From AI That Generates Good Answers to AI That Verifies What It Is Justified in Saying
Series Overview
Part 1 | Why Does AI Get Things Wrong So Convincingly?
Part 2 | Why the OXMV Loop and Evidence Ledger Are Necessary
Part 3 | Mapping Knowledge for AI with Ontology
Part 4 | OpenAI to End ChatGPT’s Custom GPT Era on December 11, 2026
Part 5 | Taking Aim at ChatGPT’s Chronic “Hallucination” Problem
By Manripungchwi Jung Yeon-tae
Generative artificial intelligence may be entering a new phase of competition. Until now, much of the race in AI has focused on which systems can generate more fluent, capable and sophisticated answers. But as AI moves deeper into journalism, law, medicine, finance, research and corporate decision-making, another question is becoming increasingly important: Can AI verify what it is saying—and recognize when the evidence does not justify a confident answer?
An OXMV-based GPT, built around a verification-centric framework called OXMV, is attempting to address that problem from a different direction. Its most striking principle is surprisingly simple: “Do not automatically trust even the premise of the user’s question.”
An AI That Questions the Question
Consider a user asking, “Why did A cause B?” The question already contains an assumption: A caused B. Or consider, “A study proved that this treatment works. Why is it effective?” That question may contain several assumptions—that the study exists, that it actually reached the claimed result, and that its findings justify the word “proved.”
OXMV is designed not to automatically promote such assumptions into verified facts. Instead, potentially material factual assumptions contained in a prompt can be separated into what the system calls Prompt Premises and evaluated before they are allowed to serve as factual foundations for subsequent reasoning. The governing principle is: Prompt Presence ≠ Factual Authority.
In other words, simply because a claim appears in a user’s question does not make it true. This distinction may sound obvious, but it addresses one of the more subtle failure modes of conversational AI: a model can accept a false premise embedded in a perfectly natural question and then produce an eloquent explanation of something that never happened.
Breaking Answers Down Into Verifiable Claims
OXMV also treats sentences not merely as blocks of text, but as collections of potentially independent claims. A single sentence may contain a factual assertion, a number, a comparison, a causal relationship, an interpretation and a recommendation. Under the OXMV architecture, these can be decomposed into Atomic Claims so that each material proposition can be evaluated separately.
Important boundaries—such as time, population, scope, conditions, exceptions, quantifiers and modality—are intended to remain attached to the claim during verification. That matters because hallucination does not always take the form of an entirely fabricated fact. Sometimes the error is subtler.
“An association was observed” can become “A causes B.” “May improve” can become “has been proven to improve.” “Some patients” can become “patients.” “A result observed under specific conditions” can become a universal conclusion.
OXMV treats such semantic strengthening as something that must be controlled rather than dismissed as stylistic rewriting.
A Source Is Not the Same as Evidence
Another central OXMV principle is: Source ≠ Evidence. Finding a paper, government report, court decision or official webpage does not automatically establish that the source supports the claim being made. OXMV therefore separates several questions that are often collapsed into one.
Does the source actually exist? Is it the correct source and version? Was the relevant original material actually accessed? What does the relevant passage say in context?
Does that passage directly support the claim? Does it support the same population, time period, scope and conditions? Is the evidence sufficient for the strength of the claim? Is the source current?
Are supposedly independent sources actually derived from the same underlying material? In this architecture, retrieval is not verification. A search result, snippet, abstract, press release or AI-generated summary is not automatically equivalent to examining the underlying evidence.
A Correct Calculation Does Not Make the Input True
OXMV applies the same distinction to numbers. Suppose a user provides an incorrect revenue figure, but the AI performs a mathematically flawless calculation using that number. Was the calculation wrong? Not necessarily.
Was the resulting real-world conclusion verified? Also not necessarily. OXMV explicitly separates these issues through the principle: Calculation Correctness ≠ Input Truth.
The framework similarly distinguishes percentages from percentage points, absolute from relative change, counts from rates, and actual results from estimates, forecasts, scenarios and targets. It also seeks to prevent statistical association from being automatically transformed into causation.
Not Every Error Is an “AI Hallucination”
One of OXMV’s more distinctive features is its attempt to separate different origins of error rather than labeling every incorrect statement a hallucination. Its architecture distinguishes categories including:
PROMPT_PREMISE_ERROR — an error already present in the user’s factual premise.
SOURCE_ERROR — an error already present in a retrieved source.
ANALYTICAL_ERROR — an error arising in calculation, statistics, logic, causal analysis or domain application.
HALLUCINATION — unsupported factual or semantic content newly introduced during generation.
MIXED_ERROR — multiple error origins affecting the same conclusion.
UNKNOWN_OR_UNRESOLVED — cases where the available evidence does not justify determining the origin.
This distinction has practical consequences. If a retrieved source itself contains an incorrect number and the AI accurately reports, “This source states X,” the upstream error and the model’s generation behavior are not necessarily the same problem.
But if the model takes that source and adds an unsupported causal conclusion, removes an important condition, expands a finding to a broader population, or converts uncertainty into certainty, that additional transformation may constitute a separate generation error.
The OXMV Loop: Verification Does Not End When the Answer Is Written
The heart of the architecture is what can be described as a closed-loop verification process. Many systems can retrieve information before generating an answer. OXMV attempts to go further.
If new evidence or counterevidence appears, a source changes, a calculation is corrected, or an important assumption is revised, the system is designed to identify which claims depend on the changed information. It can then determine which claims require reverification, reassess the permitted strength and scope of those claims, regenerate the relevant output, and inspect the generated answer again.
In simplified form:
Verify → Generate → Inspect the Generated Output → Detect Change or Error → Trace Dependencies → Reverify → Regenerate
The important idea is that verification is not intended to be a one-time checkpoint before generation. The generated answer itself becomes another object that must remain within the boundaries authorized by the evidence.
The Evidence Ledger: Keeping Track of What the AI Actually Knows
If the OXMV Loop represents the process, the Unified Verification Ledger represents the state-control architecture behind it. The framework is designed to maintain relationships among claims, evidence, sources, counterevidence, gaps, unresolved verification work, mutations, dependencies, output constraints and verification snapshots.
One especially important principle is that “Verified” is not treated as a permanent label. A law can change. A scientific paper can be corrected or retracted. A clinical guideline can be updated.
A market price can change within seconds. Software and AI models can move to new versions. A claim verified under one set of conditions therefore may not remain verified indefinitely under another. OXMV treats verification as context-bound to factors such as time, source version, scope, population and conditions.
How Is This Different From ChatGPT, Gemini or Claude?
The distinction requires care. General-purpose AI systems such as ChatGPT, Gemini and Claude are themselves evolving rapidly. Their developers continue to improve reasoning, retrieval, tool use, grounding, factuality and other mechanisms intended to reduce errors. It would therefore be inaccurate to claim that mainstream AI systems “do not verify information.”
The more meaningful distinction is architectural emphasis. General-purpose frontier AI systems are designed to solve an extremely broad range of tasks. OXMV, by contrast, places an explicit verification and state-control architecture around the question: What is the maximum claim that the currently available evidence actually authorizes the AI to make?
The difference can be summarized this way:
“If conventional generative AI primarily asks, ‘What answer should I generate?’, OXMV adds another question: ‘After checking the evidence, how much of that answer am I actually justified in saying?’”
This does not by itself establish that OXMV is more accurate than ChatGPT, Gemini or Claude. That is an empirical question requiring controlled comparative testing. But it does represent a different design philosophy: verification is treated not merely as a supporting feature, but as a control layer over generation.
Hundreds to Thousands of Hallucination-Inducing Test Patterns
According to the developer, OXMV has been tested repeatedly against hundreds to thousands of hallucination-inducing question patterns, including comparative tests involving general-purpose AI systems. The adversarial cases are designed to include such traps as false numerical premises, nonexistent sources, incorrect descriptions of real sources, outdated premises presented as current facts, causal assumptions embedded in questions, unsupported comparisons, and forecasts presented as actual results.
These are not merely hypothetical categories. OXMV’s own adversarial framework explicitly defines test families for false numeric premises, nonexistent-source premises, wrong-source descriptions, stale premises, causal presuppositions, comparative presuppositions and forecast-as-actual errors. The developer says that, to date, the OXMV-based GPT has passed the prepared hallucination-inducing internal tests.
That claim, however, should be understood within its proper evidentiary boundary. It is a report of developer-run testing, not yet equivalent to independent third-party blind validation. Importantly, OXMV itself does not claim that hallucinations have been eliminated with 100 percent certainty.
Its own framework explicitly distinguishes development and regression testing from independent blind or holdout evaluation and states that its “Target 99” concept is a development objective—not a measured guarantee of 99 percent performance. An important next step would therefore be independent testing in which OXMV and systems such as ChatGPT, Gemini and Claude are evaluated under identical conditions, with model versions, prompts, tools, scoring rules and test sets controlled in advance.
Why Verification-Centric AI Matters Now
The significance of such an approach goes beyond whether an AI gets an individual trivia question right or wrong. AI systems are increasingly being considered for tasks involving journalism, law, medicine, finance, taxation, scientific research, corporate strategy and public policy. The cost of error in these fields can be dramatically different from the cost of an incorrect restaurant recommendation or casual factual mistake.
As AI enters higher-stakes environments, users may increasingly need answers to questions beyond “What did the AI say?” They may need to know: What assumptions did it rely on? Which sources were actually checked? Did those sources really support the claim?
Was counterevidence considered? Which facts remain unverified? What conditions or exceptions apply? If a source changes tomorrow, which conclusions must be rechecked?
This is where verification architecture could become economically important.
Potential Impact on Journalism
Journalism provides an especially clear example. Suppose a reporter asks an AI, “Why did Company A conceal the incident?” The question already presupposes both that an incident occurred and that Company A concealed it.
If those assumptions are not established, an AI that accepts the question literally could produce an articulate explanation of motives for an event or concealment that has not been verified. A verification-centric system would first identify the embedded premises.
For newsrooms, this could make Claim–Evidence–Source traceability as important as the quality of generated prose. The issue becomes even more important when AI-generated errors enter articles, social media and search results and are later recycled as apparent “sources” for subsequent AI systems.
Law, Medicine and Finance Could Be Even More Significant
The potential consequences become greater in high-risk domains. In law, the existence of a statute is different from whether it applies to a particular case. In medicine, evidence that a treatment works in a study population is different from determining that it is appropriate for a specific patient.
In finance, a forecast is not an actual result. In taxation, the existence of a deduction is not the same as a particular taxpayer being eligible for it. In safety engineering, “no reported incident” does not mean “no risk.”
OXMV V3.4 contains separate verification paths for legal and regulatory analysis, scientific and medical evidence, finance and investment, accounting and tax, policy, and safety.
The Next AI Race May Not Be About Knowing More
As AI systems become more capable, one of their most valuable abilities may paradoxically be knowing when not to make a stronger claim. Not inventing a missing fact. Not treating “not found” as “does not exist.” Not assuming that the absence of counterevidence proves a proposition true.
Not mistaking a correct calculation for proof that its inputs were correct. Not converting a hypothesis into reality. Not converting association into causation. And when something cannot be verified, preserving that uncertainty instead of silently filling the gap.
That leads to perhaps the simplest description of what OXMV is trying to build:
“Not an AI that simply generates more answers, but an AI designed to say only as much as the evidence permits.”
Whether this architecture can consistently outperform mainstream general-purpose AI systems remains a question for controlled, independent evaluation. But the larger question raised by OXMV may be even more important: Will the next stage of generative AI competition be defined not only by intelligence, speed and fluency—but by verifiable trust?
About the Author
Manripungchwi Jung Yeon-tae is a professional-knowledge AI planner, designer, and GPT development specialist. He currently serves as President of the Construction Industry Information Institute and develops professional GPTs across construction, law, taxation, finance, management, medicine, science, history, and philosophy. His research focuses on verified professional-knowledge AI combining ontology, XAI, the OXMV Loop, and the Evidence Ledger.
