On August 27, 2026, Anthropic opened a research preview of the Model Hardware Standard, a shared specification that lets AI agents operate lab instruments like microscopes, liquid handlers, and robotic arms. When it came out, I wrote that we need the same thing for the human side of the bench.
Since then I've talked with people who run research identifier registries and people who lead lab automation programs. I'm now convinced someone has to start it. So I'm going to.
The half of the lab no standard covers
If you run a liquid handler, your protocol is a file. It has a defined structure, it can be versioned and shared, and a machine can execute it the same way twice.
If you run a protocol by hand, which is still how most biology gets done, your protocol is a PDF, a notebook page, and your memory. The methods section of your paper tells readers what you meant to do. It doesn't tell them that step 4 ran 14 minutes instead of 10, which lot was in the tube, or that you repeated step 7 because the pellet didn't look right.
I've seen what this costs
I've spent about ten years building genomics tools, at Qiagen, at Bio-Rad, and then as a co-founder of Fluent BioSciences. At Fluent we did what every tools company does. We wrote a 60-page user guide, recorded a video tutorial, and handed the product to customers. We had a couple of field application scientists to support everyone.
When a run failed, the support case usually started with about three sentences from a frustrated scientist. Reconstructing what actually happened at the bench could take weeks of back and forth. Nobody was doing anything wrong. There was simply no record of the execution to look at.
The handoff is about to get harder
Labs are moving more methods from people to instruments. Someone has to translate a handwritten protocol, full of gaps the expert fills in without thinking, into something a machine can run.
The instrument side now has a shared language taking shape. The human side has nothing to hand over. That translation will only ever be as good as our record of what the expert actually did.
What I'm proposing
An open, vendor-neutral standard for protocol execution records. A record would capture three things:
Protocol execution record
- Which protocol, and which version of it, was followed
- Which reagents, lots, and instruments were used
- What happened at each step: order, timing, pauses, repeats, and deviations
It would reuse identifiers the research community has already built, like RRIDs for reagents and PIDINST for instruments, rather than reinventing them. It would never capture experimental results. It is not an electronic lab notebook, and it isn't meant to replace protocol repositories. It is a structured, computer-readable account of how a run was executed.
Why this has to be open
No one wants a single company to announce a standard and expect everyone to adopt it. I would like to form a working group within the Research Data Alliance so the standard can be built in the open, with academics, core facilities, automation teams, repositories, and manufacturers at the table.
Full disclosure: I'm the founder of Adept Scientific, where we build a voice copilot that guides scientists through protocols and produces execution records. That is exactly why this standard can't belong to us. Adept will be one implementer among many, and I want it held to the same spec as everyone else.
Coming up in this series
- Your methods section records intent, not execution
- When the expert walks out the door
- Don't reinvent the wheel: building on RRIDs and PIDINST
- What a protocol execution record should never contain
- Help us write version 0.1