3D scenes for AI tools
Cogniform
A 3D engine that lets software and AI tools change a scene, then inspect images and object data from the same scene version.
Early, unreleased engine. The linked capture is a small local cube example, not a general AI benchmark.
Color
Depth
World normals
Entity identityLatest verified change · CF081 / PR #81 ↗ · Merged
A runnable rendered-observation example
CF081 provides a runnable local example of revision-linked rendered observations. Color, depth, surface normals and entity identity stay tied to the source scene and frame. It remains a bounded example, not a general agent benchmark.
Source checked . Documentation and retained evidence review, not a new runtime test.
Read change evidence ↗Captured output / 10 September 2026
One cube. Four ways to understand it.
A picture tells only part of the story. Select a view to inspect appearance, distance, surface direction or object identity from the same scene.
How does a tool connect a pixel to an object?
All four observations belong to frame 1 and scene revision 0. A tool can connect a pixel to the world it is allowed to change.
Original 64 × 64 diagnostic pixels captured on 10 September 2026, enlarged without smoothing. A single local example, not a performance or cross-device benchmark.
In plain terms
Give an AI tool a scene it can inspect, including what each object is and how far away it is.
Why it matters
A picture alone does not tell a software tool which object it is looking at. Cogniform returns a view of the scene together with data that helps a tool interpret it.
Follow the example, step by step
Describe a scene
Supply objects and their positions through a structured command.
Keep a numbered version
The engine checks the change and records the resulting scene version.
Inspect the result
Color, distance, surface direction and object identity all refer to the same version.
Validate a patch before changing the world
Documented approach
The accepted atomic-world decision uses preflight over touched entities, followed by a prepared commit plan. Stable external IDs remain separate from recyclable internal ECS handles.
Alternative and scope
The record considers cloning the full ECS before each patch and applying changes before attempting rollback. It rejects whole-world copying for sparse edits and the partial-mutation risk of undo-based recovery.
Cost of the choice
Entity and idempotency capacities are bounded and fail closed. Allocation failures and internal invariant panics are process-fatal, rather than recoverable rejected patches.
What the evidence establishes
This is the accepted CF002 decision record, not a fresh benchmark or a statement that later features were absent forever. The current profile and README describe the subsequently expanded engine.
Read the decision source (opens in a new tab)A world and its evidence
Atomic world revisions and revision-linked observations sit behind bounded local adapters. The envelope detects corruption; it does not authenticate a remote caller.
Intent becomes an accountable world
Cogniform is built around a strict causal loop: bounded intent becomes an atomic scene revision; rendering produces color, depth, normal, entity-ID, and visibility observations; each result stays linked to the revision that produced it.
That boundary gives human and software-agent callers the same inspectable world model without requiring a window or a remote service.
Ownership is explicit across the engine
The Rust workspace separates protocol, deterministic compilation, content-addressed assets, procedural scene patches, authoritative world state, replay, rendering, orchestration, storage, and the local CLI.
- Transactional mutation, canonical events, logical hashing, and deterministic replay
- Bounded GLB admission with caller-driven decode, upload, persistence, and rehydration
- A headless GPU renderer with revision-linked machine-readable observations
- Local typed service, inherited-stream sessions, and a bounded MCP stdio adapter
Evidence includes the boundary
The repository checks in dependency sources for offline locked builds and documents validation, recovery behavior, threat boundaries, architecture decisions, and release readiness alongside the implementation.
The project remains an early core. Its supported imported materials and headless observations are bounded subsets; remote transport, automatic rollback, and a published release remain outside the documented baseline.
One frame, four observations
A fresh run of Cogniform's built-in headless cube example produced these four unmodified 64 × 64 diagnostic images. All belong to frame 1, scene revision 0 and the same camera, as recorded in the original manifest.
Compare the four captured observations ↑
Captured on 10 September 2026 with the Vulkan backend on an NVIDIA GeForce RTX 5070, using a debug build of source revision 7e607c515b50. This single local example is not a performance benchmark, protocol payload, conformance result or cross-adapter baseline.
Read the pinned capture guide (opens in a new tab)Related writing
Related updates
Inspect the checked source
Evidence reviewed . Individual decisions and captured examples retain their own source revisions and limitations.
Read the checked README (opens in a new tab)