Tokyo study
Explore the Station district, compare actual gameplay captures, and inspect what each optimization cost.
The 100 KB goal is unmet. The retained JP-13 road selection, supplemental geography and profile use 2,861,052 bytes before gzip. Charging the categorized mixed source raises that to 2,941,318 bytes. Shared code and existing-file changes are tracked separately.
Gameplay and resolution comparison
Recorded from the running native game. Both settings keep the HUD and output at 640 × 360. The lower setting changes only 3D rendering. Videos play at 20 frames per second; that is the recording rate, not a device-performance claim. Buildings with unknown heights, facades and interiors are inferred.
| Route | 3D scale | Median draw calls | Median capture interval | Maximum visible names |
|---|
What is in the map
The road packet keeps all 158,235 selected source ways, including the remote Tokyo islands. This is the frozen JP-13 selection, not all of Greater Tokyo. Its inherited selection excludes service roads and footpaths. Building footprints concentrate around Tokyo Station; rail, water and parks cover selected reference areas. The orbital city patch spans at most 5.376 km before the coarser Earth view takes over.
There are 100 selected landmark anchors, with 25 landmark names and 25 street names stored. At most six names appear at once, and names can be switched off. These are orientation aids, not 100 bespoke or enterable landmark models.
Bytes and reuse
| Category | Bytes | Accounting |
|---|---|---|
| Road packet | 2,735,678 | City data |
| Supplemental geography | 124,649 | City data |
| Tokyo profile | 725 | City data |
| Identified Tokyo content in source | 5,135 | Additional city charge; included in mixed source below |
| New generic runtime modules | 57,783 | Tracked separately from city data |
| New mixed runtime source | 80,266 | Conservatively charged in the 2,941,318-byte figure |
| Existing runtime changes | +10,477 | Net source growth; not proven entirely generic |
| Offline build and measurement tools | 151,502 | 22 files, excluded from runtime |
Source bytes, compiled resources, engine download, and generated memory are different measurements. The byte ledger and export audit preserve those distinctions. The research report keeps its earlier inventories; the final profile hash adds 84 city-data bytes. The 2,941,318-byte figure is conservative for the categorized new files, not a proven exhaustive upper bound on all city-specific changes.
New Free Space results
Three exact offline experiments reduced roads to 2,436,242 bytes and supplemental features to 97,600 bytes. Together with the unchanged profile, that is 2,534,567 bytes before embedded Tokyo source. This is an accounting comparison, not an integrated game package. The playable build retains the verified bounded decoder.
Local street prediction with counted corrections reached 2,613,938 road bytes. All experiments kept the original runtime records and geography. Full reproduction and GIF archives accompany this delivery; the current report identifies each file.
Experiments worth keeping
Spatial blocks capped at 256 ways added 26,975 bytes and reduced measured cold-query p95 from 25.351 to 14.666 ms. Interior batching reduced median draw calls from 4,908 to 1,317 across the same 80-frame stair route. At most 42 of 230,400 pixels changed in any compared frame. These results are reproducible measurements from this environment.
The road grid sweep, footprint and yaw experiments, sparse-name selection, classic encodings, modern formats, and late Free Space experiments are in the research report. Coarser road coordinates saved only 17.2% between 0.5 m and 4 m and introduced many more rounded-coordinate collisions. The shipped build keeps 0.5 m roads and 0.25 m supplemental polygons.
Review material
- Final capture and resolution findings
- Research and measured results
- Original Astra Ultra execution plan
- Evidence index
- Final byte and reuse ledger
- Exported-resource audit
- Current delivery report and remaining byte gap
- Astra / Ultra execution prompt
- Geographic sources and rebuild tools
- Independent fidelity evidence
- Final review fixes and profile builder
Recovery and verification
The recovered runtime matches all 327 pinned source files. Roads and supplemental features rebuilt to their original hashes. The new 8,969,068-byte game pack has the same 308 resource paths; compiled scripts, assets, map data and project settings match. Fresh headless starts reached Tokyo, Tulsa and Hamburg. Generated scene/UID metadata differs after reimport.
All 14 original videos survived and decode completely. The separate capture bundle includes six original GIFs and eight clearly marked GIF replacements derived from those videos. Original source-frame sequences were not recovered.
What remains unaccepted
The full code verifier passes, including 664 Tokyo assertions. The source, export and capture receipts identify the reviewed revision. Visual parity with Hamburg and Tulsa, physical-phone and MacBook performance, and the 100 KB goal remain unaccepted. The browser available for this work could not start WebGL2, so native gameplay captures do not certify browser gameplay. The complete documentation build is blocked by missing historical media, although its content checks and static page generation passed.
The resolution pair compares the 3D render setting. The planned procedural surface-detail sweep and a perceptual recognition study remain unrun. Astra Ultra prompting guidance is dated research and an execution prompt, not a model benchmark.