Purpose and approach
Thoughtlines is an interactive philosophy atlas and an experiment in sustained, AI-assisted research. Its aim is to turn individual investigations into a reusable body of knowledge that people can inspect, navigate and develop over time.
Reviewed 14 October 2026 · Prototype documentationA useful description is a mixed-initiative research environment: people set questions and priorities; software maintains records and repeatable workflows; a language model contributes interpretation, development and writing. This is our characterization of the project, informed by the related approaches below.
Product and methodological goals
A navigable history of ideas
The primary product connects thinkers, ideas, works and specific treatments of ideas. It should help readers investigate how a problem changes, compare accounts, follow supported historical relationships, and discover connections that deserve closer examination.
Chronological views and graph-linked stories provide different ways into this material. A story offers an authored route; the Explorer permits readers to pursue alternatives and inspect the underlying accounts and sources.
A reusable research method
A second goal is to evaluate how language models can contribute to the gradual construction of a knowledge resource. Research should produce persistent, structured outputs: treatments, source references, dates, relationships, reusable drafts and recorded decisions. Subsequent investigations should benefit from that work rather than reconstruct it from a conversation.
The software itself also develops through collaboration with Codex. Improvements to the model, validators, workflow and interface are versioned separately from additions to philosophical content.
Division of work
| Participant or layer | Responsibilities | Durable result |
|---|---|---|
| Human direction | Choose questions, scope, priorities and editorial standards; review outcomes and resolve consequential ambiguity. | Collection scope, accepted design decisions and editorial assessments. |
| Deterministic software | Identify reusable records, traverse the graph, acquire bounded sources, validate structure, save outputs and calculate coverage. | Versioned records, source captures, validation results and run state. |
| Codex as researcher | Interpret acquired text, summarize substantive accounts, identify components and assess possible relationships within an agreed scope. | Structured research packets with evidence, dating and limitations. |
| Codex as developer | Implement and test changes to the application and its repeatable research tools. | Reviewable code and documentation changes in version control. |
| Human or AI-assisted author | Select a route through existing research, compose an article and review its claims and transitions. | An editable story, scene plan and saved companion graph. |
These roles have different standards of verification. A schema check can establish that a citation and date are present; it cannot establish that an interpretation is philosophically adequate. Source inspection and editorial judgment remain necessary.
How knowledge guides further work
- Define a question. A collection identifies seed records and a research scope.
- Consult accumulated knowledge. Graph connections suggest relevant pairs. Research history identifies existing treatments, reusable drafts and earlier searches that found no support within their stated scope.
- Carry out bounded research. Scripts prepare and validate the work; Codex interprets the selected material.
- Preserve the outcome. Validated records enter the graph or draft library, while decisions and exceptions update the research history.
- Explore and communicate. Coverage reports, the Explorer and stories expose the results and help formulate the next question.
Model reasoning takes place within this cycle, but the conversation is not the sole record of progress. Saved artifacts allow work to resume, be checked and be reused across overlapping collections. Browsing and reading published stories use those saved results without a live model call.
Scope and boundaries
InPhO provides an initial structure for discovery. Local extensions add topics, people, works and treatments beyond that structure, with their origin retained in the data. The Stanford Encyclopedia of Philosophy (SEP) and Internet Encyclopedia of Philosophy (IEP) provide the reference base for broad initialization; wider scholarly and primary sources can support deeper research under an explicit source policy.
The historical scope is open. Hegelian attention to tension and transformation, and Deleuzian attention to difference, repetition and branching, inform possible lines of inquiry across the history of ideas. They do not restrict the atlas to either philosopher’s corpus or establish relationships on their own.
The current product supports saved research, bounded workflows, editorial review and two story prototypes. A single unattended research controller, automatic story generation from a user-facing form, and deployment into an external Quarto site remain development goals.
Coverage is selective. Graph proximity and shared topics support discovery, while historical influence requires appropriate evidence. A negative research decision means that the assessed sources did not support an account under that search policy; it does not establish that no such account exists.
How progress should be evaluated
The method should be assessed alongside the interface. Relevant criteria include source support, useful detail, date quality, identity consistency, avoided duplicate research, reusable outputs per source, time and usage, and the ease of correcting an account.
For readers, the central questions are whether the atlas clarifies a topic, supports productive comparisons and makes further reading accessible. More nodes or more generated prose are useful only when they improve those outcomes.