K-Dense BYOK: An Open-Source AI Research Assistant That Runs Locally

Summary

K-Dense BYOK is a free, locally runnable, MIT-licensed AI research assistant. By combining a hash-chained, append-only lab notebook with an observation log that the agent cannot write to, it ensures scientific claims are not merely asserted but verifiable.

Why it matters

As AI research tools become increasingly cloud-dependent and opaque, a locally managed open-source alternative addresses real problems. It tackles concerns regarding data sovereignty, long-term readability, and reproducibility. The 'observe rather than trust' log design offers a pragmatic and structural response to the well-documented issue of exaggerated claims by large language models (LLMs). Thanks to the flat, folder-based project model, results persist even after the tool itself becomes obsolete.

Highlights

• Hash-chained, append-only Living Lab Notebook; inputs are linked into a chain of argument and never deleted
• Observation log with no write access for the agent; targets the model's exaggerated claims at the architectural level
• Outperformed two managed platforms in terms of scientific quality and research execution across 20 interdisciplinary prompts

Critical take

The benchmark relies solely on 20 prompts and a comparison with two managed platforms. This competitor set is quite weak for claiming superiority. The authors acknowledge that the observation log has not yet captured the software environment. Consequently, the guarantee of 'verifiable claims' remains significantly incomplete. This study, self-reported by a single team and based on a fixed evaluation scale, raises questions about selection bias in both the prompts and the scoring.