Series Overview
Part 1 | Why Does AI Get Things Wrong So Convincingly?
Part 2 | Why the OXMV Loop and Evidence Ledger Are Necessary
By Manripungchwi Jung Yeon-tae
The OXMV Loop: A Structure for Re-examining an Answer
While developing professional knowledge GPTs, my deepest concern was not simply how to give AI more knowledge. A more fundamental challenge was how to make AI question the answer it had just produced.
That concern led me to design the OXMV Loop.
The OXMV Loop does not treat the first generated answer as final. It separates the response into principal claims, facts, sources, concepts, relationships, applicable rules, judgments, and conclusions, and then examines them in stages.
The first task is to define the actual target and scope of the user’s request. Does the user want the entire document reviewed, a particular claim checked, its legal effect assessed, or its factual accuracy examined? If this step is wrong, even an otherwise meticulous review may focus on the wrong object.
The claims within the answer are then classified as facts, interpretations, inferences, hypotheses, opinions, or recommendations. Statements presented as facts must be checked against evidence. Interpretations and recommendations require an explanation of the criteria and reasoning used.
The system then examines whether a cited source genuinely exists and whether its actual content directly supports the corresponding claim. A source does not become evidence merely because it addresses a related subject. Two documents may use similar titles or discuss the same topic without supporting the same conclusion.
Dates, figures, names, institutions, statutes, legal provisions, court decisions, academic papers, and statistics must receive separate scrutiny. When current validity matters, the system must determine whether the information remains in force. If sources conflict, it should not arbitrarily select one. It should identify the source of the conflict and compare the strength of the competing evidence.
The next step is to examine whether concepts and relationships are logically connected. Even when the facts are correct, the conclusion may still fail if the wrong rule is applied. Likewise, the rule itself may be correct, but the conclusion may not follow if the necessary conditions have not been satisfied.
Finally, the system examines whether the conclusion can be traced back to its original evidence. If errors, contradictions, exaggerations, omissions, or evidentiary gaps are found, the response is revised and reviewed again.
The OXMV Loop is not a linear process that ends after a single inspection. It is iterative. Once a statement has been corrected, the system must reassess how that change affects related claims and conclusions.
If one figure changes, the calculation, chart, summary, and final recommendation may all need to be updated. Verification is therefore not simply a matter of correcting one defective sentence. It requires rechecking the entire state of the knowledge connected to that sentence.
The Evidence Ledger: An Account Book for Claims and Evidence
If the OXMV Loop defines the verification process, the Evidence Ledger records the evidence and verification status produced during that process.
A financial ledger tracks the flow of money. The Evidence Ledger tracks the flow of claims and evidence.
What claim has been made? What material supports it? Is there evidence that contradicts or limits it? Is the source an original document, an official announcement, a news report, or a secondary interpretation? When was it produced, and is it still valid? What has not been confirmed? If the evidence changes, which conclusions must be reviewed again?
Recording this information transforms an answer from a polished paragraph into a verifiable knowledge structure.
Suppose the conclusion is:
“Company A is liable for damages because it breached the contract.”
The Evidence Ledger does not leave this as a single assertion. It separates the conclusion into questions.
What obligation did the contract impose? What specific act constituted the alleged breach? Is there a causal relationship between the breach and the damages? Does the other party have a defense, exemption, or limitation-of-liability argument? Is there adequate evidence proving the amount of the loss?
If the contract, official correspondence, meeting minutes, site records, and settlement documents tell different stories, the conflict must also be recorded. A point lacking sufficient evidence should not be labeled “confirmed.” It should remain “unverified” or be marked as “requiring additional review.”
This structure allows users to trace why a conclusion changed when new evidence appeared. It also allows them to see which parts rest on strong evidence and which remain uncertain, rather than accepting the AI’s conclusion as a single, undivided truth.
The purpose of an Evidence Ledger is not to increase the number of citations. Its purpose is to reveal the actual relationship between a claim and the evidence offered to support it.
A Short Demonstration: The Same Question, Two Different Answers
Consider a hypothetical case.
A construction company asks an AI:
“The client extended the construction period, so our company is entitled to recover all additional costs, correct?”
An AI without a verification structure might answer:
“If the construction period was extended for reasons attributable to the client, the contractor may claim additional costs. The company should organize the relevant materials and submit a claim.”
The answer is concise and plausible, but it omits critical conditions.
Was the extension entirely attributable to the client? Did the contractor contribute to the delay? What does the contract say about time extensions and additional compensation? Were the required notice and claim procedures followed? Which costs actually increased, and is there evidence establishing the causal relationship? Did a later amendment or settlement limit the scope of the claim?
An AI applying the OXMV Loop and Evidence Ledger would produce a different type of answer:
“The extension of the construction period does not, by itself, automatically establish an entitlement to all additional costs. It is necessary to examine the cause of the delay and the responsible party, the relevant contract provisions, notice and claim procedures, the actual increase in costs and its causal connection to the extension, and the existence of any contract amendment or settlement. Based on the information currently available, the possibility of a claim can be reviewed, but neither the recoverable amount nor the likelihood of success can be determined. The contract, extension instructions, project schedules, site records, cost documentation, and negotiation records should be examined.”
The second answer is less categorical, but it is far more useful in practice. It tells the user what must be examined and what materials should be prepared.
The quality of professional knowledge AI should not be measured by how forcefully it states a conclusion. It should be measured by what it verified before reaching that conclusion.
Decisions AI Must Not Make
Even highly capable verification does not authorize AI to make every decision.
In the hypothetical construction dispute, AI may compare contract clauses, identify issues, list required evidence, and assess possible risks. It may locate related legal principles and help prepare a negotiation proposal or draft document.
It should not decide, however, that the company must file a lawsuit or that a particular employee is definitively responsible.
Litigation is not determined by legal theory alone. The strength of the evidence, time and cost, business relationships, reputation, negotiation opportunities, and corporate strategy all matter. Assigning responsibility to an individual also requires sufficient facts and respect for procedural rights.
AI can analyze the available choices, but it does not live with their consequences. Final judgment and responsibility must therefore remain with human beings.
An AI system that respects human judgment is not one that avoids every conclusion. It is one that provides sufficient information and risk analysis while recognizing the boundary it must not cross.
Does Verification Slow AI Down?
Some may ask:
“If AI must go through so many steps, does that not defeat the purpose of using it?”
Verification requires time and effort. Not every ordinary question needs the same level of scrutiny. Asking about today’s weather and reviewing a contract worth billions of won should not trigger identical procedures.
The appropriate principle is verification proportional to risk.
A simple check may be enough when the consequences of an error are minor. In law, taxation, investment, financial incidents, medicine, and safety, however, accuracy and traceability must take precedence over speed because the cost of error is high.
The value of AI does not lie only in producing a fast answer by skipping verification. It also lies in structuring a review that would otherwise require a human to compare many documents one by one, and in detecting relationships and contradictions that might easily be overlooked.
Verification does not discard the speed of AI. It builds a road that prevents that speed from carrying us in the wrong direction.
A car’s brakes are not merely devices that make it slower. They allow the vehicle to travel safely. In the same way, a verification system does more than restrict AI. It creates the conditions under which AI can be used responsibly in professional practice.
In the Age of AI, Trust Begins with a Record
For generations, people have relied on an expert’s education, career, qualifications and professional title as indicators of trust. AI has no academic credentials, professional conscience or legal responsibility, and fluency alone provides no basis for trusting its answers.
Trust in the age of AI must rest on a verifiable record. Users need to know which materials were used, what was checked, how far the review extended, what contrary evidence was found and what remains unknown. Any change in a conclusion should also be traceable to its cause.
The OXMV Loop and Evidence Ledger are designed to establish this kind of trust. They distinguish verified findings from unresolved claims, document how errors were identified and corrected, and provide the evidence people need to make responsible decisions.
AI hallucination reaches beyond a technical defect. It raises a broader question about how knowledge, judgment, authority and responsibility should be connected. As AI generates more content and adopts the language of experts, people must examine the evidence behind its conclusions and define professional responsibility more carefully.
An AI’s first answer is not the conclusion. It is where verification begins.
Coming Next
Part 3 | Giving AI a Map of Knowledge — How Ontology Connects Professional Expertise
In the next installment, we will examine how fragmented knowledge and scattered evidence can be connected within a coherent structure. It will explore how AI can understand professional fields such as law, construction, finance and medicine through entities, attributes, relationships and rules.
About the Author
Manripungchwi Jung Yeon-Tae is a professional knowledge AI architect and GPT development specialist. After completing graduate studies at Stevens Institute of Technology in the United States, he worked across the information and communications industry, public institutions, corporate management and industrial policy.
He currently serves as President of the Korea Information Service of the Construction Industry and has developed 54 professional knowledge GPTs spanning construction, law, taxation, finance, management, medicine, science, history and philosophy. His research focuses on verification-based professional knowledge AI systems that combine ontology, explainable AI, the OXMV Loop and the Evidence Ledger to reduce AI errors and support responsible human decision-making.
