The OpenResearch Bargain: Local Control, Remote Power, and the Limits of Reproducibility

OpenResearch is a local-first workspace for research agents and autoresearch, but its real subject is the relationship between speed and trust. This final synthesis asks what the product's full arrangement secures: a project and Git history own the source, an experiment tree preserves conditional decisions, isolated agent sessions protect active edits, a committed snapshot defines what crosses to compute, supervised logs preserve what happened, and artifacts carry the interpretation forward. The judgment is conditional. OpenResearch is a strong fit when a researcher values inspectable lineage and local control, but it is not a substitute for controlled environments, valid data, meaningful metrics, or scientific judgment.

The product's bargain is easy to misunderstand in either direction. It is not merely a convenient dashboard around a coding model. The local dashboard matters because it coordinates records that otherwise live in separate tools. But it is also not an autonomous laboratory whose status badges certify truth. Its contribution is narrower and more useful: it makes the path from an agent's proposed change to a claim more explicit, and it preserves enough of that path for a researcher to inspect where confidence should stop.

A local control plane around a moving compute boundary

The operating model can be stated as one sentence: OpenResearch keeps research identity and evidence coordination local while allowing a recorded source snapshot to cross into a chosen compute boundary and return as inspectable output. Every important component supports that sentence.

The project is the ownership anchor. It identifies the repository, baseline branch, run command, and optional paper or publication context. The experiment tree turns a change into a hypothesis with a parent, branch, description, and fixed comparison contract. An agent session works in a private worktree, so parallel exploration does not require several processes to edit one checkout. Shared Git refs let sessions discover one another's branches without giving them permission to claim the same working directory.

The local store and dashboard connect those surfaces. The dashboard can show projects, experiments, runs, logs, diffs, commits, files, artifacts, chats, and events. The store preserves relationships among them, while the repository remains the source of branch and commit identity. A playbook gives the harness project context and operating procedures, but it does not become the only record of what the project is. A local model can change where inference occurs, but it does not change the branch, run, or evidence contract.

A completed experiment then moves through a second boundary. OpenResearch resolves the branch's recorded commit, creates a content-addressed source archive, and gives the selected adapter a durable snapshot. The backend may be the local machine, a user-owned SSH host, a scheduler or serverless provider, a Tinker target, or an ephemeral OpenResearch box. The adapter knows how to preflight, transport, submit, poll, stream logs, cancel, and clean up. The local experiment still owns the research identity and the run still records the source revision.

Evidence returns through the run log and result records. The supervisor preserves observation across a detached process, a remote handle, or a dashboard restart. The researcher then interprets the output, connects it to literature, and stores a report, figure, dataset, or PDF in the artifact surface. The next hypothesis is not supposed to begin from a vague summary; it descends from what the previous evidence actually established.

flowchart LR subgraph Local["Local control and evidence"] project["Project and Git baseline"] --> tree["Experiment tree"] tree --> session["Agent session and private worktree"] session --> commit["Committed experiment source"] commit --> snapshot["Immutable source snapshot"] log["Supervised run log"] --> artifact["Durable artifact"] log --> next["Next decision"] next --> tree end snapshot --> compute["Chosen compute boundary"] compute --> log limits["Security, environment, data, and scientific limits"] -.-> snapshot limits -.-> compute limits -.-> log limits -.-> artifact

The diagram also shows what the product does not own. The arrows among project, tree, session, snapshot, compute, log, and artifact are system relationships. The surrounding limits are conditions that can weaken a claim even when every arrow is recorded. Local control does not make remote infrastructure trustworthy by itself.