decepticons is a backend-neutral research kernel that provides predictive primitives for building machine learning systems. It is designed to offer a collection of core mechanisms—such as substrates, memory, gating, routing, and readouts—that downstream systems can combine into trained models without modifying or forking the kernel itself. The platform enforces a clear boundary by including only mechanisms that can be generalized and reused across multiple descendant systems, while excluding project-specific policies, training frameworks, or orchestration components.
The kernel features substrate dynamics, memory primitives, controller summaries, pathway gates, routing, modulation, feature views, and reusable readouts. It supports a variety of substrate modes, including frozen, learnable decays, learnable mixing, learned recurrence, and gated retention, as well as memory attachments like n-gram priors and exact-history caches. The architecture is modular, with substrate, memory, control, routing, and view modules, and supports configuration-driven dispatch for substrates. The kernel also includes lightweight runtime evaluation, artifact accounting, export helpers, and deterministic substrate builders. Causality verification is a core part of the platform, with every substrate mode checked for future-leak under perturbation.
11 or higher. The core kernel depends only on numpy, but optional backends for PyTorch and Apple MLX can be installed to add additional functionality, such as the PyTorch CausalBankModel or the MLX backend for Apple Silicon. Installation is available via PyPI, and a command-line interface is provided for fitting and sampling predictive models over UTF-8 corpora. The kernel is distributed under the MIT license, and its source code and documentation are available on GitHub.
This tool is aimed at developers and researchers building custom machine learning architectures who require reusable, backend-neutral primitives for predictive modeling. It does not include a training framework, fleet orchestration, policy logic, or audit packaging, focusing instead on providing a stable foundation of mechanisms that can be integrated into a variety of downstream projects.
In the Frameworks & SDKs space, decepticons takes a focused approach. It focuses on providing reusable predictive primitives for building machine learning systems without forking core kernels. decepticons is an open-source project aimed at machine learning developers. The project is open source (MIT). It runs on the command line.
decepticons first shipped in 2026. The project is developed in the open on GitHub with 94 commits in the last 90 days. Among its 5 catalogued features are predictive primitives, memory modules, and gating and routing. decepticons is currently in beta.
Latest indexed changes and source events
Other apps tracked under the same category.
