Industrial Twin Lab Manifesto
Build machine intelligence in the twin before trusting it in the machine.
Industrial Twin Lab is an engineering research environment for learning about a machine before an algorithm is trusted near it. Its central rule is: Never let AI perform its first experiment on the physical machine. It is not a digital-twin dashboard, predictive-maintenance package, AutoML platform, IoT monitor, LLM wrapper, or asset-management application. The object of the work is evidence-backed machine knowledge, not merely a model.
Machine knowledge sequence
- Machine
- Observation
- Hypothesis
- Experiment
- Simulation
- Evidence
- Knowledge
- Decision
Principle 01 — Twin Before Intervention
AI should experiment with a validated digital environment before any recommendation reaches the physical machine.
Principle 02 — Physics Before Pure Correlation
Where engineering knowledge exists, machine learning should complement physics rather than ignore it.
Principle 03 — Evidence Before Deployment
A model should not be promoted because its accuracy metric looks good.
- generalization
- robustness
- stability
- uncertainty
- operational value
- explainability
- safe deployment behavior
Principle 04 — Local First
Critical industrial data, engineering knowledge, experiments, and inference should be capable of operating locally.
Cloud services may be optional extensions, never mandatory foundations.
Principle 05 — Isolation by Design
The AI experimentation environment must be logically and architecturally isolated from OT control environments.
Principle 06 — Human in Command
AI may:
- observe
- analyze
- hypothesize
- simulate
- experiment
- recommend
but deployment or operational intervention must remain governed by explicit engineering authority.
Principle 07 — Models Compete
No algorithm should be assumed to be best.
Models should compete under identical experimental conditions.
Principle 08 — Features Compete
The system must evaluate not only algorithms but also the sensors and engineered features that make predictions possible.
Principle 09 — Failure Can Be Simulated
Rare industrial failures should be studied through validated synthetic scenarios, fault injection, historical replay, and simulation.
Synthetic data must never automatically be treated as ground truth.
Principle 10 — Every Experiment Becomes Knowledge
Failed experiments are still valuable.
Every experiment should contribute to organizational machine knowledge.
Principle 11 — Learn from the Fleet
Knowledge learned from one asset should be testable on similar assets.
The ultimate goal is not one intelligent machine. It is an intelligent fleet.
Principle 12 — Trust Must Be Measurable
Every recommendation should expose:
- evidence
- model confidence
- uncertainty
- data quality
- assumptions
- limitations
- provenance
These principles are gates, not slogans. A P-101 bearing-degradation candidate that performs well on one synthetic split has not yet established portability, stability, physical plausibility, or safe deployment behavior. Its Evidence Package must preserve the hypothesis, data and twin versions, feature pipeline, validation regime, uncertainty, limitations, and provenance. A human engineer retains authority to reject, repeat, constrain, or stop the work.
The sequence ends in a decision rather than an automatic actuation. Training a model, recommending an investigation, approving an inference model, and changing a physical setpoint are four different authorities. Industrial Twin Lab addresses the first two and provides evidence for review; it grants no system authority over machinery.
Continue with the safety architecture, the Experiment Fabric, or the open research questions.