sddzlsc/sciforge is a local-first research agent harness that covers deep learning research from literature search and deep reading through idea generation, GPU experiments, paper writing, and publication figure creation. The latest release is v0.2.0. The provided record does not state what changed in this version.
The project organizes the research flow with Skills, Tools, Pipelines, and a local web console. It is DeepSeek-based and supports a user-specified OpenAI-compatible LLM as well as remote GPU work over SSH. Idea cards produced during literature search carry over into the experiment stage, and project state is shared across research-state.yaml, findings.md, and research-log.md. GPU jobs are connected through SSH with tmux or Slurm, and the local console binds to 127.0.0.1. LLM call logs can be audited, and experiment figures are designed to be cross-checked against results/ files. It is written in Python and uses the MIT license.
The recorded design alone, however, cannot confirm the accuracy of research results or the quality of experiments. No material showing actual verification quality was provided, and most of the basis is the README's own description. The structure that goes beyond literature and conversation to actually running experiments is worth referring to, but without separate evidence it is premature to judge the reliability of the outputs.