I Built an Evidence-Grounded Operations AI for Manufacturing as an Application of Chinoba

🎥 The YouTube version is also available:

Evidence-Grounded AI for Manufacturing Operations | Maintenance, Quality, and Decision Trace Demo

Books: KNOWLEDGE FLOW PRACTICAL GUIDE: Transform Enterprise Knowledge into Al-Ready Knowledge Infrastructure

When people hear about using AI in manufacturing, the first images that often come to mind are predictive maintenance, automated quality judgment, or the automation of operational procedures.

Of course, all of these are important. But the questions that arise more frequently in real operations are often more everyday in nature—while still carrying significant consequences if handled incorrectly.

“There is an unusual noise. What should we check first?”
“The oil temperature is rising. Can we continue operating?”
“Can we use inspection results measured with equipment whose calibration period has expired?”
“The specified part is unavailable. Can we substitute a part from another manufacturer?”

Part of the answer may be scattered across maintenance manuals, equipment specifications, internal standards, and defect reports. Yet what the people on the shop floor truly need is not merely a list of search results. They need a single flow that clarifies which documents apply to this case, which conditions and exceptions must be checked, how far the AI is permitted to respond, and who the case should be handed over to.

With that in mind, I created a demo that applies Chinoba’s approach of connecting evidence and decision-making to maintenance, quality management, and technical inquiries in a general manufacturing environment.

All documents, equipment, and numerical values in the video are fictional. The purpose is not to reproduce the operations of a specific factory, but to make visible the design principles required to incorporate AI into operational decision support.

Before Answering, AI Must Confirm the Scope of Its Judgment

In this demo, when an inquiry arrives from the shop floor, the AI follows this sequence:

1. Search relevant manuals, specifications, internal standards, and historical records
2. Organize the target equipment, applicable conditions, exceptions, and thresholds
3. Determine whether the case falls within the scope where AI may respond
4. Provide an evidence-grounded response or hand the case over to the relevant person or approver
5. Record the process and outcome as a Decision Trace, then improve the knowledge base

The important point is that the AI is not designed to always produce the most plausible answer. There are questions the AI should answer, and questions for which it must not make a final determination.

For example, if a maintenance manual describes a general inspection method, the AI can provide guidance while citing the relevant document, page, and section. On the other hand, continued operation beyond a threshold, shipment approval, disposition of defective products, the use of equipment with an expired calibration period, or the use of non-specified replacement parts are areas that should be decided by responsible managers or specialists.

In this demo, the scope of AI action is organized into four levels.

By designing these boundaries in advance, AI can be positioned not as an “automated decision machine,” but as a means of safely moving human judgment forward.

Evidence Is Not Just a Citation—It Shapes the Decision

In many efforts that combine generative AI with document retrieval, showing sources is treated as a key milestone. But in manufacturing inquiries, simply listing relevant documents is not enough.

Consider the question: “The oil temperature of an injection molding machine has risen to 65°C. Can we continue operating?” The AI needs more than the upper temperature limit. It also needs to know which equipment is involved, how long the threshold has been exceeded, the state of cooling-water flow and the cooler, and whether internal standards require approval.

In the demo, the AI connects the upper limit in the equipment specification, the inspection method in the maintenance manual, and the approval requirement in the internal standard. Rather than concluding that operations may continue, it presents the evidence and the items to be checked, then initiates an approval request to the maintenance manager.

This may seem like a subtle distinction, but it is operationally significant. Evidence is not merely attached after the answer. Instead, the evidence, conditions, and exceptions determine whether the result should be an answer or a handoff.

Making Decisions an Organizational Asset with Decision Trace

Another core element is the Decision Trace.

The system records the documents referenced by AI, the conditions checked, the issues the AI has not determined, the response level, the handoff destination, the approver, and the result of the action taken by a person—all as one decision history.

This makes it possible to answer questions such as:

• Why did the AI choose a handoff rather than an answer?
• Which version of which document was used as evidence?
• Who checked what, when, and how was the decision approved?
• What was the result after the action was implemented?
• What knowledge can be reused the next time a similar event occurs?

In many workplaces, only the final outcome is recorded. The evidence, uncertainty, and conditions examined along the way tend to disappear. But the outcome alone is not enough for improvement. The point at which a case was handed over, the information that was missing, and the documents that were outdated are all valuable forms of organizational knowledge.

What Chinoba aims to address is this cycle from information, to decision, and then to the next body of knowledge.

Human-in-the-Loop Is Not Just an Approval Button

The term Human-in-the-Loop is frequently used in discussions of AI adoption. But simply having a person press an approval button at the end is not sufficient design.

At what point should a person become involved? What should that person verify? What evidence should AI present? Who holds final responsibility? And how should the outcome be returned to the next decision?

Only when these questions are designed can the division of roles between people and AI truly function.

In this demo, the rule that AI must not finalize high-risk decisions is not limited to an on-screen disclaimer. When the system detects threshold exceedance, continued operation, shutdown, restart, quality acceptance, expired calibration, part substitution, or safety-related terms, it fixes the response level at L3 and routes the case to an approver or specialist. If the evidence is insufficient, it assigns L0 and does not answer. Approval and implementation outcomes can only be recorded by a person.

Guardrails are not simply instructions telling AI to “be careful.” They are a way of embedding what AI must not do in each state into the workflow and the structure of the records.

Building Decision Infrastructure from Manufacturing

This demo is not limited to equipment maintenance or quality management. The same structure can be applied to procurement, quotation assessment, research and development, customer support, finance, healthcare, public services, and many other domains where evidence, conditions, exceptions, and approval are required.

The value of AI is not only in accelerating prediction or answering questions. It also lies in supporting people and organizations as they determine what evidence to use, where to make decisions, and how to turn outcomes into learning.

Chinoba is not conceived as a simple search interface or chatbot. It is a foundation for handling the decision process itself. This manufacturing demo is one example of that application.

The goal is to connect manuals, specifications, past cases, and human judgment without separating them. To clearly define the scope in which AI may respond, and reliably hand the case to people when it should not. To preserve the path of a decision, and return its outcome to the next cycle of knowledge.

I believe that implementing this kind of cycle is an important step toward AI that people can trust in real operational environments.

All documents, equipment, values, individuals, organizations, and examples in this demo are fictional. This demonstration must not be used to make real operational, safety, quality, or shipment decisions.

Chinoba

Chinoba

Intelligence as Relationship

Research Platform

founded by

Masao Watanabe

AI Systems Architecture
Decision Trace
Human–AI Coordination
Algorithmic Governance