A Research Run Is More Than an Answer
Feynman’s runtime and workflow now point to a more useful unit of analysis than the response printed at the end of a session. A question enters through a controlled Pi environment, moves through planning and source choices, becomes evidence and artifacts, and eventually meets a verification decision. The practical judgment this model supports is simple: a result can be present without being complete, organized without being verified, and inspectable without being scientifically correct.
That distinction is the hinge between Feynman as a collection of research commands and Feynman as a research system. The runtime explains where work happens. The workflows explain how a question moves. The source layer explains what kind of evidence can enter. The artifact and manifest layers explain what remains after the model stops. Taken together, they form a chain of states, not a single confidence score.
The research run as a linked state
Consider what happens when a researcher asks a broad question. The command does not jump directly from prompt to conclusion. Feynman first establishes a session and workspace, then selects a workflow scale. A narrow question may stay with the lead agent and a small set of targeted searches. A broad question may use bounded researcher delegation. That choice changes the expected evidence path before any synthesis exists.
The run then moves through source acquisition. Academic search, web retrieval, code inspection, dataset inspection, and domain-specific connectors can contribute different kinds of material. A paper might provide a method description; a code repository might clarify implementation; a dataset card might reveal a schema or access restriction; a release note might establish a changed behavior. These inputs are not interchangeable merely because they arrive as text. Their source type affects what they can support.
The next transition is from material to findings. Extraction identifies claims, methods, entities, or limitations that can be used in analysis. Synthesis connects those findings into a readable answer. The artifact layer preserves the output and, in richer workflows, the plan, score audit, graph, reproduction note, or provenance record around it. Verification then asks whether the important claims fit the evidence and whether the proposed or recorded checks actually occurred.
The transition structure matters because it tells the reader where to look when a result is weak. If the source path was narrow, the problem may be coverage. If a legal full-text route was unavailable, the problem may be access. If a model synthesized before delegated outputs were consumed, the problem may be workflow coordination. If an artifact exists without a completed check, the problem is verification. A single final paragraph cannot reveal those distinctions unless the system preserves them along the way.
Feynman’s typed research manifest formalizes this chain. Its versioned research-run contract has separate fields for the workflow, topic, status, research jobs, sources, papers, extracted entities, tool runs, artifacts, next actions, verification, and storage constraints. This separation is not bureaucratic decoration. It gives later tools a place to ask different questions: What jobs were attempted? Which sources entered? What artifacts were produced? What remains to be checked? What was intentionally not stored?