Writing
Notes from the bench
Protocol execution, reproducibility, and what changes when every run leaves a record.
The archive
10 pieces2026
-
The capital life science tools companies ship out the door
Manufacturers invest in two kinds of capital. The third kind, created every time a customer runs the product, is the one they throw away.
-
Even a verified robot needs an execution record
Google's VERID shows that even an automated lab needs a record of how each run actually went. The bench needs one too.
-
The tool makers have to learn too
A continuously learning lab needs two loops: one for the scientist at the bench, and one for the people who build the tools.
-
Your customers are asking AI how to run your kit. Whose answer are they getting?
Scientists already use general AI at work. When a chatbot fills a gap in your user guide, your product inherits the result.
-
Help us write version 0.1
I'm forming a Research Data Alliance working group to write version 0.1 of an open standard for protocol execution records. Here's why now, and who we need.
-
What a protocol execution record should never contain
Scientists will only adopt a protocol execution standard if they feel supported, not monitored. Here is where I think the lines should be.
-
Don't reinvent the wheel: building a protocol execution standard on RRIDs and PIDINST
RRIDs already identify reagents and PIDINST already identifies instruments. A protocol execution standard only needs to define what is genuinely new.
-
When the expert walks out the door
The protocol stays in the shared drive. The judgment that makes it work leaves with the person.
-
Your methods section records intent, not execution
Papers say what we meant to do. Results depend on what we actually did at the bench.
-
We standardized the robots. Now let's standardize the humans.
Instruments now have a shared standard for AI agents. The human side of the bench still has nothing.