From Evidence to Literature and Durable Artifacts

OpenResearch connects experiments to a broader research record in two directions: literature gives a new hypothesis context, and artifacts give a completed run a durable form that others can inspect. This chapter focuses on the boundary between discovery, interpretation, and presentation. The practical judgment is that a paper search can shape a question, and a report can preserve a result, but neither a retrieval snippet nor a polished figure can substitute for claim-level reading and recorded run evidence.

That boundary is essential in an agent-assisted workflow. An agent can find a promising paper, propose a variant, run an experiment, and write a persuasive summary without any one step being deliberately deceptive. The failure comes from collapsing different evidence roles. A search result is treated as if it were a verified finding. A report is treated as if it were the measurement. A figure is read without the configuration that produced it. OpenResearch's literature and artifact capabilities are most useful when they keep those roles connected but distinct.

Literature changes the question before it changes the code

OpenResearch includes discovery and paper-retrieval primitives for alphaXiv, OpenAlex, and bioRxiv. The literature workflow describes source-specific discovery, ranking, focused follow-up, and paper reading. A project can retain paper context, including an arXiv identifier or paper PDF, so a research effort can preserve where its starting question came from. These capabilities make literature part of the project boundary rather than an unrelated browser tab that disappears when an agent session ends.

The value of that context is not simply that it produces more citations. Prior work changes what a new experiment means. It can reveal that an apparently novel change has already been tested, that a baseline is not the field's usual control, that a metric has a known limitation, or that an implementation choice is tied to a condition the new run does not reproduce. It can also sharpen a hypothesis by showing which variable earlier work held fixed and which trade-off remained unresolved.

That use of literature is more disciplined than asking an agent for a generic “related work” paragraph. The relevant question is what the source changes about the experiment: its baseline, mechanism, comparison, data interpretation, or boundary. OpenResearch's project and experiment records can then hold the resulting context while the tree carries the new hypothesis. Literature informs the parent and question; it does not become evidence that the new branch has succeeded.

Discovery and reading are separate stages. Discovery services return candidates, metadata, and, depending on the source, snippets or reports that explain why a paper was found. Those outputs are useful for ranking and deciding what to inspect next. They are not enough to support a substantive claim about a method, result, limitation, or comparison. Claim-level synthesis requires reading the selected paper representation and preserving the source link that actually supports the statement.

The distinction becomes practical when a search returns several papers that use similar language. A candidate title may be relevant because it shares a method name, but the full text may reveal a different dataset, objective, or evaluation condition.