03 / Shinraya Tech Labs
Research reports, experimental findings, and implementation notes. These are repository-hosted materials; no peer-reviewed publication or conference acceptance is claimed.
The repository documents pretrained GRPO training and save/reload runs, independent GSM8K evaluation, model qualification, and a frozen execution package.
The primary hypothesis is unestablished. Pilot runs and software fixtures do not establish improved reasoning or reduced reward overfitting. The final GPU study remains externally blocked.
Released CPU studies preserve configurations, per-seed results, checkpoints, and checksum manifests. Small-system comparisons include failed optimizations and harmful mitigation outcomes.
No quantum advantage, GPU speedup, or hardware performance advantage is demonstrated. CUDA and IBM QPU runs were not executed for v0.1.0; the measurements are small-system CPU simulations.
Held-out HotpotQA and MuSiQue evaluations did not establish an improvement over semantic retrieval. Validation selected cosine-only retrieval as the primary V2 path.
Evaluation uses internal held-out subsets of public development data, not official hidden-test scores. The result concerns the tested formulations, not all topology methods. Local prototypes have no established privacy guarantee or independent security audit.
The public README describes the architecture, models, interfaces, and simulation commands. These document intended and implemented scope, rather than a measured scientific result.
The reviewed README does not establish biological validation, human-brain-scale performance, or verified MPI/CUDA scaling. No such results are claimed here.
Reading the evidence
Benchmarks are meaningful only with their datasets, configurations, comparison budgets, and uncertainty. The linked reports retain those details. This index does not turn a pilot, implementation, or point estimate into a broader claim.
An example: ATLAS claims & limits