Verified knowledge kernels
Hornbeam
Knowledge your agents can stand on — every answer cites checked evidence.
Hornbeam turns what language models extract into versioned knowledge kernels: a schema, an evidence graph that records where each fact came from, and typed tasks. Kernels are checked by deterministic gates and optional formal solvers, repaired in bounded steps, and frozen once accepted — so answers can be audited and replayed months later.
What it does
Model output is untrusted data
LLM output is parsed and checked, never executed as code.
Evidence on every answer
Each answer carries citations, and each citation is checked as supported or not.
Formal checks when it matters
Optional OWL reasoning, SMT, Datalog and answer-set solvers validate a kernel's logic.
Deterministic promotion gates
Kernels are promoted or rejected by gates in code — not by another model's opinion.
Bounded repair
Failures are repaired in typed, bounded steps with regressions blocked.
Immutable and replayable
Accepted kernels are content-addressed, portable and replayable.
Domain packs
New domains plug in without touching the core.
Evidence
- 5,000+automated tests
- Stricttype checking across the core
- Zeroruntime dependencies in the core
Services around it
Production agentic systems
From prototype to production: we design and build agentic systems whose work is verified independently, bounded by budgets and approved by people before it lands.
Formal assurance for AI systems
For the rules that must never be broken, tests are not enough. We specify safety properties formally, prove them, and check that production code conforms.