world/translation/T99_Translation_Terrain.md
Status: ACTIVE v1.0 (flipped 2026-07-15 — critic verdict PASS-WITH-FIXES with all six findings fixed forward, then the contracts-corrections pass: the §0 flip condition).
Translation Doc — terrain/world-chain leg. Tier 99 operational (the runtime-inference layer per
T1_Build_Pipeline_Contracts §0); not world canon. Owns the platform-specific inference rules that
turn a region page's Section 1 + Zone Catalog into a built, populated UE 5.8 Landscape — the
parameter layer the Pipeline Contract and Region Page Template are silent on (exact request
shapes, resample formulas, the PARAMETERS BLOCK schema). Subordinate to the CVD, to
T1_Region_Page_Template [ACTIVE v2.0], and to T1_Build_Pipeline_Contracts §1 (UE5); where this
doc's inference conflicts with either, they win [SRC: T1_Build_Pipeline_Contracts §0].
Derivation basis (no web research this pass; every claim below traces to a committed repo doc,
tagged [SRC: file §section]): GEODATA_TERRAIN_CHAIN.md (the spine), RESEARCHED_STACK.md
(Chain 2 + the adjudicated-v2 addendum), ALTERNATIVES_MATRIX.md (Slots 1, 2, 4),
ADVERSARIAL_ADJUDICATION.md (the corrections — Lyra 2.0's license, the C++ import-shim severity,
the capture-rights gate; the briefs live in docs/pipeline_review/ + its tech_research/), `T1_Region_Page_Template
[ACTIVE v2.0] §1 + Zone Catalog, T1_World_Geography_Master [ACTIVE v2.3]` §14 (the
area_bounds/zone_id schema this doc extends), and the capability-boundary sections of
COSMOS3_DEEP_DIVE.md / VIDEO_TO_3D_BRIDGES.md (docs/pipeline_review/tech_research/). First
version, authored during the Phase D correction pass ahead of the Flores Island vertical slice;
versions per the standing Translation Doc discipline as region builds surface edge cases.
Specifies the terrain/world-generation leg of the UE 5.8 build chain for the 69 real-Earth
regions (69 is the CVD/CLAUDE cardinal; §14.8 additionally elevated four secondary sites — Japan,
Korea, Angkor, Rapa Nui — to first-class region pages, lifting the populated T0_Region_Index to
71 rows and the soft page-target to 71 without touching the canonical count
[SRC: T1_World_Geography_Master §14.8]). Answers RESEARCHED_STACK.md's
acceptance chain 2 for the terrain/GIS leg — download through populated Landscape — up to the
handoff into the engine build proper, which T99_Translation_UE5_Build owns from Landscape-import
onward.
Four ruled positions, adjudicated 2026-07-15, binding on every step below:
Gaea detail synthesis → 16-bit heightmap → UE Landscape is the only deterministic,
reproducible, license-clean, fully-batchable route across all 69 regions — the DEM already IS
the geometry [SRC: GEODATA_TERRAIN_CHAIN.md §7; ALTERNATIVES_MATRIX.md Slot 1 RECOMMENDATION].
into an editable, paintable, foliage-populated Landscape — materials/textures do not survive
the static-mesh conversion, and Cesium World Terrain is not available offline
[SRC: GEODATA_TERRAIN_CHAIN.md §1]. This demotes a Contract §1.2-named subsystem's role
exactly the way Contracts §0's platform-volatility clause anticipates: the WHAT (real-world
coordinate anchoring, global terrain reference) is unchanged, the HOW moves to the DEM
backbone, terrain_pipeline_type's enum stays schema-valid — a Translation Doc call, not a
Contract amendment (the same framing UE5_Build §1 applies to this tool split). Its role is
reference/preview/loading-screen only, now
including streamed gaussian splats (Cesium's April 2026 3D-Tiles-2.0 support) — a distribution
channel for an already-generated splat, not a tool for producing one
[SRC: VIDEO_TO_3D_BRIDGES.md §3].
Lyra 2.0 (§4) may generate reference imagery for fantastical vril-sites only, behind explicit
labeling and §17.1 — never real cultural terrain, never final playable geometry
[SRC: ALTERNATIVES_MATRIX.md Slot 1 RECOMMENDATION; ADVERSARIAL_ADJUDICATION.md Docket 1a].
✅ RULED (D-MPV, Josh 2026-07-21): a tiered ladder replaces the earlier one-number-per-region
framing — 1 m/vertex hero set-pieces (caves, boss pockets, ritual grounds) · 2 m standard
playable · 5-10 m ambient/vista — assigned per AOI on each region's terrain pass. Each AOI's
value is decided at the exact moment §2 Step 3 resamples that AOI's DEM; every other number
(Z-scale, tile count, World Partition cell count, erosion detail budget) falls out of it, now
per-AOI [SRC: GEODATA_TERRAIN_CHAIN.md §7]. The Liang Bua 30 m value stays a SMOKE artifact.
Input is a region page's header plus Zone Catalog: chapter_set, per-zone and region-ambient
area_bounds, and the terrain/gameplay_functions/ambient_threat/vril_density/climactic_
affordances/protected fields the spine's §3b rows populate verbatim, plus the page-authored
geometry the spine leaves downstream [SRC: T1_Region_Page_Template §Zone Catalog, §The
Consumption Boundary].
Area_bounds per zone is a polygon vertex array or center_plus_radius in WGS84 lat/lon (the
"Cesium coordinate convention," which is the same geodetic frame the Step 2 DEM API takes, so no
reprojection is needed) [SRC: T1_World_Geography_Master §14.2]. The region-level bounding box is
the union of every zone's area_bounds plus region-ambient rows — that union is the AOI Step 2
requests. Per-zone area_bounds carries forward to Step 4 (carving) and Step 8 (weightmap
alignment). Artifact out: one south/north/west/east decimal-degree bbox per region.
Primary call, one per region, scripted:
GET https://portal.opentopography.org/API/globaldem?demtype=COP30&south=<lat>&north=<lat>&west=<lon>&east=<lon>&outputFormat=GTiff&API_Key=<key>
— a free, instantly-issued key; Copernicus GLO-30 is global, free, explicitly commercial and
non-commercial without time limit [SRC: GEODATA_TERRAIN_CHAIN.md §2]. Fallback: same endpoint,
demtype=NASADEM (SRTM reprocessed, better void-filling); demtype=AW3D30 (JAXA, attribution
required) is the third cross-check tier where COP30/NASADEM disagree. The automated trigger for
falling back is not specified anywhere in the source research (only "where Copernicus has a gap
or artifact") — flagged, not invented, §5 item 5. Same-session calls also fetch ESA WorldCover
10 m and Sentinel-2 L2A true color over the same AOI, feeding Step 8
[SRC: GEODATA_TERRAIN_CHAIN.md §5]. Provenance — the DEM dataset actually used, the mandatory
Copernicus modified-data notice (Step 4 always modifies the DEM), and any AW3D30/JAXA attribution
— records via the page header's "Sources processed this session" field and Section 20's
population log; no separate provenance store is needed [SRC: T1_Region_Page_Template §Region
Identity header, §Section 20]. Artifact out: one raw GeoTIFF DEM (float, EPSG:4326), one
landcover GeoTIFF, one Sentinel-2 set — per region, native projection/resolution.
Deterministic and batchable across all 69 regions: reproject to a consistent local UTM/
equidistant projection (gdalwarp); mosaic DEM tiles crossing the AOI (gdalwarp/gdal_merge);
clip to the Step 1 bbox; resample to the region's meters-per-vertex (`gdalwarp -tr <x_res>
<y_res>`) [SRC: GEODATA_TERRAIN_CHAIN.md §4 Leg A]. The load-bearing formula:
meters_per_vertex = real_AOI_footprint_m / target_landscape_vertex_count, where the target is
either one 8,129-vertex Landscape actor or an N-tile World Partition grid (Step 6). The ratio
between area_bounds' real km² and the chosen footprint IS the scale-compression policy — one
value per AOI, not a manual redesign [SRC: GEODATA_TERRAIN_CHAIN.md §7]; the concrete values are
✅ RULED as the D-MPV tiered ladder (§5 item 1 — 1 m/vertex hero set-pieces · 2 m standard
playable · 5-10 m ambient/vista, assigned per AOI on each region's terrain pass). 30 m source pixels do not upscale
into finer real detail by resampling alone — a finer meters-per-vertex than native 30 m smooths/
terraces rather than manufacturing detail; Step 5 closes that gap, not the resample step
[SRC: GEODATA_TERRAIN_CHAIN.md §3]. Artifact out: one analysis-ready elevation GeoTIFF,
pixel-aligned with the landcover and imagery GeoTIFFs for Step 8.
rasterio.features.rasterize burns each fantasy/vril-site footprint into the elevation raster,
followed by a numpy blend/feather at the boundary to avoid a visible seam
[SRC: GEODATA_TERRAIN_CHAIN.md §4 Leg A] — the scripted, reviewable, re-runnable leg; every
alteration that must apply identically and unattended across all 69 regions belongs here, never
as a hand edit. Input contract: a per-region, page-authored table — one row per zone whose
canonical_anchors reference a T0_Vril_Site_Registry site_id, joined to that zone's own
area_bounds for the carve polygon. The registry supplies WHAT (site_name, site_class,
vril_function, mastery_tier_required); the page's Zone Catalog supplies WHERE (the registry
carries no footprint column) [SRC: T1_World_Geography_Master §14.7; T1_Region_Page_Template
§Zone Catalog]. Neither this doc nor the carve script re-authors site canon. Era modifications
(sea level, coastline, glaciers, deforestation reversal, river courses, volcanic activity) apply
here too, as a further chapter_set-era-keyed pass using the same mechanism
[SRC: T1_Region_Page_Template §Section 1]. The reversible/hand leg (post-import Landscape Edit
Layers / Patches / Splines) is T99_Translation_UE5_Build's, not this doc's
[SRC: GEODATA_TERRAIN_CHAIN.md §4 Leg B]. Artifact out: an altered elevation GeoTIFF per §3's
alteration list.
The resampled+altered DEM routes through Gaea's erosion/thermal/snow graph (.tor, CLI/
automation), running physically-scaled ("Real Scale") erosion to synthesize sediment and
meandering rivers a raw 30 m source cannot contain, exporting back to 16-bit
[SRC: GEODATA_TERRAIN_CHAIN.md §6] — the middle synthesis step only, not download (Step 2) or
import (Step 7). Licensing: Gaea Indie, $99 one-time perpetual, covers a solo dev under $100K
revenue; its 8,192×8,192 export cap lines up with the ~8,129-vertex Landscape ceiling for
single-tile regions. CLI-tier unknown, flagged not assumed: whether headless automation needs
Professional ($199) instead of Indie is UNCONFIRMED — the edition page names no CLI restriction,
but one search result claimed Pro/Enterprise is required; confirm with QuadSpinner before pinning
$99 for an automated pipeline [SRC: GEODATA_TERRAIN_CHAIN.md §6; ADVERSARIAL_ADJUDICATION.md
Docket 1c] (§5 item 2). Gaea 3.0's shipped-status is also UNCONFIRMED — target Gaea 2.3 (confirmed
shipped June 2026) until 3.0 is confirmed live (§5 item 3). What flips this step: a benchmark
showing Gaea's erosion is not perceptibly better than UE Landmass + auto-material at the chosen
scale drops it entirely [SRC: ALTERNATIVES_MATRIX.md Slot 1 "WHAT FLIPS IT"] (§5 item 9). Artifact
out: a detail-synthesized elevation raster, ready for 16-bit emission.
Format conversion: gdal_translate -ot UInt16 (or QGIS's Heightmap Export plugin) → 16-bit
grayscale PNG or RAW16+JSON [SRC: GEODATA_TERRAIN_CHAIN.md §3]. Z-scale: `Z Scale = (max height
in meters) × 100 (m→cm) × 0.001953125 (=1/512)`, from the region's real min/max elevation plus
headroom for anything Step 4 added; at the Landscape default Z-scale of 100 vertical range is
±256 m, so any region spanning more must set Z-scale explicitly. Tiling math: a single Landscape
actor caps at 8,129×8,129 vertices (127 quads/section × 2×2 sections × 32×32 components); the raw
heightmap import FILE ceiling is far larger (up to 43,351×43,351 px) specifically so a larger
region slices into multiple 8,129-capped actors under World Partition. Real-world World Partition
reports ~32 km × 32 km (1,024 km²) with foliage before performance gets heavy — past ~64 km² it's
worth asking whether sub-meter resolution is needed everywhere [SRC: GEODATA_TERRAIN_CHAIN.md §3].
Artifact out: the 16-bit heightmap(s) + per-layer weightmap inputs + Z-scale + tile layout — the
complete packet T99_Translation_UE5_Build's Landscape step consumes.
This doc's chain ends at the emitted heightmap packet; landing it in Unreal is
T99_Translation_UE5_Build's step, noted here only to mark the handoff and one dependency Step 6's
timeline must respect. The dependency: stock UE Python has no one-call heightmap-to-Landscape-
creation path — landscape_import_heightmap_from_render_target() only writes into an
ALREADY-EXISTING Landscape at render-target precision. Creating the actor from a heightmap
requires a small C++ shim wrapping ALandscape::Import() (or a plugin such as TAPython's
unreal.PythonLandscapeLib) — the single highest-priority engineering dependency in the whole
chain, correctly scoped as bounded engineering, not scripting [SRC: RESEARCHED_STACK.md Chain 2
item 4; ADVERSARIAL_ADJUDICATION.md Docket 2] (§5 item 10 — owned by T99_Translation_UE5_Build,
sequenced pre-Ch2-5-slice). Import target: classic heightfield Landscape, not UE 5.8's
experimental Mesh Terrain — real (true XYZ-mesh, overhangs, non-destructive modifiers, native PCG/
World-Partition) but experimental this release with an undocumented scripting surface; tracked
for a later re-target [SRC: ALTERNATIVES_MATRIX.md Slot 1 Alternative C] (§5 item 11). Seam: per
the three-seam routing rule, Landscape import is BULK/STRUCTURED — native Python + Remote-Control
RECOMMENDATION].
Reclassify WorldCover's 11 classes (tree cover, shrubland, grassland, cropland, built-up, bare/
sparse vegetation, snow/ice, permanent water, herbaceous wetland, mangrove, moss/lichen) into the
project's material-layer taxonomy; rasterize each into its own 0-255 weightmap on the SAME grid
as Step 6's heightmap; import via landscape_import_weightmap_from_render_target()
[SRC: GEODATA_TERRAIN_CHAIN.md §5] — with the sibling caveat carried: writing weightmaps onto an
EXISTING Landscape via Python is exactly the seam UE5_Build Step 5 flags as having no first-class
public painting API (the confirmed paths are weightmaps at import time, the auto-material, or PCG
writing layers); this step's render-target call runs at that import-time point in the chain, and
the two docs share one caveat, not two stories. PCG Biome Core (engine-bundled since 5.6,
carried into 5.8 — but Experimental, v0.2 per UE_BUILD_AUTOMATION §3.1; UE5_Build Step 5's
"verify runtime performance before shipping it" caveat governs here too)
reads landscape layer weights as a generator subtype. Zone-catalog gameplay affordances should drive WHICH PCG graph runs and how it's
parameterized per zone, via Bounds/Volume actors keyed to zone polygons, not landcover alone. Flag:
the landcover→layer→PCG-spawn chain is documented; the zone-polygon→PCG-bounds-actor→graph-
selection binding has no existing product feature to point at — it has to be built (§5 item 12).
Sentinel-2 true color can additionally feed the Landscape's base albedo/diffuse texture using the
same reprojection-to-grid treatment, no extra API needed. Artifact out (terminal state for this
doc): a populated, playable Landscape with paint layers and PCG-spawned foliage/props/encounter
volumes matching the zone catalog; NPC/quest/cinematic population is out of scope.
Section 1 of every region page must carry the following, machine-readably, for Steps 1-8 to run
unattended. The Zone Catalog (T1_World_Geography_Master §14.2) already owns zone_id, area_bounds,
gameplay_function, ambient_threat_level, vril_density, climactic_affordance_tags, and
canonical_anchors — this block adds only the chain-specific fields Section 1's prose calls for
(bbox, meters-per-vertex, the alteration recipe, water plane, era modifications
[SRC: T1_Region_Page_Template §Section 1]) without schematizing them field-by-field, which is
this Translation Doc's job per BPC §0 [SRC: T1_Build_Pipeline_Contracts §0].
| Field | Type | Source | Notes |
|---|---|---|---|
| region_bbox | south/north/west/east, decimal degrees | derived — union of zone area_bounds + region-ambient rows | Step 1 output; not separately authored |
| meters_per_vertex | float | page-authored per AOI, per the D-MPV ladder (§1; §5 item 1 — RULED 2026-07-21) | the scale-compression parameter, assigned per AOI (Step 3) |
| target_footprint | enum {single_tile, world_partition_N} + tile layout | derived — meters_per_vertex × bbox area vs. the 8,129 cap | feeds Step 6 tiling |
| alteration_list | ordered table: zone_id, site_id (FK → T0_Vril_Site_Registry), operation {carve, composite, raise, flatten}, feather_radius_m | page-authored, one row per FK'd zone | Step 4 input |
| dem_source_used | enum {COP30, NASADEM, AW3D30} + attribution_notice | recorded post-download, Step 2 | provenance rides Section 20 |
| water_plane | float, meters absolute + reference {sea_level, region_relative} | page-authored | named in Section 1 prose, unschematized until this doc |
| era_modifications | structured: sea_level_offset_m, coastline_shift_geometry, glacier_coverage_pct, deforestation_reversal_bool, river_course_shift_geometry, volcanic_activity_state | page-authored | a Step 4 pass, keyed to chapter_set's era |
| gaea_pass | bool + .tor graph reference | page-authored or defaulted | Step 5; false only where relief is negligible |
Every field is either already-canonical (read from the Zone Catalog / T0_Vril_Site_Registry,
never re-authored) or a build parameter this doc adds at Section 1 — consistent with the page
template's CONSUME/FEED split [SRC: T1_Region_Page_Template §T0 Registry Contract].
None of the lanes below touch the shipped-asset path (§1's third ruled position); all exist for
fantastical vril-site concept/blockout reference only, gated as noted.
trajectory → 60s/720p spatially-consistent video, genuinely clean and RTX-5090-capable (a
distilled NVFP4 variant does 60s in 34s) [SRC: COSMOS3_DEEP_DIVE.md §3]. Video-only — no native
3D reconstruction; a mood tool, not a geometry source, unless paired with a reconstruction step
that carries its own consistency risk.
Nano is a tight single-5090 fit (29GB+ BF16 alone, offload required, no confirmed quantized
consumer path). Output is video/image/audio/action, never mesh/depth/point-cloud/USD; "action"
is robot control-signal space, not a cinematic camera path, so despite the surface similarity to
SANA-WM it is not a fly-through generator [SRC: COSMOS3_DEEP_DIVE.md §1, §CAPABILITY BOUNDARY].
At most a low-value moodboard aid; no game-dev use was found anywhere in the source research.
Model matrix corrected to 16B/64B here — a sibling brief's stale 2B/14B figure does not apply
[SRC: ADVERSARIAL_ADJUDICATION.md Docket 1c].
§1.5], Marble Pro $35/mo minimum for commercial rights [SRC: ALTERNATIVES_MATRIX.md Slot 1
Alternative D]) — the one Family-B lane exporting actual
mesh + gaussian splats + video, not just video. The sanctioned narrow previz lane for
fantastical set-pieces because it returns real geometry, gated behind schema-bound + §17.1,
never real cultural terrain, never final geometry [SRC: ALTERNATIVES_MATRIX.md Slot 1
Alternative D].
(engineered against spatial-forgetting/temporal-drift; outputs an actual .ply splat), but its
WEIGHTS are licensed "NVIDIA Internal Scientific Research and Development" — non-commercial, no
production use, no works for sale. RULED correction: a sibling brief's "Apache 2.0, code and
weights" claim is factually wrong (same HF card; text fetched and quoted verbatim)
[SRC: ADVERSARIAL_ADJUDICATION.md Docket 1a]. Previz-only at most, same tier as Hunyuan3D and
gated Marble — never shipped-path — and its 43GB VRAM floor sits above the 5090's 32GB
regardless, so even previz use is an unconfirmed spike, not a working recipe today
[SRC: VIDEO_TO_3D_BRIDGES.md §1.4.2].
producing one; consistent with §1's Cesium ruling [SRC: VIDEO_TO_3D_BRIDGES.md §3].
with no documented working recipe, and is a strictly lossy round-trip for terrain AUTHORING —
the DEM already IS the geometry. Its only legitimate scope, if ever built, is concept art /
cinematic flythroughs silhouette-locked to the real DEM, never a terrain shortcut
[SRC: VIDEO_TO_3D_BRIDGES.md §4]. Flagged, not adopted.
The §17.1 rule (binding on every lane above and on real-capture reconstruction alike): never
ship AI-*generated* "documentation" of a real sacred/cultural site. Reconstruction tooling
(Postshot, Polycam, gsplat, RealityCapture) is cleared ONLY for genuine captures — real photos or
video of the real place. The moment the video itself is AI-generated, running it through a
splat/reconstruction pipeline produces an asset wearing photogrammetry's credibility as a costume
— every visual trust signal of a survey, none of the ground truth — a SHARPER version of the
§17.1 risk on file for raw generative architecture, not a mitigation of it. Adjudicated CORRECT
and load-bearing; keep it verbatim [SRC: VIDEO_TO_3D_BRIDGES.md §"genuinely adds value" item 1;
ADVERSARIAL_ADJUDICATION.md Docket 3].
The capture-rights gate (new — first formal specification of this for the terrain/previz
lane): real-capture reconstruction of a real cultural or sacred site additionally needs a per-site
capture-rights/consent check before the resulting mesh/splat ships — heritage sites and museums
may prohibit commercial 3D capture, and crowd-sourced tourist photography is third-party
copyrighted. "Real capture" clears the §17.1 representational-care bar; it does NOT by itself
clear legal provenance. Name the capture's rights basis (own capture, licensed stock, public-
domain/CC, museum-permissioned, or unconfirmed) per asset before it ships, the way the
book-substrate lane uses use_class [SRC: ADVERSARIAL_ADJUDICATION.md Docket 3] (§5 item 13, owner
Josh/art-pipeline).
Pulled from GEODATA_TERRAIN_CHAIN.md's seven residual unknowns plus the adjudication's
terrain-relevant additions, triaged into RESEARCHED_STACK.md's three classes.
| # | Item | Class | Trigger | Owner |
|---|---|---|---|---|
| 1 | Target meters-per-vertex — ✅ RULED (D-MPV, Josh 2026-07-21): per-AOI tiered ladder — 1 m/vertex hero set-pieces (caves, boss pockets, ritual grounds) · 2 m standard playable · 5-10 m ambient/vista, assigned per AOI on each region's terrain pass; the Liang Bua 30 m value stays a SMOKE artifact | RULED | assigned at each AOI's Step 3 resample | Josh — traversal pacing, area_bounds |
| 2 | Gaea CLI tier ($99 Indie vs. $199 Pro) | INSTALL-NOW | before any headless Gaea run is scripted | direct QuadSpinner confirmation |
| 3 | Gaea 3.0 shipped-status | WATCH | before pinning any 3.0-only feature | whoever scripts Step 5 next |
| 4 | "An NVIDIA world-gen tool" premise — no matching shipped product found | JOSH-RULES | before assuming a Gaea-replacement is live | Josh — name it or drop the premise |
| 5 | DEM fallback trigger — no automated void/artifact-detection rule specified | ENGINEERING-SPIKE | first real region GDAL run surfaces a visible void | pipeline/harness engineering |
| 6 | PCG Biome Core exact node-graph wiring | ENGINEERING-SPIKE | before Step 8 is scripted, not just doc-read | UE5-build engineering |
| 7 | Mesh Terrain as a future Landscape re-target | WATCH | Mesh Terrain reaching Beta/Production with a documented Python surface | track only |
| 8 | Landscape-import C++ shim (ALandscape::Import()) | ENGINEERING-SPIKE | pre-Ch2-5-slice | T99_Translation_UE5_Build |
| 9 | Zone-polygon → PCG-bounds-actor → graph-selection binding — no existing product feature | ENGINEERING-SPIKE | after Step 8 lands for the first region | UE5-build / PCG engineering |
| 10 | Capture-rights gate needs an operational process, not just a named rule | ENGINEERING-SPIKE | before any real-capture reconstruction asset ships | Josh / art-pipeline |
| 11 | Gaea-vs-UE-Landmass erosion-quality benchmark | BENCHMARK-DAY | 5090 landing | 5090 bench-test protocol |
| 12 | 69-region GDAL/import throughput — asserted "batchable," never quantified | BENCHMARK-DAY | before the Ch 2-5 slice is called validated for throughput, not just correctness | pipeline engineering |
Item 1 was the single most consequential open item in this doc — ✅ RULED (D-MPV, Josh
2026-07-21) as a per-AOI tiered ladder (1 m/vertex hero set-pieces such as caves, boss pockets,
and ritual grounds · 2 m standard playable · 5-10 m ambient/vista), per-AOI assignment riding
each region's terrain pass and the Liang Bua 30 m value kept as a SMOKE artifact; every other §2
number (Z-scale, tile count, World Partition cell count, Gaea's erosion budget) derives from each
AOI's chosen value once set [SRC: GEODATA_TERRAIN_CHAIN.md §7]. Items 5, 6, 8, and 9 are the genuine engineering-spike cluster
on the terrain chain's critical path: correctly specified, not yet costed or de-risked
[SRC: ADVERSARIAL_ADJUDICATION.md Docket 2].