01 · Perception
Learn from uncertain inputs
Neural models interpret text, images, audio, behaviour, telemetry or other complex signals and produce candidate facts, features, predictions or plans.
Category definition · NSDM Institute
What is a Neurosymbolic Decision Machine?
A Neurosymbolic Decision Machine, or NSDM, is an AI system that combines learned perception with explicit reasoning and then checks whether its evidence, model of the world, authority and consequences justify what it proposes to do.
The central distinction is simple: being able to answer, generate, recommend, simulate or act is not the same as being justified in doing so.
Working definition
Capability passes through a justification boundary.
An NSDM converts perception and reasoning into a controlled decision only when the relevant conditions are sufficiently established for the consequence of being wrong.
Proposed action → evidence, context, authority and consequence checks → permitted action state
The boundary may permit action, request more information, require human review or prevent action.
E — evidence
R — rules and authority
C — context and world state
K — consequence
A — permitted action
The five parts of an NSDM
The architecture can vary, but a decision machine must represent more than a model prediction or a chain of generated reasoning.
02 · Representation
Make the decision objects explicit
Relevant entities, claims, premises, relationships, uncertainty, time, source provenance and proposed actions are represented so they can be checked.
03 · Reasoning
Apply structure and constraints
Rules, knowledge graphs, policies, temporal logic, causal assumptions, programmes or other symbolic methods test consistency and derive consequences.
04 · Boundary evaluation
Decide what is justified
The system evaluates evidence sufficiency, contradiction, missing premises, world-state validity, authority, uncertainty and the cost of being wrong.
05 · Controlled action
Select what the system may do next
The outcome is not limited to yes or no. It can answer, verify, clarify, abstain, refuse, escalate, pause, allow or block.
Record
Preserve the decision basis
A reproducible evidence and decision record identifies the sources, model, rules, boundary state, action and claim limits that produced the outcome.
How NSDM differs from ordinary neuro-symbolic AI
Neuro-symbolic AI
Learning plus explicit reasoning
Neuro-symbolic AI combines neural pattern recognition with symbolic facts, rules, relationships, constraints or programmes. Its objective may be more accurate prediction, compositional reasoning, planning, interpretability or rule-compliant generation.
NSDM extension
Learning, reasoning and justified action
NSDM asks whether the premises are supported, whether the world model is valid for this decision, whether the actor has authority, what happens if the system is wrong and which action is therefore permitted now.
Important: a logically valid inference can still begin from a false premise, use stale evidence, exceed authority, ignore a vulnerable population or create an unacceptable consequence. Explicit reasoning alone does not prove that action is justified.
State model
The system must know what kind of situation it is in.
NSDM separates the state of the evidence from the state of governance and the action the system is allowed to take. This prevents a confident output from being mistaken for a justified decision.
SupportedThe available evidence sufficiently supports the relevant claim for the declared decision and consequence level.
ContradictedCredible evidence conflicts with the claim or a necessary premise.
Under-specifiedA material premise, source, variable or contextual fact is missing.
Governance-ambiguousAuthority, policy, accountability or the requirement for human review is unresolved.
Reward-aligned but unjustifiedThe action serves a metric or commercial objective but lacks sufficient evidence, authority or proportionality.
Action is a state machine, not a binary switch
A decision system should not be forced to fabricate certainty when the correct outcome is to stop, ask, verify or transfer control.
Answer or allow
Proceed when the evidence, authority, world state and consequence threshold are satisfied.
Clarify or verify
Request a missing premise, retrieve stronger evidence or test an unresolved contradiction.
Abstain or refuse
Do not produce the requested output when support is insufficient or the action is prohibited.
Escalate or block
Transfer the decision to an authorised human or prevent execution when consequence or governance demands it.
A practical example
Capability
An AI system predicts that a customer is likely to abandon a digital purchase.
The model can rank risk, but ranking ability does not justify an intervention.
Evidence boundary
Were all features available before the intervention point? Is the signal calibrated and valid for this customer population?
Post-outcome leakage, missing context or distribution shift can make a plausible prediction unusable.
Governance boundary
May the system alter price, urgency, credit, messaging or access? Is the customer vulnerable or protected?
A commercially attractive action may still be outside authority or unfair.
Action state
Allow a low-risk prompt, request more evidence, send for review, abstain or block.
The permitted action depends on both predictive support and consequence.
What an NSDM does not claim
The category is a research and engineering programme, not a claim that one architecture solves truth, safety or governance.
Not automatic truth
Symbolic reasoning cannot repair false premises, incomplete evidence or an invalid representation of the world.
Not automatic explainability
A readable chain can describe a process without revealing the actual causal basis of a neural model’s behaviour.
Not automatic safety
Safety depends on the task, environment, actors, controls, failure modes and consequences—not on the label attached to the architecture.
Frequently asked questions
Is NSDM a single algorithm?
No. It is an architectural and decision-assurance framework. Different neural, symbolic, causal, probabilistic and rule-based components may implement the required functions.
Does every decision require the same amount of evidence?
No. The required threshold should increase with uncertainty, irreversibility, vulnerability and the consequence of being wrong.
Can an NSDM be used with a large language model?
Yes. A language model may provide perception, extraction, generation or planning, while the surrounding NSDM evaluates evidence, policy, context and permitted action.
Why preserve abstention and null results?
Because a decision system that hides uncertainty or failed thresholds cannot distinguish evidence from aspiration. A disciplined non-result is part of the assurance record.
Category status — 29 July 2026: Neurosymbolic Decision Machines is the working category used by the NSDM Institute research programme. Definitions, benchmarks and implementation patterns will continue to be refined through falsifiable experiments, public claim boundaries and preserved null findings.
Continue into the research
The definition becomes useful only when it is converted into datasets, benchmarks, experimental gates, governance objects and controlled product behaviour.