Agentic Research Workspace

A paper tells the story of a result. The research that produced it is a much larger object.

There are intermediate calculations, evolving questions, code versions, experimental setups, failed approaches and decisions about what to try next. Some become supplementary material. Much of the path between them remains scattered across notebooks, folders and conversations.

Agentic Research Workspace starts with a question: what if we preserved those relationships as research happened, then published a navigable representation of the work?

A PDF could remain a useful way to read the argument. It would no longer have to be the container for everything we need to inspect, reproduce or extend it.

A research pack would be a curated, versioned publication: the narrative together with the claims, evidence and artifacts that give it substance. It should answer not only “what did you find?” but “how did you get here, what supports this, and what can I reuse?”

The system that grows with a research project needs to retain more than the final pack. It may hold rough notes, competing hypotheses, abandoned experiments, unfinished derivations and private material.

Publishing should be a deliberate transformation: select the relevant records, pin their versions, check their dependencies, handle rights and redactions, and state what is missing. Capturing something must never automatically make it public.

The intended operating model is one project-wide coordinator agent. It can retrieve context across the authorized workspace, create focused tasks or subagents, and integrate their returns into durable project state. Continuity should survive the end of a chat.

That broad view does not grant unlimited authority. Scientific acceptance, external actions and publication remain separate decisions. The current mapper’s deterministic job coordinator is not this future workspace agent.

Before preserving research as it happens, we need to understand what already exists. The first implemented part reconstructs source-grounded structure from published literature.

The useful unit is not just a paper or a summary. It is an assertion, a reasoning step, a piece of evidence, and the relationships that make those parts intelligible.

Superdeterminism is the testing corpus, not the identity or domain limit of the project. It gives us a concrete collection of papers against which to test evidence routing, argument structure and cross-source navigation.

The pictured 52-paper baseline contains 6,872 records, 14,442 source-local links and 644 potential cross-source relationships. These are counts of retained output, not measures of completeness or correctness. Audit depth is uneven, and many sources remain provisional.

All four screenshots on this page show that private research viewer. They illustrate the direction; they are not a live demo of an application included in the alpha.

The first public release is an experimental literature-mapping baseline. Its easiest entry point is a model-free, source-checkout example: an invented source and manually authored dossier compile into six records and nine links, with zero model calls. It tests the record-to-graph path; it does not demonstrate live extraction.

The release is not a complete research IDE, hosted service or research-pack publisher. The pictured viewer, a self-contained package installation, and later long-form ingestion work are not included.

The next design milestone works backward from publication: build one small reference pack that a person can read, an agent can navigate, and someone else can use to trace a result to its inputs. That should tell us what the living workspace needs to capture.