Approach
How we link neural networks and human reasoning.
NeuroLink AI builds one link that works in two directions. Experts can follow how a system reached a recommendation, and the system can represent what experts intend. This page sets out the two gaps, the layer that closes them and the rules we work by.
The problem
A fluent answer is not yet a decision.
In finance and healthcare every decision has an owner. The owner must be able to follow the reasoning and stand behind the outcome. A fluent answer offers neither, and the connection between people and AI breaks in two places.
| Aspect | AI to peopleThe explanation gap | People to AIThe intention gap |
|---|---|---|
| What breaks | The system gives an answer, and how it got there stays hidden or arrives as fluent text that cannot be checked. | The system cannot see what the expert means: the goal behind a request, the rule that matters in this case, the judgement applied without being said. |
| How it shows in the work | A reviewer opens the source files and rebuilds the case to be sure. The system becomes a first draft and is trusted for nothing more. | A request is reworded again and again. The output is fluent, misses the policy that applies and forgets the correction by the next case. |
| What it costs | The verification tax. The time the system saves is spent checking it, so it is moved, not saved. | Rework and drift. Every case starts from zero, and the system never learns what this team prioritises. |
| Why a bigger model does not fix it | A larger model writes more fluent text. It does not produce a record of how a recommendation was reached. | A larger model knows more in general. It cannot know what one team prioritises unless that is stated, kept and checked. |
| What closes it | Reasoning you can follow Evidence, rules and steps, in the expert’s order and terms. | Intent the system understands Goals, constraints, priorities and judgement as explicit structures. |
The link
What travels in each direction.
The name states the design. NeuroLink stands for the link between neural networks and human reasoning. Reasoning travels one way and intent travels the other, and each must arrive in a form the receiver can use.
AI to people
Reasoning you can follow
The system shows its work, in the terms and the order the expert already uses.
A trace, not a paragraph
Each recommendation carries what a reviewer needs to retrace it.
- The evidence, each piece with its source
- The rules that applied, and those that did not
- The signals the model learned from
- The steps between them, in order
- The outcome
Shaped to the reader
The same reasoning is shown at different depth and in different terms. A reviewer sees the policy check first, a team lead sees the exceptions, a new colleague sees the definitions. Role and experience set the order and the vocabulary.
Confidence and limits
The system says how sure it is and why. It names what was missing and flags a case that lies outside what it has seen. Doubt is shown, not smoothed over.
Without it
A fluent paragraph. It sounds right, cannot be checked and sends the reader back to the source documents. That is the verification tax.
In between: a shared reasoning layer.How the layer works
People to AI
Intent the system understands
What the expert means becomes a structure the system can use, not a sentence it has to guess at.
Goals, constraints, priorities
Written once as explicit structures: what the decision is for, which limits must hold and what outranks what when they conflict. The system checks its work against them and reads them back for the expert to confirm.
Examples and corrections
Past cases and fixes show what words leave out. The system generalises from them, states what it generalised and takes a correction to that as well.
Overrides
When an expert changes or rejects a recommendation, they give a reason. The reason is stored with the case and adjusts the system, so the next similar case starts closer to their judgement.
A prompt alone
A prompt is one sentence. It is not kept, cannot be checked against a rule and says nothing about priorities. People rewrite it for every case, and the system still misses the point.
The shared layer
Where learning, rules and expertise meet.
Between the two directions sits one shared reasoning layer. Three kinds of knowledge work on the same decision, and what people decide feeds back into all three.
How to read the figure
- Neural learningLearned from documents and records.
- Symbolic reasoningRules, policies and constraints.
- Cognitive modelHow this expert weighs and decides.
- ExpertSets intent and decides.
- AI to peopleReasons: evidence, rules, steps.
- People to AIIntent: goals, limits, priorities.
- FeedbackOverrides adjust the layer.
Blue arrows carry reasoning to the expert. Amber arrows carry intent back. The dashed amber line is feedback across decisions: each override adjusts the layer for the next case.
| Part | What it contributes | What it cannot do alone |
|---|---|---|
| Neural learning | Reads language and finds patterns in documents, records and history, including patterns nobody wrote down. | State why it reached a result in terms a reviewer can check. |
| Symbolic reasoning | Applies explicit rules, thresholds and policies exactly, and leaves a trail people can inspect. | Read untidy text, or notice what no rule anticipated. |
| Cognitive models | Describe how a group of experts attends to evidence, weighs it and decides, so the system orders and words its reasoning to match. | Supply the facts of a case or the text of a policy. They model the expert and rely on the other two for the rest. |
| Feedback and meta-learning | Turn each override and correction into a change in the other three, so the system learns how this team reasons. | Decide anything. The decision stays with a person. |
The loop
One decision, five steps.
The system reads, weighs and drafts. A person decides. What the person decides teaches the system.
System
Observe
Read the documents, records and policies that bear on the decision.
System
Model the expert’s reasoning
Weigh them the way this expert or team would: what matters, and which rules apply.
System
Recommend with reasons
Propose an outcome with the evidence, rules and steps behind it, at the depth this reader needs.
Expert
Expert decides
An accountable person accepts, amends or overrides. The decision is theirs.
System
Learn from feedback
Record why the expert decided as they did, and adjust the model of them.
Feedback returns to the start. The next case begins from what this one taught.
CogniFlow, the platform we are developing, runs this loop inside existing expert workflows.About CogniFlow
Principles
Five rules we work by.
Adapt the machine, not the human.
Experts have spent years learning how to weigh evidence. We do not ask them to rephrase that work for the software. We model how they reason, and shape the order, depth and vocabulary of every explanation to match.
Reasons before answers.
We design the reasoning first and the recommendation second. A recommendation without its evidence, rules and steps sends the reader back to the start, so the system does not offer one.
People stay in command.
AI proposes and accountable people decide. Nothing is final until a person approves it. An override is a signal to learn from, not an error to correct.
Earn trust with evidence.
Each recommendation shows its sources, the rules applied and how sure the system is. A system that says it is unsure is more useful than one that is always certain. Trust built case by case lasts longer than trust that is asked for.
Research-grade, production-hard.
Ideas from cognitive science, neuro-symbolic machine learning and human-AI interaction have to work in a real workflow: real data, real constraints and people with other work to do. We read the research, and we test it against the work.
Regulation
Regulation as context, not as a claim.
The EU AI Act sets requirements for high-risk AI systems. Three of them, in Articles 12 to 14, describe questions we treat as design questions from the start.
| Article | What it asks of high-risk systems | How our design treats the concern |
|---|---|---|
| Article 12Record-keeping | Systems that automatically record events over their lifetime, so that their operation can be traced. | Inputs, steps, versions and human approvals are recorded for every decision. |
| Article 13Transparency | Operation transparent enough for those who deploy a system to interpret its output and use it appropriately. | Evidence, rules, steps and confidence arrive with every recommendation, in the reader’s terms. |
| Article 14Human oversight | A design that lets people oversee the system effectively while it is in use. | People decide. The system proposes, shows its limits and treats an override as input to learn from. |
This is context, not a claim of compliance. Whether a system meets the Act depends on its purpose, how it is deployed and an assessment of that use.
Start with one decision.
Tell us about a decision your experts make every day. We will model how they reason and test the link in your workflow.