://store.steampowered.com/hwsurvey/ram?platform=pc — Steam October 2026 survey
[10] https://www.w3.org/TR/device-memory — Device Memory API
[11] https://wicg.github.io/performance-measure-memory — Measure Memory draft
[12] https://dataspace.copernicus.eu/explore-data/data-collections/copernicus-contributing-missions/collections-description/COP-DEM — Copernicus DEM
[13] https://support.microsoft.com/en-us/surface/models/surface-laptop-3-specs-and-features — Surface Laptop 3 specifications
[14] https://raw.githubusercontent.com/KhronosGroup/glTF/main/specification/2.0/Specification.adoc — glTF 2.0 specification source
[15] https://gitea.lego-cloud.eu/corp-v1-aerosim/corp-v1-aerosim/src/commit/16144cc74cac4008e72dcb4aa4b4cad14d74df2a/packages/flight-core/src/index.ts — AeroSim packages/flight-core/src/index.ts
[16] https://gitea.lego-cloud.eu/corp-v1-aerosim/corp-v1-aerosim/src/commit/16144cc74cac4008e72dcb4aa4b4cad14d74df2a/applications/web/src/flight/physics/production-physics.ts — AeroSim applications/web/src/flight/physics/production-physics.ts
[17] https://gitea.lego-cloud.eu/corp-v1-aerosim/corp-v1-aerosim/src/commit/16144cc74cac4008e72dcb4aa4b4cad14d74df2a/packages/flight-core/src/physics/evaluate-flight-forces.ts — AeroSim packages/flight-core/src/physics/evaluate-flight-forces.ts
[18] https://gitea.lego-cloud.eu/corp-v1-aerosim/corp-v1-aerosim/src/commit/16144cc74cac4008e72dcb4aa4b4cad14d74df2a/applications/web/src/flight/FlightCanvas.tsx — AeroSim applications/web/src/flight/FlightCanvas.tsx
[19] https://gitea.lego-cloud.eu/corp-v1-aerosim/corp-v1-aerosim/src/commit/16144cc74cac4008e72dcb4aa4b4cad14d74df2a/applications/web/src/flight/flight-render-policy.ts — AeroSim applications/web/src/flight/flight-render-policy.ts
[20] https://gitea.lego-cloud.eu/corp-v1-aerosim/corp-v1-aerosim/src/commit/16144cc74cac4008e72dcb4aa4b4cad14d74df2a/applications/web/src/flight/world/world-space-contract.ts — AeroSim applications/web/src/flight/world/world-space-contract.ts
[21] https://gitea.lego-cloud.eu/corp-v1-aerosim/corp-v1-aerosim/src/commit/16144cc74cac4008e72dcb4aa4b4cad14d74df2a/assets/packages/mountain-valley-100km/runtime-manifest.json — AeroSim assets/packages/mountain-valley-100km/runtime-manifest.json
[22] https://gitea.lego-cloud.eu/corp-v1-aerosim/corp-v1-aerosim/src/commit/16144cc74cac4008e72dcb4aa4b4cad14d74df2a/assets/packages/training-airfield-100km/training-airfield-runtime-manifest.json — AeroSim assets/packages/training-airfield-100km/training-airfield-runtime-manifest.json