`DETAIL_LOT_METRES = 160` was one number for a board that stopped being looked at from one distance, and it was wrong at both ends of the range it is now looked at from. Measured with `scripts/cost-at.mjs` on the merged board against a 400,000-triangle cap: | pose | before | after | lot | | --- | --- | --- | --- | | Los Angeles, 13 km | **688,536** | 343,372 | 400 m | | Los Angeles, 12.5 km | **678,546** | 343,372 | 400 m | | Los Angeles, 11 km | **654,656** | 390,602 | 250 m | | Los Angeles, 9 km | **635,806** | 380,762 | 250 m | | **Los Angeles, 7.7 km** | **621,866** | **375,196** | 250 m | | Los Angeles, 2.5 km | **415,780** | 345,340 | 200 m | | Los Angeles, 0.9 km | 297,148 | 363,454 | 40 m | | San Francisco, 7.7 km | 222,286 | 300,782 | 80 m | | San Francisco, 2.5 km | 216,292 | 279,158 | 80 m | | San Francisco, 0.6 km | 200,114 | 347,830 | 40 m | | San Francisco, 400 km | 355,763 | 355,763 | — | | San Francisco, 45 km | 270,377 | 270,377 | — | **Six poses were over the cap and five of them had never been measured.** Only the 7.7 km one was on record; 2.5 km over Los Angeles was at 415,780, and the whole band from 9 km up to where `DETAIL_STANDOFF_M` switches the metros off at 13.3 km ran 635,806 to 688,536, because the frustum there holds essentially the entire Southland. None of them is over now. San Francisco goes the other way — 4.2x the buildings at the poses the complaint was about, and 15,613 at the closest one against 849. No cap moved. **Zero new draw calls, and that is a bound rather than a sample.** It is the same `InstancedMesh`, the same geometry and the same material; a district is built at 40, 80, 160, 200, 250 and 400 m, all six rungs live in the store, and the live buffers hold whichever one the board chose. Draws are unchanged to the digit at every pose above. A test asserts the mesh identity across four rung changes. The lever is a **lot count, not a lot size**, because a lot size is the wrong thing to hang on the camera: at the same 22.6 km of engine stand-off San Francisco has 1,987 lots in frustum and Los Angeles has 23,000, so any rule in metres of lot or kilometres of reach is a different bill in the two cities. The board packs the finest rung whose *visible* sum fits 19,000 lots, which is 400,000 minus the 184,000 to 205,000 triangles everything-but-the-city measured at across the whole band this runs in — flat to 10%, which is why there is no stand-off ramp and why the one that was here first came out. Costs, honestly. `createBlocks` goes 264 ms -> 1,236 ms and its store 4.7 MiB -> 18.9 MiB, because every rung is walked and kept; the live instance buffers get *smaller*, 59,166 -> 19,000, since the mesh no longer has to be able to draw a rung nothing can afford. A rung change pops — each rung is its own survey, not a subdivision — and cross-fading would cost the draw call this board does not have. Los Angeles at 7.7 km is visibly thinner than the 621,866-triangle version it replaces. `TODO.md`'s LA section is rewritten around what is left: no budget cell stands at any of these poses, which was the first item on it and still is. 1,723 tests.
Tera
The map view of Lumbridge Simulate — California from above, in three.js. Its other half, Spaces, is the offices you walk into: one engine and one asset library, seen from outside and from inside.
Apache 2.0. Runs at tera.lumbridgecorp.com.
What it is
An engine plus data packs. The default board joins Los Angeles and San Francisco with live deterministic traffic on US-101 and on the honest I-5 → I-580 → I-80 approach. Choose either route chapter to follow the procedural black Model X.
The engine renders terrain, coastline, built cities on authored street grids, bridges, roads, markers, road traffic and air traffic. A city pack is pure data — coastlines, hills, districts, landmarks, camera chapters — so adding a city is a data contribution anyone can review, not a fork.
California, the detailed Bay Area, and Los Angeles / Orange County / Riverside ship today. The corridor is intentionally sparse; detailed cities remain their own boards rather than forcing a 600 km world into one full-resolution mesh.
A plan view sits top right: the board drawn flat, with the footprint of the camera's own frustum on it, so you can see where you are looking from outside the shot. Click or drag it to move the camera; scroll it to dolly. It is a 2D canvas rather than a second WebGL context, drawn from the same city pack, and it follows the sun into the night along with everything else.
Who sees what
Three tiers, resolved once at boot by src/access.ts:
| anonymous | signed in | admin | |
|---|---|---|---|
| the map, the plan view, the named chapters | ✅ | ✅ | ✅ |
| observed weather and live aircraft | ✅ | ✅ | ✅ |
| the office | public depth — shell, furniture, viewpoints, nobody home | full depth, with presence | full depth |
| the marker feed | per TERA_MARKERS_ACCESS |
✅ | ✅ |
the godmode panel (G) — date, season, weather override, counters, pose editor |
— | — | ✅ |
The sky is public on purpose. Cloud cover over San Francisco is a government sensor reading, and the aircraft are broadcasting their positions unencrypted to anyone with a receiver; neither is something an account can grant you access to. Gating them cost the only moment that makes this project land — real fog rolling off the Pacific onto a city you recognise, at the real time of day, on a first visit.
The markers are the one feed that can carry something private, so the server
decides. TERA_MARKERS_ACCESS is members by default and an operator has to
say public out loud, which /api/v1/health then announces in degraded[].
The default is the safe answer rather than the common one, because the failure
mode is silent: nothing errors, nothing looks broken, the data is just readable
by the internet.
These are drawing decisions, not a security boundary, and src/access.ts
says so at length. Live data and office presence are withheld by the API, from
a caller it does not recognise; the client tier stops the app asking for
something it will not get. Admin is granted only by TERA_ADMIN_SUBJECTS on the
server — never inferred in the browser, and never from an API that failed to
answer. A deployment with no API at all is open, because "clone it and it works"
is the promise; it is not "clone it and you are an administrator".
Quick start
npm install
npm run dev
Play controls
The bottom mode dock is the local-player source of truth: View, Drive, Explore,
Fly, or office Walk. A transition clears stale held input and atomically hands
the follow camera to one subsystem. WASD is movement; Q/E is vertical or
yaw, I/K pitches the crow, Space is the primary action, G glides, P
resumes assistance, R resets, and C switches the driving camera. A standard
gamepad maps both sticks, triggers, shoulders, and rising-edge action buttons.
Touch play uses a pointer-ID analogue stick at lower left and only the actions that apply to the current mode at lower right. The Map button remains available during possession. Touch, keyboard, and gamepad state are independent, so a released or cancelled finger cannot clear another source that is still held. The UI and follow camera are presentation adapters only; they never enter Arena observations, rewards, snapshots, traces, or simulator hashes.
Headless RL environments
Tera also exports a versioned, renderer-independent Arena contract with five deterministic environments: US-101/I-5 driving, Frontier Valley office navigation, seeded SF/LA office robot jobs, crow waypoint flight, and California electric-aircraft flight. They share the client controllers and office plan, but require no canvas, DOM, Three.js scene, network service, or new runtime dependency.
Import them from @lumbridge/tera/arena. Seeded train/dev scenarios,
component rewards, safety terminals, maximum steps, snapshots, checksummed
traces, exact replay and executable inaction/scripted baseline proofs are
documented in ARENA.md.
The visible SF and LA office robots use that same fixed-step job state. Their patrol, parcel, inspection, and charging loops are authored demonstration scenarios—not presence, telemetry, or evidence of real company work—and the UI labels them as a seeded simulation.
Using the engine
import { createScene } from "@lumbridge/tera/engine/scene.ts";
import { createStage } from "@lumbridge/tera/engine/stage.ts";
import SAN_FRANCISCO from "@lumbridge/tera/cities/sf.ts";
// One stage per canvas, for the life of the page. Cities are put on it and
// taken off again; a renderer per city leaks its shadow map on every switch.
const stage = createStage(canvas);
const scene = await createScene(stage, {
city: SAN_FRANCISCO,
markerPalette: { hiring: 0x4ade80, closed: 0xef4444 },
});
scene?.setMarkers([
{ id: "1", lat: 37.7765, lng: -122.4241, label: "Somewhere", colorKey: "hiring" },
]);
createScene is async because the heightfield is built in a Worker — half a
million samples, about 730 ms on the Bay Area, and not on the main thread. It
resolves to null if the build was abandoned through options.signal, which is
what makes switching city mid-build cheap.
The engine renders Marker[] and looks colours up by colorKey in a palette
you supply. It does not know what your markers mean — that mapping lives in
your adapter. This is what lets one renderer serve a private map coloured by
one scheme and a public map coloured by another, without either being a fork.
Adding a city
Write src/cities/<id>.ts exporting a City. Trace the coastline and parks by
hand, place hills as radial peaks, and give each district its street bearing.
Two rules, and they are not stylistic:
- Do not import geometry from OpenStreetMap. OSM and Nominatim output is ODbL — share-alike, and incompatible with this repo's licence.
- Do not commit logos or brand assets. They are trademarks, not code.
See ARCHITECTURE.md §3 for the full reasoning, and
NOTICE for the attribution and data-provenance statement, and
PROVENANCE.json for the machine-checked shipped-artifact and
original procedural-lineage ledger. Run npm run provenance, npm run licenses,
and npm run sbom before accepting assets or dependencies.
Aircraft
The engine takes a FlightSource. Two ship here: SimulatedFlights (original,
flies real approach and departure corridors) and AdsbFlights (open community
ADS-B feeds such as adsb.lol).
FlightRadar24 is deliberately absent — their terms forbid scraping and forbid redistributing their data, so a client for it cannot live in an Apache-2.0 repository. Commercial sources belong in private deployments. The best long-term answer is an RTL-SDR receiver: first-party data with nothing to comply with.
Layout
src/engine/ renderer — terrain, blocks, structures, markers, flights, scene, minimap
src/cities/ data packs — pure geography, no code
src/transport/ serializable route packs and renderer-independent simulation
src/arena/ versioned headless RL contract, scenarios, traces and environments
src/assets/ original procedural asset library
src/adapters/ where outside data plugs in
src/tools/ instruments — god-only, dynamically imported, never statically
engine never imports cities; neither imports adapters.
Nothing under src/tools/ may be reached by a static import from the app. It is
loaded by one await import() behind access.can.debug, so a visitor who is
not an admin does not download the code at all — which is the strongest
available reading of "nothing here runs for a non-god visitor": not a hidden
panel, not a disabled panel, no panel. src/tools/index.ts states the rule and
what silently undoes it.
Licence
Apache License 2.0 — see LICENSE and NOTICE.
The ordered build plan and parallel work lanes live in BUILD_PLAN.md.
