The platform
A no-code environment where your expertise or know-how becomes a decision engine.
rAIson turns your expertise or know-how into a deterministic reasoning agent — whether you provide it through a natural-language dialogue, as plain text, or as numerical data. You work in the language of your own application; the platform handles the logic underneath.
From specification to service
Four stages, one of them human.
Your specification is captured
You provide the specification of the decision problem — however you choose to give it — and its structure (defaults, exceptions, overrides) is turned into formal rules.
There are four ways to provide it — see Workflows.
Structured preference-based argumentation (PAF) code is generated
Each decision and each preference is expressed as a decision or preference argument rule — with a labelled explanation. The code is generated by the platform itself in the dialogue route, and by a large language model (Claude) in the text, data, and done-for-you routes.
A human approves it
Before anything is deployed, the expert/user reviews each rule and its explanation in natural language, then approves, edits, or rejects it. The approved program — not any automated draft — is what the engine commits to.
Symbolic AI engine executes every decision
The PAF program compiles to a propositional CNF formula; a modern SAT solver computes conflict resolutions in milliseconds, returning a verdict with the rules that fired and the counter-rules they defeated.
Why translate policy text for a symbolic AI engine?
Text tells humans what to do. Our symbolic AI engine lets machines decide — and explain why.
Text stays the source of truth. The symbolic AI reasoner is how that text becomes actionable.
Automation
A regulator's text becomes an executable decision service. Ask the engine a case, get a verdict.
Explainability
Every decision is backed by a chain of fired rules and defeated counter-rules — an audit trail by design.
Revisability
When the text changes, only the affected rules change. No retraining, no model drift.
Consistency
The same case never gets two different answers. Contradictions become explicit conflicts with preferences.
Expressive power combined with computational power
One language rich enough to model how decisions are really made.
rAIson combines the high expressive power of our structured preference-based argumentation framework (PAF) with the high computational power of SAT solving — so a single model can capture the full texture of a real policy, not a flattened approximation of it.
Non-defeasible knowledge
What holds unconditionally — the ground truth a decision can always rely on.
Defeasible, contestable knowledge
Knowledge that is derived, defeasible, and overridable by a stronger argument — used when different sources give conflicting assessments of the same information, and the conflict itself requires preference resolution.
Reasoning under missing information
Information whose truth value cannot be determined at decision time, so the engine assumes it — exploring both worlds, where the assumption holds and where it does not, and deriving a decision for each.
Which rule wins a conflict
When rules collide, preferences decide the outcome — the heart of non-monotonic reasoning.
Preferences over preferences
Context-dependent priority — which preference applies depends on the situation itself.
The full decision, faithfully
Facts, beliefs, assumptions (abducibles), preferences and meta-preferences in one program — then solved deterministically.
What is policy-style text?
Text that prescribes what an agent should do under competing conditions.
The platform is built to read the linguistic signatures of policy — the patterns that make a document a set of rules rather than prose.
- Modal verbs — shall, must, may, should, will.
- Default-and-exception — "generally X, unless Y, however if Z, except when W".
- Nested qualifications at multiple levels of specificity.
- Explicit overrides — "regardless", "notwithstanding", "in any case".
- Conditional triggers on observable facts or beliefs.
Where it appears
- Laws & regulations (EU AI Act, GDPR)
- Clinical guidelines (NICE, WHO)
- Corporate policies (risk, HR)
- Agent scenarios (buying, routing, negotiation)
- Case law & judicial reasoning
- Standards (ISO, codes of practice)
The symbolic AI engine
From rules to a Boolean formula — and why that matters.
Every PAF program — expressed in propositional or first-order (FOL) form — compiles to a propositional CNF formula whose satisfying assignments are the program's admissible extensions (i.e. winning decision arguments and therefore supported decisions). Once it is in CNF, three decades of solver engineering come for free.
Reasoning, offloaded to SAT — under the admissibility semantics of our PAF
Modern SAT solvers handle millions of clauses. The hard part is solved once.
SAT is the lingua franca
Deadlines, resource limits, safety properties — each becomes another CNF clause, solved together by the same engine. No new machinery to build.
Assumptions (abducibles) become first-class
Missing information is modelled as free Booleans whose truth is explored — exactly what SAT solvers were built to do.
Completeness
It tells you when the answer is genuinely ambiguous.
Per-rule probing classifies each verdict as sceptical — a single winner — or credulous — a genuine dilemma — by applying admissibility semantics. You see whether the system has a unique answer or has flagged ambiguity for human review.
Verdict shape
Every result carries its own reasoning.
- The decision reached
- The rules that fired
- The counter-rules they defeated
- Whether the answer is unique or contested
Where it sits
A position the mainstream alternatives can't occupy.
Retrieval is not decision
RAG fetches the relevant clause but leaves the decision to the LLM that reads it — stochastic, and with no rule trace to audit.
Rules without reading
Classic engines need rules authored by a knowledge engineer, and they don't read the source — encoding cost dominates. And their reasoning is monotonic: a derived conclusion stays valid regardless of what comes after, with no mechanism for handling contradictions.
No theory, no audit
LLM agents that decide through prompting are inspected after the fact, by sampling. No theory to review, no diff to audit when the model changes.
See it on your own policy.
Bring a regulation, a guideline, or a decision scenario. We'll show you the agent it becomes.
Request access