What it is
Lab Assistant is the second half of a pair with Science Machine. Science Machine finds a research question that looks open; Lab Assistant is meant to carry that question to a result someone else can check. Today it is two written documents. One is a design for a desktop assistant. The other, the Checkable Science protocol, is a set of numbered rules for people without research credentials who work with an AI agent and still want their work to hold up.
How it works
As designed:
- Start with a question handed over from Science Machine.
- Talk or type to the assistant; it keeps a running lab log.
- Write the question so it could turn out wrong, and freeze the plan before any data exists.
- Take readings straight from instruments through a tool connection, so numbers are never retyped.
- Package the plan, the log, the data, the code and the write-up for a reviewer.
The protocol's rules apply even without the software: log every search for prior work, say how much of each source you actually read, and record whether a person or an agent did each step.
What exists today
- A design document (first draft)
- The Checkable Science protocol (first draft)
- Worked material from one real question: a dataset check, a methods study pack, a dossier and a task list
- A small human labeling task with printable sheets
- Two podcast episodes about that question, made with PodForge
What isn't there yet
- No software at all; the assistant, the log and the instrument hookups are all plans
- The protocol has not been tested on a complete project and expects revision after the first one
- No templates or forms that a newcomer could pick up and fill in
- No public version of the worked example
Under the hood
- Plain Markdown kept in an Obsidian notebook, under version control
- Shares Science Machine's core rule: code produces anything countable; the AI never supplies a measurement, timestamp, id or statistic
- Designed to plug into the same agent tooling Science Machine uses