Checkpoint: The Control Surface of a Research Session
OpenResearch is easiest to judge before a run starts, when the project, experiment branch, agent session, and evidence plan can still be inspected together. This checkpoint connects those already-established pieces into one practical question: where do the source, hypothesis, edits, execution contract, and intended evidence live? The answer gives a reader a preflight test for an agentic research action—and also shows which questions no amount of local orchestration can settle.
The point is not to repeat the parts separately. A project without an experiment is only a registered source. An experiment without a controlled session is a branch with an intention but an uncertain editing boundary. A session without a recorded branch is a conversation that may never become a research result. And a launch without an evidence plan can produce a terminal status without producing a defensible measurement. OpenResearch's design becomes coherent when these records are treated as a chain with distinct owners.
One prepared hypothesis, several owners
Consider a researcher who wants an agent to test one change in a model or analysis script. The project owns the repository and baseline branch. That is the answer to “which source does this work belong to?” The experiment tree owns the hypothesis and its parent. That is the answer to “what earlier decision does this change build on?” The session owns the private worktree and the interaction context. That is the answer to “where did the agent make and inspect the edits?” The fixed run command owns the intended evaluation recipe. That is the answer to “how will this branch be compared?” The eventual run record and log will own the source commit and observed output.
These owners are connected but not interchangeable. The project can contain several experiments. A single session can discuss an idea before creating a node, or can create several branches over time. A branch can outlive the session that edited it. A run can continue after the dashboard turn that launched it has ended. The system's integrity comes from preserving these distinctions while linking them, not from pretending that one object is the whole research process.
The connections also make different failures diagnosable. If the repository path is wrong, the project boundary is wrong. If a branch has no clear parent or description, the hypothesis boundary is weak. If two sessions edit one branch, the worktree boundary has failed or been bypassed. If the command differs between supposedly comparable nodes, the execution contract has changed. If a run is complete but its log lacks the effective configuration or final metric, the evidence boundary is incomplete. These are not the same problem and should not be repaired in the same way.
Imagine the same proposed change viewed through each owner. The project asks whether the source belongs to this research effort. The experiment asks whether the change builds on a known parent. The session asks whether the agent is editing the branch it is meant to edit. The harness asks whether the agent has current project context and the right operating procedures.