03 / RESEARCH NOTES
Follow the
evidence.
Questions, controlled comparisons and the findings that survived the experiment. Including the negative ones.
01 / Research · negative resultCan graph structure or persistent topology improve multi-hop retrieval beyond semantic similarity?
The repository reports HotpotQA all-support@5 of 0.5100 for cosine and 0.4700 for graph + topology (400 held-out questions). The tested additions did not establish a replicated improvement. These reported measurements have not been independently rerun for this portfolio.
02 / Pilot complete · GPU study pendingDo reasoning policies learn correctness, or exploit noisy and inconsistent verifier rewards?
The repository reports a completed framework and pilot, including pretrained save/reload runs. The primary research hypothesis remains unestablished; the final GPU study is externally blocked.
03 / Research release · no advantage shownUnder controlled budgets, where do exact simulation, learned states and variational algorithms actually help?
The v0.1.0 README reports 569 records across seven small-system CPU studies. No quantum advantage is demonstrated. These results are repository-reported, not independently reproduced here.
EXPLORATORY / NOT A COMPLETED STUDYNeuroSim
Biophysical neuron models, RK4 integration and network simulation with model tests. Comparative benchmarks and validated HPC scaling have not been established, so this remains a simulator note rather than a completed research case study.
Explore the simulator source ↗These are summaries of published repository records, not independent replications, peer-reviewed publication claims or new experiments.