Gradientfall is an in-development 3D action-adventure that turns machine-learning concepts into geography, combat, quests, and exploration. The current build is an honest vertical slice: a playable Datasedge Meadows environment with a third-person controller, data-driven content, real-time combat foundations, the Bit companion, and a deliberately ambitious code-authored visual stack.
Explore: documentation map · current roadmap · worldbook · runtime architecture · run locally
| Day | Dusk | Night |
|---|---|---|
![]() |
![]() |
![]() |
These are in-engine progress captures, not target renders. They document the current visual stack and make later regressions easier to recognize.
- Godot 4.7 project with a typed GDScript architecture and autoload-based state/event boundaries.
- Third-person movement and camera controls in a generated 3D landscape.
- Melee/ranged combat foundations, enemy spawning, health, damage shards, and combat HUD.
- Bit companion behavior and landmark reactions.
- Bootstrap town with 13 approved villagers, lit buildings, and proximity dialogue.
- The Perceptron Vault's playable 2-3-1 neural-network dungeon and Gatekeeper encounter.
- A 700-bloom Iris field whose specimen measurements feed an in-game classifier compendium.
- Dynamic sky, celestial layers, clouds, water, terrain, grass, vegetation, particles, and painterly shaders.
- Schema-validated JSON pipeline for quests, NPCs, items, monsters, quizzes, lore, and points of interest.
- 111 approved content entries across all seven content types; the staging inbox currently validates empty.
This repository does not claim that the ten-region, 40–80 hour design is complete. The vision is documented in the GDD; the shipped scope and next gates live in the roadmap and devlog.
- Quality over speed. The single most important factor of this entire project. Milestones ship when they are verified, documented, and understandable by a stranger — never to keep pace. The full contract lives in CLAUDE.md.
- Documentation is a first-class deliverable. Every script carries a doc header and member docs to the standard in docs/ARCHITECTURE.md; every session appends a dated entry to the devlog; roadmap checkboxes and content-budget ticks are updated in the same commit as the work.
This locally generated image is an art-direction mood study, not gameplay and not locked canon. It explores the patient scale, warm twilight, and open traversal language of the future Tensor Desert; the Worldbook remains the authority for the region's golden dunes, buried matrix ruins, grid-patterned sands, and MNIST rune architecture.
content/ JSON schemas, inbox batches, and approved game content
docs/ GDD, worldbook, roadmap, devlog, and content workflow
game/ Godot project, scenes, typed scripts, and code-authored assets
tools/ content validation and authoring utilities
The runtime never hardcodes authored content. ContentDB loads validated entries from content/approved, while GameState owns durable state and EventBus handles cross-system signals. See docs/ARCHITECTURE.md and docs/CONTENT_PIPELINE.md.
Requirements: Godot 4.7.1 and Python 3.11+.
python tools/validate_content.py --inbox
python tools/validate_content.py --all
godot --editor --path gameThe local verification gate is:
godot --headless --editor --path game --quitGodot can return zero even when its output contains a script parse error, so CI also rejects SCRIPT ERROR, Parse Error, and failed-script-load messages.
Gradientfall was extracted from the danieltalbert profile repository with its path history intact, then updated from a verified source-only snapshot of active local work. Generated engine caches and agent-only metadata were deliberately excluded.
Original code, documentation, and code-authored assets are available under the MIT License. Dataset extracts will be added only with explicit source and license records before release.






