Recursive Discovery
A compact system for recursive mathematical and empirical scientific discovery.
Overview
Recursive Discovery is a research system for scientific problems that need both mathematical reasoning and empirical evidence. A model chooses what to investigate; external tools run formal checks and experimental protocols; a persistent graph records the proposals, observations and relationships between them.
A result can also become a representation, a named concept, or a callable instrument that later work inherits, so the system accumulates reusable resources as it runs.
Capabilities
- SQLite artifact graph with content-addressed blob storage
- HMAC-authenticated execution records with inputs, outcomes and provenance
- Mathematical checker adapters for Python, SymPy, Lean, Z3 and cvc5
- Reproducible empirical laboratory with seeded runs and statistical summaries
- Context compiler with authority labels and graph-proximity retrieval
- Reusable instruments, including model-defined expression instruments
- Sealed prospective evaluator with a local vault and a service adapter
- Literature search and source reading across arXiv, OpenAlex and Crossref
The kernel
The kernel is the component that should remain understandable as models and instruments become more capable.
| Element | Responsibility |
|---|---|
| Artifact | Content-derived identity and typed references for proposals, results, instruments and decisions |
| Ledger | Retains scientific state and its lineage, using SQLite and content-addressed files |
| Execution kernel | Runs a declared tool, captures its outcome and provenance, and signs the record |
| Frontier | Derives pending work from missing records: a claim needs a check, an experiment needs a run |
| Promotion | Records an abstraction as a basis for a subsequent research world |
Scope
Version 1.0.0 provides the research machinery. A user-supplied model bridge connects the model provider.
The default process worker requires deployment isolation; execution is predominantly serial.