\\
polibio.ai
An open template for building a domain-specific agentic research system on Claude Code. It iterates hypotheses, builds datasets to the design, prepares fieldwork, and probes findings with theory and quantitative evidence — it is scaffolding you fill with your field's commitments, sources, and canon.
General-purpose models apply every theoretical tradition with equal weight, and they leave intact the oldest constraint of empirical work: the data dictates the study.
Models re-derive the analytical apparatus from scratch each time, weighting traditions equally. A discipline's commitments, moves, and canon never enter the loop — and so never shape the work.
Without machinery to build evidence to a design, the researcher bends the question to whatever statistics happen to exist. The study becomes a function of the dataset, not the inquiry.
Polibio.ai inverts both.
Discover and gather before interpreting. Theory before statistics. Fieldwork between C and D — not in the system, but in the researcher's hands.
A conjecture meets a networked evidence base — APIs and MCP — and is revised in minutes, not weeks.
The study specifies variables, units, and period; the system assembles them across sources with full provenance.
Triangulate preliminary evidence with literature — your library first — and the documentary record into an operational plan.
Apply your field's frameworks first, then quantitative tests. The system returns a diagnosis, never a verdict.
Polibio.ai finds, builds, prepares, and probes. What it discloses is scaffolding; what you bring is your field — its commitments, moves, sources, and canon.
Four analytical layers and their agents, the slash commands, the data toolkit, the provenance-logged DataPool, and the CLAUDE.md skeleton.
Your field's intellectual commitments, its analytical moves, its evidence sources, and its canon — the "interlocutor library." These are the three sections of CLAUDE.md you rewrite.
A set of connectors to common social-science sources. Every observation traces to the call that produced it. Usable standalone, even without Claude Code.
Everything lands in a local DuckDB DataPool whose schema enforces — by design — the conditions for empirical work that can be audited and reproduced.
The phases leave a trail of auditable artifacts — provenance-logged datasets, charts, a probe diagnosis. Below is the output of one worked example: Italy's industrial districts in the AI–robotics transition, the case the companion paper develops. The phases turned a lagging-adoption story into a question about absorption — and built the evidence to ask it well.
The pattern is general. The included example — institutional law & economics — is one instantiation, not the contribution. Three conditions apply.
Your field's analysis separates into discrete moves that chain. An agent can be given one move at a time and held to it.
Your sources are queryable — or at least machine-readable. If the data is in a building, no system substitutes for getting on a plane.
There is a defensible order in which the work proceeds — discovery before interpretation, theory before statistics — and you can name it.
A paper and a repository. Pick the one that matches how you read systems.
Polibio.ai — Agentic AI for Field Research and Inductive Theory in the Social Sciences.
Read on SSRN↗The four-layer scaffolding, the slash commands, the data toolkit, the DataPool, and the CLAUDE.md skeleton. Fork and fill.
Open on GitHub↗