When I describe execution records to scientists, the first question is almost always some version of "Are you watching my results?"

It's the right question. A standard that records what happens at the bench will only be adopted if scientists feel supported, not monitored. That can't just be a promise in a privacy policy. It has to be built into the format.

Record the process, never the results

An execution record should describe how an experiment was run: which steps, in what order, how long, with which reagents and instruments, and where it deviated. It should never contain experimental results, data, or conclusions. Those belong to the scientist, and they're governed by the lab's own rules and the publishing process.

Principles I'd propose for the standard

  • No results, by design. The schema has no place for experimental outcomes.
  • Identity is optional. A record can be useful without naming who ran it.
  • Consent before sharing. A record is published or deposited only when the scientist chooses to.
  • Aggregate before analyzing. Any analysis across many records should use anonymized, aggregated data.
  • Process locally where possible. Sensitive raw inputs, like audio, shouldn't have to leave the lab to become a record.

A design choice we made at Adept

When we built Adept's voice copilot, we knew academic institutions would, rightly, block anything that streamed lab audio to the cloud. So we built the app around an open-source speech-to-text model that runs on the device. Voice isn't transmitted off the device. It cost us some convenience, but it meant privacy wasn't a feature we asked people to trust. It was part of the architecture.

I think the standard should encourage the same thinking. The format should make the private path the easy path.

The harder question: lot data

Lot numbers are one of the most valuable things a record can hold. If many labs report that a specific lot is underperforming, everyone benefits from knowing sooner, including the manufacturer. But lot-level data raises fair concerns about how it's reported, verified, and interpreted. I don't have the full answer, and I don't think anyone should design it alone.