An open-source AI research assistant called K-Dense BYOK keeps a hash-chained lab notebook that logs what the AI actually did, not what it claims it did.
K-Dense BYOK is free, MIT-licensed software that runs on a researcher's own machine. Researchers supply their own model access - the BYOK stands for bring your own keys - and the app supplies scientific procedure libraries, workflow templates, data catalogs, and reviewer and writer roles the agent can hand work to. Its Living Lab Notebook only appends entries, and the record of what happened comes from watching the agent's actions rather than the agent's own account, through a log the agent has no tool to write to. In a twenty-prompt interdisciplinary benchmark scored against a rubric fixed in advance, K-Dense BYOK beat two managed AI research platforms on scientific quality and research execution, according to its creators.
The target failure here is model overclaiming - agents asserting they ran code or produced a result when they didn't. Watching behavior instead of trusting self-reports is a genuine fix for that. But the paper's own account shows where the fix stops: the environment details and the commands said to regenerate K-Dense's results were files the agent itself wrote, not entries in that tamper-proof observed log, because the log doesn't yet capture the software environment at all.
So the one feature meant to make claims checkable doesn't yet check the most basic reproducibility question: what software ran this.