Jacobian's documentation follows the Diátaxis framework, organized by what the reader is trying to do. Start with a tutorial when learning the system, use a how-to guide for a specific task, consult reference material for exact contracts, and read the explanations for design rationale.
Jacobian is a toolbox of atomic math tools for AI agents: find them with
math.find, run them with math.run, get mathematical results, and
compose those values across turns. Checker tools are optional additional
catalog IDs—separate from producers. Catalog entries are often still called
capabilities in the API. The product model
and Search and execute define
the contract.
Jacobian is pre-stable. The product, architecture, operation-library, and tool documents define the current contract; the installed catalog defines which operations are available in a particular checkout. Evaluations guide portfolio quality and do not grant formal authority.
These documents define the current product contract:
| Question | Document | Status |
|---|---|---|
| What is Jacobian? | Product model | Product and ownership model |
| How is it structured? | Architecture | Dependencies and trust boundaries |
| What does MCP expose? | Tool surface | Fixed MCP projection |
| What operations are installed now? | Runtime capability://catalog |
Current installation-specific inventory |
| What work is open? | GitHub issues (e.g. architecture epics) | Implementation priorities live in issues, not a parallel goals doc |
Tutorials are guided learning paths. They assume no prior Jacobian experience and build toward a complete result.
- Compute and independently check a determinant shows the boundary between an exact producer and a separate checker.
- Retrieve a Lean theorem and check a proof composes computed declaration retrieval with independent Lean replay.
How-to guides assume you already understand Jacobian's basic model and need to complete a specific task.
- Discover, invoke, and check domain math tools
- Configure an agent from a source checkout
- Install native and formal providers
- Troubleshoot Z3 installation on macOS
- Run the MCP visibility evaluation
- Deploy the remote MCP server
- Author a Harbor benchmark task
- Run agent evaluations
Reference documents define exact interfaces, records, gates, and test expectations.
Cross-cutting references:
- Tool surface — MCP resources, tools, and invocation contracts
- Domain operation library — built-in producer, bounded-search, artifact, and exact-replay contracts
- Native Python API — supported native-value modules
- Provider runtime — backend availability, compatibility, and identity
- Persistent state format — supported migration floor and fresh-store transition
- Testing strategy — validation layers, commands, and CI responsibilities
Domain-owned references: Capability references grouped by owning domain (graphs, matrix, polynomial, Lean, SAT/SMT, finite math, number theory, linear algebra, topology, geometry). Adding an operation or provider does not require editing a central list; each domain owns its own subdirectory.
Evaluation references: Benchmark contracts and evaluation methods — Harbor task contracts, dataset inventory, validation gates, model observations, performance measurement, and regression policy.
Reference scenarios: Worked cases — mathematical scenario catalog and certified-homology case.
Use the runtime capability://catalog and math.find for the
installed capability inventory and exact operation schemas.
- Product model — what the product is
- Architecture — host shape, search/execute, ownership and durable execution
Do not add parallel “direction”, “goals”, or portfolio-planning novels under
explanation/. Product intent lives in those two documents; open work lives
in GitHub issues.
Read CONTRIBUTING.md before changing code or public documentation.
Harbor benchmark authoring and verifier work uses the repository-local
harbor-benchmarks skill.
Source-grounded reliability probes based on recently resolved conjectures use
recent-conjecture-evaluations.
For hosted operation, follow
Deploy the remote MCP server; ignored tmp/
records are host evidence, not source of truth.
When adding a document, place it by reader need:
tutorials/— guided learninghow-to/— one taskreference/— contracts and lookupexplanation/— only product model and architecture unless a feature needs a dedicated operational reference that does not fit architecture
Do not mix product intent with supported release behavior. Concrete work lives in GitHub issues.