An independent research line on self-improving cognitive systems, building the Continuous Improvement Engine (CEI).
The Continuous Improvement Engine (CEI) is a research effort on self-improving cognitive systems. It studies how an agent can refine its own behaviour across a recurring cognitive cycle — perception, deliberation, action, and reflection — under pressure that is not merely tolerated but converted into structural improvement (antifragility), with learned rules accumulated, revised, and pruned over time rather than encoded in advance.
This research line originated within applied work in precision poultry genetics (Gallora) and now operates as an autonomous, independent research effort.
Two contributions to energy-efficient manycore computing: a simulation-validated decentralized thermal-state routing algorithm achieving 92.1% thermal variance reduction under 4× localized overload, and a theoretical ENAQT-inspired CMOS topology proposed to convert Johnson-noise thermal fluctuations into charge-transport assistance.
Cite this work
@misc{barceloscosta2026thermal,
author = {Barcelos Costa, Cleber},
title = {Decentralized Thermal-State Load Routing and an {ENAQT}-Inspired
Circuit Design Principle for Energy-Efficient Manycore Architectures},
year = {2026},
doi = {10.5281/zenodo.19857070},
publisher = {Zenodo},
url = {https://doi.org/10.5281/zenodo.19857070},
license = {CC BY 4.0}
}
A hybrid retrieval method combining SimHash content addressing with weighted majority-vote aggregation of 1-hop graph neighbor signatures, achieving MRR 0.914 on a 392-node typed knowledge graph — a 2.13× improvement over a 384-dimension neural baseline using 48× less storage per node, with no GPU, neural training, or embedding API required.
Cite this work
@misc{barceloscosta2026sdm,
author = {Barcelos Costa, Cleber},
title = {Topology-Aware Binary {SDM} for Knowledge Graph Retrieval:
A Multi-Architecture Empirical Study with Neural Baseline
and Quantum Walk Analysis},
year = {2026},
doi = {10.5281/zenodo.19645323},
publisher = {Zenodo},
url = {https://doi.org/10.5281/zenodo.19645323},
license = {CC BY 4.0}
}
