New event or task
Shared facts & approved rules
Our approach
Make the basis of
a decision visible.
Neural models can propose useful actions. Their outputs alone do not show which facts and rules justify a decision.
meander makes that basis explicit. It connects a shared knowledge graph, formal reasoning, and source history so an agent can check a proposed action and retain a record of how the conclusion was reached.
meander in action
Continuous reasoning
& action evaluation.
An agent uses meander to work out what follows from the current facts and what evidence it still needs. New information can change the next step.
For a proposal such as “Refund €70,” meander checks the requirements and explains the result. The application decides whether to execute.
Reviewed cases become examples for new rule proposals. People test and approve changes before subsequent decisions use them.
Update the facts
Add new evidence, retain its source
Reason about a case
Derive conclusions & identify missing facts
Evaluate the action
Determine whether requirements are met
Action or review
The system executes or asks a human
Record outcome
Feed actual results into the next cycle
World model
Shared facts
& approved rules
How meander learns and reasons
Learning from cases.
Explicit, inspectable rules.
meander combines a company’s policies with rules it learns from cases. Inductive logic programming and predicate invention produce concepts and candidate rules that can be tested and inspected.
Examples & counterexamples
Reviewed cases and verified outcomes
Inductive logic programming
Predicate invention
New concepts
Explicitly defined
General rules
Apply across cases
Specific rules
Obligations / permissions
Test candidate rules
Compare with verified outcomes
Review & approve
People approve changes before subsequent decisions use them.
Built on FactGraph
Knowledge, reasoning,
and history belong together.
A graph stores the relationships. Logic derives conclusions. Provenance connects them to evidence. Together, they let you inspect the basis of a decision at the time it was made.
Knowledge over time
Reconstruct what information was available at a given moment and when it was valid. New facts do not erase the context of an earlier decision.
Reasoning with uncertainty
Represent confidence explicitly through probabilistic reasoning, so uncertain evidence can be considered in an evaluation.
Permissions and obligations
Use deontic logic to express what an agent may, must, or must not do, relative to the encoded rules and available facts.
A trace back to the evidence
Retain the source evidence, rule versions, and derivation steps. Review how a conclusion was reached instead of relying on an explanation written afterwards.
A few practical questions.
Does meander replace the language model?
No. The model interprets requests and proposes actions. meander supplies explicit knowledge and reasoning that your application can use to evaluate those proposals.
Does a proof guarantee the action is right?
A proof shows what follows from the encoded rules and the facts used. It does not make incomplete evidence or incorrect rules true. Those inputs remain part of what needs to be reviewed.
Where would we start?
Start with one decision your agent needs to make: the available evidence, the applicable rules, and the cases that need human review. We can help you map the workflow and assess an integration.
Our consulting servicesWork with us
Have a workflow in mind?
Tell us what your agents need to do, which rules they should follow, and where you need a traceable decision.
Book a meeting