The Judgment Feynman Can Support, and the Judgment It Cannot Replace

Feynman is designed to support a research decision without becoming the authority that makes the decision for you. Its distinctive design is an inspectable loop: Pi supplies a controlled runtime, source-specific tools gather material, workflows shape the investigation, ranking or execution produces evidence and artifacts, and the workbench preserves lineage and checks. The practical judgment is conditional. The design is most evidently suited to cases where a question is broad, evidence is distributed, and the cost of losing the path from source to conclusion is high. It is less necessary when a narrow answer needs little durable context. In neither case does a successful Feynman run guarantee scientific correctness.

The distinction matters because the repository contains both automation and restraint. It can choose a research scale, route searches, construct a local paper graph, expose score components, record a plan, attach a producer to an artifact, and surface an unverified claim. Those are real mechanisms. The model, provider, source, and human still determine whether the resulting judgment is sound. Feynman’s contribution is to make that division visible enough to act on.

One model: research as a stateful decision system

The whole system can be understood as four connected layers. The first is the runtime layer. Feynman wraps Pi and passes it the session, prompts, extensions, model controls, workspace, and environment it needs. This establishes where the research loop runs and who owns the session machinery.

The second is the evidence layer. AlphaXiv, web search, Hugging Face inspection, and domain-specific connectors provide different routes into papers, repositories, datasets, documentation, and scientific records. The routes are not interchangeable: a paper’s metadata, a repository’s implementation, and a dataset’s schema answer different questions. Full-text resolution adds another distinction between a legal access candidate, a landing page, and a fetchable source-specific text route.

The third is the decision layer. Deep research decides whether a question is narrow enough for direct lead-owned work or broad enough for bounded delegation. PaperRank decides what to read first. A literature review maps consensus and disagreement. A replication or recipe workflow asks what would be needed to test or implement a reported result. Each decision is narrower than “the model knows the topic,” which is why each can have a different evidence threshold.

The fourth is the continuity layer. ResearchRun manifests, plans, execution records, artifact versions, snapshots, claims, provenance, and verification checks preserve what happened around the answer. The workbench makes those relationships navigable over time. A result can therefore be inspected backward toward its sources and actions, or forward toward the next check.

The layers form a loop rather than a straight line. A missing source can send the run back to search. A sensitive ranking can prompt calibration or closer reading. A failed reproduction can change the next action. A privacy boundary can force the system to retain less detail and ask the researcher to retrieve the source again. The loop’s quality is therefore determined not only by how much it automates, but by whether the states and transitions remain distinguishable.

This leads to the strongest defensible claim about Feynman: it is an evidence-oriented research operating model. It gives a research question an explicit route, a set of bounded intermediate objects, and a place for unresolved dependencies. That claim is stronger than saying it is merely a chat interface and weaker than saying it is an autonomous scientific authority. The repository supports the former. It does not establish the latter.

Choose the surface for the decision you actually have

Feynman’s surfaces make sense when matched to the decision, not when used as a checklist of features.