Traditional AI decisions require gigabytes of weights and billions of tensor parameters. werr delivers instant, sub-millisecond cognitive reflexes with 0 Bytes of VRAM using a 24-byte mathematical seed.
Every decision begins as a seed c. werr iterates z → z² + c and watches the orbit swing around the plane, a wave that either settles or grows.
The moment |z| passes 2, the orbit can never come back. That escape step is the error signal, and it costs a few multiplications instead of a forward pass through billions of weights.
The window is cut into four quadrants and each one is measured the same way. How the escapes spread across the quadrants becomes a typed answer: noul, choice or score.
Without dense technical jargon, explained through intuitive real-world analogies:
LLMs like ChatGPT write long deliberative essays (System-2). werr acts like your hand pulling away when touching a hot stove (System-1). It is an ultra-fast, reflexive smart 'Yes/No' decision engine embedded directly inside software.
Traditional AI must store billions of weights (multi-GB files) in VRAM. werr stores 0 weight matrices! Decisions are synthesized on-the-fly from the infinite Mandelbrot fractal pattern in microseconds.
No waiting for cloud API calls or network hops. Runs locally on any cheap CPU, mobile device, or low-power microcontroller with zero electricity waste.
1. noul: Yes/No ('Should I unlock the door?')
2. choice: Multiple Choice ('Route to Emergency or Clinic?')
3. score: Ranking ('How critical is this threat? 0 to 3 scale')
Feed your app's variables (e.g., User role, request frequency, failed logins).
Ask werr: 'Is this operation safe?' or 'Which microservice should receive this?'.
Get deterministic, type-safe probabilities in < 1 ms to branch your code.
Execute smart if-statements with 0 VRAM and zero cloud latency.
// Run evaluation to view serialized telemetry record
| Dimension | TypeSafe AI (Jev) | OpenJev / NanoJev (Qwen 4B) | werr (Fractal System-1) |
|---|---|---|---|
| Weight Tensor Memory | Multi-GB (Cloud Hosted) | ~8 GB VRAM | 0 Bytes (Zero Tensor Memory) |
| Model Seed | Cloud API Token | Model Weights Checkpoint | 24 Bytes only (cx, cy, zoom) |
| Typical Latency | ~100 ms (HTTP API) | ~15–30 ms (GPU) | < 1.0 ms (Pure Local CPU) |
| Hardware Requirement | Internet Connection | CUDA-capable GPU | Any CPU / Microcontroller |
| License & Cost | Proprietary API ($/token) | Open Weights | Free for Research & Edu (BSL 1.1) |
Execute real benchmark suites on the live WERR kernel instantly — 100% in-process or via live REST endpoint with 0 byte tensor memory.
git clone https://github.com/pCwOrM/werr && cd werr && python -m unittest tests.test_windtunnel_webmcp_isolated
Papers, machine-checked proofs, open datasets and the systems built on the same reflex. Drag to turn the constellation; every star opens its source.