Separation of Logic: Why the AI Reasons but Never Writes the Rules
The Principle
Separation of Logic, defined by Babak Shafiei through PyxonData Inc., is the architectural principle inside Lumina Cortex that draws a hard line between the AI model and the rules it reasons with. The AI provides language and inference. The rules, thresholds, physics models, regulatory limits, and engineering standards are owned by domain experts and live outside the model. Change the rule, change the behavior. Audit the rule, audit the system.
Most AI systems entangle the model and the logic it operates on. The behavior of the system emerges from the weights of the model itself, plus a long, opaque prompt. When the system gives an answer you disagree with, there is no rule to inspect, no threshold to change, no document to audit. There is just a model, and the only intervention available is "ask it again."
That is unacceptable in industrial environments. Engineers, regulators, and auditors must be able to point to the rule that produced a recommendation and the person who owns it. Separation of Logic makes that possible.
Two Sides of the Line
The AI Provides
- Language understanding and dialogue
- Pattern recognition across messy data
- Reasoning chains that combine evidence
- Explanations a human can read
- Routing across specialized agents
Domain Experts Own
- Alarm thresholds and severity bands
- Physics and process models
- Regulatory and safety limits
- Engineering standards (API, ASME, ISO)
- Budget rules, fiscal constraints, and policies
The line is enforced. The AI cannot quietly drift the corrosion threshold or invent a budget rule because it "sounds reasonable." If the rule is not in the rule layer, the AI must say so and ask. If the rule is in the rule layer, the AI must apply it as written.
How Separation of Logic Works in Lumina
Rule Packs
The logic an agent applies is held as rules rather than absorbed into model weights: vibration severity bands, corrosion growth assumptions, work-order priority logic, fiscal close rules. Because a rule is written down rather than learned, a domain expert can read it, argue with it, and test it.
Citation in Every Answer
Every answer tells you whether it came from your data or from the agent's domain expertise, and that distinction is drawn from whether a query actually ran rather than from anything the model asserts. Where a figure came from your data, you can drill through and check it against the underlying rows.
Immutable Safety Layer
Safety and compliance agents carry a veto in the Boardroom, and when one of them objects the outcome is settled in code rather than by the model chairing the debate. An agent cannot be argued out of a veto, because the decision never belongs to the language model in the first place.
Change-Controlled Edits
Because the logic is code rather than model behaviour, a change to it is a reviewable change with an author and a history, in the same way any other engineering change is. Bringing that under a formal approval flow your review board can sign off is active work on our roadmap.
Explainability by Construction
Because the rules live outside the model, an explanation traces back to logic a person can read instead of stopping at "the model just knows". A calculation that lives in code is auditable and testable, and it returns the same answer tomorrow for the same reason it returned it today.
A Scenario: Updating a Corrosion Threshold
A new inline inspection campaign reveals that a particular pipeline alloy is corroding 20% faster than the original rule pack assumed. In a typical AI system, the only way to incorporate that finding would be to retrain the model or rewrite the prompt. Both are expensive, opaque, and impossible to audit cleanly.
In Lumina, the integrity engineer opens the corrosion rule pack, raises the growth assumption from 0.10 mm/yr to 0.12 mm/yr, attaches the supporting inspection report, and submits the change. After approval, every integrity agent across the organization begins reasoning with the new value. Past recommendations are still traceable to the old rule pack version. Future recommendations are traceable to the new one.
The model did not change. The data did not move. Only the rule changed, and the entire system updated coherently around it.
Why This Matters
Auditability
Every output is traceable to a rule, a version, and an owner. Internal audit and external regulators can review the system the same way they review a procedure binder.
Change Without Retraining
Rules change with the business. Regulations update. Standards revise. Lumina can absorb those changes by editing a rule pack, not by retraining a model or rewriting a prompt.
Safety Cannot Drift
Safety and compliance rules are immutable. No clever prompt, no transferred memory, no model upgrade can erode them. The boundary is enforced by architecture, not policy.
Domain Experts Stay in Charge
The vibration analyst owns the vibration rules. The integrity engineer owns the corrosion rules. The CFO owns the budget rules. The AI is a powerful assistant, not a quiet decision-maker.
Where It Sits in Lumina Cortex
Separation of Logic is the trust foundation of Lumina Cortex. It is what allows the other layers to be safely composed:
- Experience-Transfer Memory can carry judgment between users because it cannot rewrite the rules.
- Vertical Intelligence Traversal can cascade between agents because each agent applies its owner's rules consistently.
- The Boardroom debate is grounded in cited rules, not in model opinion.
- The System of Reasoning is only a system if its reasoning is inspectable. Separation of Logic is what makes it inspectable.
Related Reading on Lumina
- Lumina Cortex: The Organizational Intelligence Fabric
- Experience-Transfer Memory
- Vertical Intelligence Traversal
- Customer Memory Space
- The Boardroom
- The System of Reasoning
- Sovereign AI for energy, oil and gas in Singapore
Learn More
For the full thought leadership and real-world context behind these concepts:
- Intelligence Integration: Why the Energy Industry Should Stop Integrating Data and Start Integrating Intelligence: the foundational article on pyxondata.com by Babak Shafiei.
- Introducing Lumina Cortex: the full article on Medium (mirror of the pyxondata.com post).
See it on your own data
Upload a dataset and ask in plain English. No setup, no pipeline, runs in your browser.
Get new insights in your inbox
No spam. Just new articles on industrial AI, data sovereignty, and the reasoning layer, sent when we publish.