Scientific work exposes a weakness hidden by ordinary knowledge tasks: producing an answer and producing a defensible result are different things.
An agent may write a protocol, schedule an instrument, analyze a measurement, and draft a conclusion. Unless those actions remain connected to their source observations and human decisions, the final document is only a story about the experiment—not the experiment’s record.
The unit of work is the evidence chain
A laboratory run begins before an instrument starts. It includes the hypothesis, sample identity, protocol version, operator, environmental conditions, calibration state, exceptions, and authority to proceed. Every downstream claim inherits uncertainty from that chain.
Treating these as attachments added after the work creates gaps exactly where scrutiny matters. The system should capture them as first-class objects while the work unfolds.
Observation and inference must remain separate
An instrument reports a value. An agent may infer that the value supports a hypothesis. Those are not the same event, and a trustworthy system should never collapse them into one confident sentence.
The record must distinguish raw observation, transformation, interpretation, uncertainty, and approval. A reviewer should be able to move backward from the conclusion to the exact state of the world that produced it.
Humans define the consequential boundary
Automation can prepare protocols, coordinate runs, detect anomalies, and assemble evidence. Humans still decide which procedures are authorized, when an anomaly demands intervention, and whether a conclusion is acceptable for its intended use.
That boundary should be explicit in the workflow rather than implied by a policy document. Approval is not a chat message; it is a signed change in what the agent may do next.
Reproducibility becomes an operating property
When context, execution, and evidence share one record, reproducibility is no longer a reconstruction exercise performed months later. The system already knows which protocol ran, which machine acted, what changed, and who accepted the result.
This does not guarantee correct science. It makes the path to a claim visible enough to challenge, repeat, and improve.
For scientific agents, provenance cannot be an export format. It must be the shape of the work from the moment intent becomes action.