RACER
▌ PROTOTYPE · MILESTONE 2 · WORKING CODENAME
A portrait, retro pseudo-3D racer for one thumb on each side of a phone. A run is five generated tracks against a checkpoint timer, with a fork and an upgrade after each. Every one of those tracks has been driven by a human-like bot, on every car you could arrive with, before it is shown to you.
"racer" is a working codename: the game has no name yet, and it's an early prototype. The iPhone beta is free through TestFlight. Android dev builds, the source and the design doc are for testers with access to the repository (dev build).
SEASONAL · OCT 19 – NOV 1 · EVERY YEAR
From October 19 to November 1 (by the same UTC day the daily run uses) every biome turns into its spooky twin. The title says HAPPY HALLOWEEN between two jack-o'-lanterns, bats cross a big moon, and the roadside is scarecrows, dead trees, gravestones and pumpkins. Only the paint changes: the tracks, the timers, the daily, ghosts and share codes are exactly the same as the rest of the year.
One run a day, the same for everyone: its seed is a hash of the UTC day, so it turns over at midnight UTC everywhere. The game builds all nine of the day's tracks in the background while the title is up, so DAILY starts at once; the title keeps today's tries and best.
seed = mix64(tag, UTC day)
Finish a track and your fastest drive becomes a ghost: a pale, see-through copy of your car replayed from its input log, with your time against it at every gate. RETRY SEED races the same five roads again; the daily's ghosts work the same way.
GHOST -0.8S
Four bundled courses, your saved layouts and an in-game editor: a
map and an elevation strip, bends, hills, width, surfaces, drift
corners, cars and gates. PROVE runs the same bots on your layout;
a proven one races timed and shares as an RL… code,
a draft as RD…. FROM SEED forks any generated track.
FIRST LIGHT · RIDGE RUN · GRAVEL PASS · LONG DUSK
A run is a day on the road: MEADOW, COAST, ALPINE, DESERT DUSK and NIGHT CITY, one per track, each with its own sixteen colours, horizon layers and roadside flora. Scenery only: no biome changes how a track drives.
6 scenery colours change · 10 stay fixed
Engine, tyres, gates, boosts, spins, the last five seconds and the title music, all generated by a standard-library Python script with no samples, byte for byte reproducible and checked in CI.
just gen-audio
Every sprite is drawn in code at startup. Cars are small solid models lit from the upper left and dithered between tones; the player's hatch has lean, drift and spin frames with brake lamps; traffic comes as a hatchback, a van and a pickup.
Art::new ≈ 6.5 ms, once
Procedural racers usually generate a road and hope. This one builds a track, then has two bots drive it, and only a track that passes every check is ever shown. Not "probably drivable": checked, on every loadout the player can reach it with.
A perfect bot is the ceiling: it plans the whole track's speed at 93% of grip and searches which corners to drift. A human-like bot is the floor: it reacts to 150 ms old information, can't see over crests, aims a little off its line, lapses now and then and overcooks the odd corner.
human: 150 ms late · 85% grip · 1% lapses/s
Each gate's time is calibrated from sixteen human-bot drives on the weakest car you could arrive in, plus a margin. Then fresh human drives, with noise the calibration never saw, must keep that margin on every reachable loadout, 15 times out of 16.
margin: 2.0 s on track 1 → 1.0 s on track 5
A draft that fails is redrawn; the best failure has its worst section softened; as a last resort comes a gentle road every car takes flat out. Drift-required corners must be passable with a drift and cost a grip-only driver at least 1.5 s.
8 redraws → 2 repairs → fallback
just balance 200 run seeds · 1,800 tracks unvalidated tracks ...... 0 (must be 0) fallback tracks ......... 1.33% (must be ≤ 2%) human runs lost ......... 3.9% (must be 3–15%) perfect bot time left ... 6–18 s (every track) human over perfect ...... 8–30% (every track)
Numbers from the M1 build, recorded in NOTES.md. CI runs the whole batch on every pull request; the human bot's numbers are being checked against real playtest runs.
About a minute a track, split by checkpoint gates. Time left at one gate carries to the next; running out is the only way to lose. After each track, pick the easier road or the harder one, each with its own upgrade: engine, turbo, tyres, brakes or charger. They stack.
Hold drift into a bend to slide wide and charge a boost; let go to spend it. Early tracks only hint at it. From track 3 some long, tight carousels are signed with chevrons and can't be made in time without a drift.
Blind crests hide what is over the top. Slow traffic in packs, one lane always free. Posts on the outside of bends, rocks in the lane, gravel, ice and puddles, and narrows on harder tracks. A hit spins you out and costs seconds.
Left, right, brake and drift pads under the road, lifted clear of Android's gesture bar. Steering ramps up over 200 ms so a tap is a nudge. Keys and a gamepad work too, and all of them reduce to the same one byte per tick, so a replay can't tell them apart.
These are the renderer's golden frames: fixed scenes whose pixel digests the test suite pins on Linux, macOS and Windows. The site build paints them fresh from the code, at 2×, nearest-neighbour.
The simulation is a Rust crate with no dependencies and no engine: a fixed 120 Hz step, its own trig instead of libm, SplitMix64 and PCG32 streams, integer tracks. Two hundred noisy bot runs, each replayed from its log, hash to the same digest on Linux, macOS, Windows, an Android emulator, a Pixel 10 and a Lenovo tablet.
0x4a5c1bcc8fc9fdfb
A run is a seed and an input log. A sim epoch names a digest over pinned tracks and a corpus of golden runs, so a change that would move a replay's outcome fails the tests until the epoch is bumped.
SIM_EPOCH = 1
The classic pseudo-3D road, rastered in Rust into one 180-pixel-wide buffer and uploaded as one texture; Godot only presents it. The height floats with the phone's aspect, so a taller phone sees more sky, not a smaller car.
180 × 320–400 · 16 colours
120 Hz on a charger, 60 on battery, and backing off when the phone runs hot, through godot-device-power. A stripped Godot engine keeps the APK near 28 MB.
charger 120 · battery 60 · hot 30
Every push to main replaces the rolling
dev prerelease with a new APK. It is a prototype: expect
rough edges. GitHub sign-in and repository access are
required; there is no store release.
Every run you finish or quit is saved on the phone as its seed and
input log. just pull-runs copies them off and sets each
track's time, closest margin and spins beside sixteen human-bot drives
of the same track: that comparison is how the bot gets tuned.
Verify with SHA256SUMS
or gh attestation verify.
Rust (pinned in rust-toolchain.toml), Godot
4.7.2 and just. The Android recipes also need the Android SDK
and NDK and JDK 21; the
README
has the details.
git clone https://github.com/cascadiacollections/racer
cd racer
just test # sim, renderer, golden runs; no Godot needed
just run # build the extension and play on the desktop
just balance # build and prove 1,800 tracksOn a phone: just android-template once, then
just android builds, installs and launches;
just android-digest checks the phone agrees with this
machine bit for bit.
| ← → / A D | Steer (ramps to full lock in 200 ms) |
|---|---|
| ↓ / S | Brake; the throttle is automatic |
| Space / Shift | Drift (hold, with some steering) |
| Esc / P | Pause |
main.Same studio, same rules: a deterministic Rust simulation that proves every level fair before you see it. Salt Line is an artillery roguelike: read the wind, break the ground, make every shell count. It has a free demo and is coming to Steam.