A scientific agent can describe code it might have run, summarize a repository it might not have opened, or report a numerical result without preserving the operation that produced it. The answer may even be correct. Correctness alone does not establish an inspectable execution.
The minimum execution record
A useful computational receipt should identify enough of the run for another qualified person to understand what happened and, when licensing and data policy permit, attempt it again.
- QuestionThe scientific task and acceptance criteria.
- InputsIdentified data, versions, checksums, or approved references.
- OperationThe code, workflow, query, or tool invocation actually executed.
- EnvironmentRelevant software, dependencies, infrastructure requirements, and configuration.
- OutputsProduced artifacts, logs, measurements, and checks.
- ExceptionsFailures, substitutions, retries, omissions, and manual interventions.
- AuthorityWho approved execution, reviewed the result, and may release it.
Reproducible does not mean universally portable
A run may depend on controlled data, licensed software, specialized instruments, or infrastructure unavailable to another team. That limitation should be stated rather than hidden behind a “reproducible” badge.
Different evidence states are useful. A method may be documented but not independently run. A computational path may be reproduced on equivalent inputs but not on the restricted original. A result may be reviewed without being public. Precision about the state is more trustworthy than one broad label.
Compare artifacts, not only answers
Two agents may return the same final statement through incompatible or invalid paths. They may also return different answers because of a meaningful difference in inputs, dependencies, or assumptions. Comparing only prose hides that information.
The object of review is not merely the answer. It is the chain from question to evidence.
A strong evaluation compares manifests, executed operations, logs, outputs, failure behavior, and scientific interpretation. It also records where comparison is impossible.
Failure records prevent false confidence
A timed-out process, missing dependency, malformed input, or failed quality check is not editorial clutter. It changes what the result can support. A system that silently repairs or retries an operation should record the intervention.
Negative execution evidence can be valuable even when no scientific conclusion follows. It can identify brittle methods, unsupported environments, hidden dependencies, and conditions another team should avoid.
From article to research capsule
When permissions allow, a Science Agent Daily item can point to a compact research capsule containing the approved context, method, environment, expected checks, outputs, limitations, and provenance. The capsule is not a promise that every reader can run restricted work. It is a structured account of what is available and under which conditions.
The publication should distinguish “described,” “executed,” “independently reproduced,” and “scientifically reviewed.” Readers and agents can then decide what the evidence actually permits them to claim.
Editorial basis: the LRN publication contract that separates claims, observations, execution, validation, and release. Concrete reproduction reports will identify their own authoritative sources and receipts.