Overview
The precursor work began in July 2026 with a practical question: can information about the work a program declares survive into execution? Early category-owned routing and concurrent local-workload experiments included a fairness run that showed essentially no advantage over ordinary concurrency. Concurrency alone was not treated as proof.
By September 28, 2026, the work had become a compiler-and-runtime contract. Nioret Semantic Language (NRL) states typed application work; NSE carries declared relationships into deterministic artifacts and can select work affected by changed inputs. CPU and CUDA are downstream backends, not the definition of that contract.
Current implementation
The 1.0.0a4 research SDK produces inspectable bundles through a deterministic CPU path. An independent C++ consumer compiled and ran its generated CPU output.
- Changed-input execution: Gate 1 passed correctness checks across five graph shapes and every nonzero input-change mask used by the suite.
- Planner development: an early exhaustive planner reached a measured scaling limit. The current sparse path avoids materializing every changed-input combination, although dense changes can still favor full execution.
- Benchmark evidence: a controlled CPU package covers eight synthetic workload families with correctness checks and paired randomized trials. Quantitative results remain under benchmark and IP publication review.
- Device backend: the implementation emits deterministic CUDA source, but the recorded environment did not compile or run it on a CUDA device.
Limitations and direction
The demonstrated graph and arithmetic model is bounded. General NRL control flow does not yet become arbitrary NSE selective execution, and policy for choosing between selective and full execution under dense change remains unfinished.
- The SDK is alpha, with deliberately narrow type and control-flow support; it is not a production platform.
- The evidence is synthetic, single-environment, and not independently reproduced. Physical CUDA qualification, production workloads, and additional machines remain outstanding.
- Compiler transformations, internal representations, dependency construction, scheduling, storage mechanisms, and processor mechanisms remain private.
The long-term research target is purpose-built NSE-oriented compute for larger AI, simulation, and scientific workloads. No native NSE processor is implemented.