Neuro-Symbolic Systems
Neural where necessary.
Symbolic where possible.
We build neuro-symbolic systems that combine the flexibility of modern AI with structured reasoning, verification, and deterministic computation — so intelligence gets cheaper and more reliable as it matures.
01 — Thesis
Intelligence is more than inference.
Most systems today route every operation through a large model. That works — until reliability, cost, and inspectability start to matter. Serious systems need more than a probability distribution over answers.
R.1
Reliability
Important operations should be inspectable and verifiable — not just plausible.
R.2
Efficiency
Expensive inference should not be spent on work that can be represented procedurally or executed deterministically.
R.3
Structure
Reasoning benefits from explicit state, relationships, constraints, memory, and tools.
R.4
Adaptation
Neural intelligence handles ambiguity and novelty. Structured systems preserve repeatable competence.
02 — Architecture
One system. Two kinds of reasoning.
Neural components interpret, generalize, and handle what has never been seen before. Symbolic components constrain, verify, and execute what is already known. The architecture decides which one does the work.
Neural Intelligence
Flexible reasoning for the unknown.
— Interpretation
— Ambiguity
— Generalization
— Novel reasoning
— Language
Symbolic Systems
Structure and certainty for the known.
— Constraints
— Tools and programs
— Verification
— Deterministic execution
— Structured memory
The Result
Neuro-Symbolic Intelligence
Adaptable reasoning · Inspectability · Reliability · Lower inference cost · Reusable procedures
03 — Reliability
Reliability is an architecture problem.
A capable system should not merely generate answers. It should have mechanisms to check, constrain, and verify its own work — and produce a trace that another engineer can read.
Execution trace
run 0413
01
Interpret request
verified
02
Retrieve constraints from memory
verified
03
Execute known procedure
verified
04
Check result against constraints
verified
04 — Efficiency
Use intelligence where intelligence is needed.
Expensive reasoning should become cheaper execution over time. As the system encounters a problem again, it moves the work down the stack.
Model reasoning
Full neural reasoning where ambiguity genuinely requires it.
Repeated competence
Structured procedure
Reasoning that has proven itself becomes a reusable, inspectable procedure.
Known operation
Symbolic execution
Deterministic work runs deterministically. No model in the loop.
05 — Principles
How we build.
01
Structure over prompt chains
Complex behavior is engineered as explicit systems — state, constraints, procedures — not as fragile sequences of prompts.
02
Verification over assumption
Every consequential action is checked against explicit constraints before it runs, and leaves a trace afterward.
03
Procedures over repeated reasoning
When the system solves a problem well, that competence is captured and reused rather than re-derived every time.
04
Models as components
A model is one part of the system, not the system itself. Orchestration, memory, and tools carry the rest.
05
Escalate intelligence only when necessary
Routing sends known operations to deterministic execution and reserves neural reasoning for genuine novelty.
06
Observable by design
State, decisions, and provenance are visible to the engineers who operate the system — by default, not by exception.
06 — Research
Technical work
Example entries. Placeholders for future publications, systems, and releases.
Efficiency
Procedural distillation: converting repeated reasoning into verified programs
Example entry. A method for observing recurring model reasoning patterns and compiling them into deterministic procedures with explicit correctness checks, reducing inference cost on repeated tasks.
Read
Memory
Structured working memory for long-horizon agent tasks
Example entry. A typed, inspectable memory substrate that lets agents maintain state, constraints, and provenance across extended task sequences without re-deriving context.
Read
Reliability
Verification layers for tool-using systems
Example entry. An architecture for checking model-proposed actions against symbolic constraints before execution, with traces that make every decision auditable after the fact.
Read
Orchestration
Adaptive escalation policies: deciding when a model is necessary
Example entry. Learning routing policies that send novel problems to neural reasoning and known operations to symbolic execution, measured by cost, latency, and error tolerance.
Read
Synthesis
Neuro-symbolic program synthesis under explicit constraints
Example entry. Combining language-model generalization with constraint solvers to synthesize programs that are correct by construction rather than correct by sampling.
Read
The next generation of AI will not be one model answering every question. It will be systems that know when to reason, when to retrieve, when to verify — and when not to use a model at all.
Ornyx — on the architecture of intelligence
07 — Company
We build systems for problems that demand more from AI.
Ornyx exists to make sophisticated intelligent systems practical to engineer — for teams whose problems cannot tolerate guesswork. We work with researchers, engineers, and organizations that need intelligence they can inspect, afford, and trust.
More capable
C.1
More trustworthy
C.2
More economical
C.3
More understandable
C.4
Easier to engineer
C.5
Build what comes after the model.
We are open to conversations with researchers, engineers, and technical partners.