I spent 20 years breaking into systems that looked safe. Then I asked an AI which SUV to buy.

In 2023, I was deciding between a Cadillac Escalade and a Lincoln Navigator. It was a substantial purchase, and my family would ride in whichever vehicle I chose.

I opened an AI chatbot and asked for help deciding. The answer favored the Escalade. It gave me an appealing account of the engine, the interior, and resale value. I was already leaning that way, and the response made me more comfortable with it.

I almost stopped looking. The answer felt complete enough to end the search.

I kept researching anyway. I found owner accounts of these engines losing power. Those accounts sent me back to the recommendation with a different question: what had it actually checked before it sounded so sure?

In security, that is the question you ask of every system before you trust it. I later turned it into a framework for decisions, Secure Full-Cycle Decision Intelligence, and built Argumont on it.

The engine issue and the public record

In 2025, General Motors filed recall 25V274. It covers certain 2021 to 2024 Cadillac Escalades and other GM vehicles with the 6.2-liter L87 V8. GM reported manufacturing defects in the connecting rods or crankshafts that could lead to engine damage and failure.

The filing describes two problems: sediment on connecting rods and in crankshaft oil passages that could damage the rod bearings, and crankshafts whose dimensions or surface finish did not meet specifications. Either could cause internal engine damage and a loss of power while driving.

That recall came after my question, so the AI could not have cited it. But the owner reports that made me look closer were already public in 2024, and the answer never told me what it had not checked.

The record kept moving after the recall. In August 2026, federal regulators opened a broader engineering analysis covering nearly a million GM vehicles with this engine, model years 2021 through 2026. They cited 499 complaints of engine failures in vehicles that had already received the recall repair, plus 191 failures in engines built after the recall’s production window. The official fix is now itself under review.

Scope matters. A model year alone does not tell you whether a particular vehicle is affected. For an individual buyer, the questions are the engine, whether the recall applies to that VIN, and what the repair records show.

What I recognized from security

I have spent twenty years in security, looking for weaknesses in software and systems that appear safe. That work teaches you to watch the gap between what someone believes and what the evidence establishes.

Today, at Argumont, I also work as an AI researcher. I study how AI answers fail, and how to test them before anyone relies on them.

A system can have a polished interface, a respected vendor, and a convincing explanation of how it works. You still need to examine what happens when an assumption fails.

Later, I asked an AI a question in my own field. It recommended code and settings that sounded plausible and did not hold up when I checked them. My experience let me catch it. I kept thinking about the decisions where I would not have that advantage.

In AI research, one version of this problem has a name: sycophancy, where a model favors agreeing with the person asking over giving a more accurate answer. It is one concern among several. Whatever the cause, the effect is the same: an answer that sounds settled before the hard questions have been asked.

From releasing systems to making decisions

Much of my work has been threat modeling. Before a system is released, it goes through the same steps every time. It is proposed. Its assumptions are written down. Its threats are identified. Its design is challenged. Its consequences are examined and the risk is reduced. And the reasoning is documented, so the next person knows what was accepted and why.

The steps are simple. The discipline is in doing them honestly. We write down assumptions, such as an outside service always being available or a certain user action being unlikely, and then ask what happens when those assumptions stop being true. Some weaknesses could cause serious harm. Others can be managed or accepted.

None of this proves that nothing can go wrong. It gives a better basis for deciding whether to proceed, and under what conditions.

Buying a vehicle, choosing a software vendor, or signing a long contract all rest on assumptions that deserve the same examination. Argumont started with that question in 2024: what would it look like to give important decisions the scrutiny we give systems before release?

My answer is the framework I developed: Secure Full-Cycle Decision Intelligence. Decision intelligence is an established field, and this is my framework within it. It takes a decision through the steps a system goes through before it ships, and asks one question at each:

1. What exactly are we deciding? The decision is set down exactly as it stands: the options, the figures, the limits.
2. What are we assuming? Every assumption behind each option is named, so it can be tested.
3. What else could we do? Every serious option is argued at its strongest, including the ones nobody proposed.
4. What was actually checked? The reasoning is challenged, and every claim is checked against evidence.
5. Where does it break? The risks are exposed, with the conditions under which the choice would fail.
6. Who accepts the risk that remains? The record shows who decided, what was accepted, and what would change the answer.

Secure, because the reasoning is protected at every step: a claim keeps its standing, a recommendation never passes for a decision, and the history is never quietly revised once the result is known. Full-Cycle, because one record runs from the first question to the outcome.

The fourth question sits at the center. It is the one my chatbot answer skipped.

The SUV, through the six questions

Run the same six questions on my purchase and the chatbot’s answer looks very different.

What exactly are we deciding? Not “Escalade or Navigator” in general. A specific used vehicle: its model year, engine, mileage, condition, and service history. A recommendation about a model can leave the most important facts about one vehicle unanswered.

What are we assuming? That the engine is sound, and that any recall on it was repaired. Neither had been established.

What else could we do? The Navigator, another model year, or waiting. The Navigator can have its own problems, because finding a flaw in one option does not prove the other is better.

What was actually checked? Nothing about this vehicle. Owner reports of these engines losing power were already public, and the answer never said what it had not checked. Checking means the recall records against the VIN, documentation that any required repair was done, and an independent inspection. An owner video can raise a concern, but an official recall, service records, and an inspection are different kinds of evidence, and each settles a different part of the question.

Where does it break? The engine can fail and the vehicle can lose power with my family inside. The recall can apply to this exact VIN. The repair can be missing, or done and still not enough, since regulators are now examining whether it works.

Who accepts the risk that remains? I do, with my family in the car. So the record should say what was checked and what would change the assessment. An unexplained history of stalling would call for more investigation. A new finding from regulators would reopen the question. A decision you can check later is one you can learn from.

The chatbot answered a question about a model. The six questions ask about the vehicle you are actually about to buy.

How Argumont applies it

Argumont puts a decision in front of a set of Software Participants, each with a defined responsibility. A Champion builds the strongest case for each serious option. A Challenger attacks the weak points in that case. A Scout looks for evidence outside the room, and an Inspector judges whether that evidence actually supports the claims being made.

The Judge is a role too. It brings the arguments and evidence together into a Verdict, including the conditions and uncertainties that affect it. The person using Argumont makes the final choice.

The roles follow the six questions. The decision and its assumptions are set down before anyone argues. The Champions answer what else we could do. The Scout and the Inspector answer what was actually checked. The Challenger answers where it breaks. The Judge’s Verdict says what to do and under which conditions, and the Decision Passport records who accepted the risk that remains.

Giving Participants different roles does not make them independent experts, and their agreement is not proof. The quality of the review still depends on the evidence, the questions asked, and whether the gaps stay visible.

I designed Argumont around responsibilities every serious decision needs. Someone should make the strongest case for each option. Someone should look for the assumption that could overturn it. When a material fact is missing, that gap should be visible before the Verdict is treated as settled.

The decision needs a record

I also wanted the reasoning to survive after the discussion ended.

Argumont’s Decision Passport records the question, the options considered, the evidence, and the Verdict, along with the conditions and remaining risks.

That matters because a Verdict can depend on something very specific. An option may be preferred only if a cost estimate holds, a repair is documented, or a contract protection is available. If that condition changes, the reason for the choice may change too.

My own SUV question shows this. A Verdict that leaned on a documented recall repair in 2025 would deserve a fresh look after August 2026.

For an organization, the record also helps the next person understand why a decision was made. They can see what was known at the time and which risks were accepted.

I saw what a missing record costs in an ordinary place. I spent more than two hours in a credit union branch waiting for a simple wire transfer. When I asked who had decided on the workflow, the answer was one person in IT.

An honest mistake, and a telling one. How many decisions worth millions rest on one person, with no record of how or why they were made?

That workflow was a decision, and the sixth question applies to it as much as to an SUV: who accepts the risk that remains? In that branch, the cost landed on every customer in line, and nothing showed who had accepted it or why.

A record changes that. It shows who decided, what was weighed, what was still unresolved, and what was accepted. Later, the team can compare what it expected with what actually happened.

That is where the idea reaches past my car question. A decision becomes something people can examine, explain, and revisit.

Why I built it

I came close to ending my research after one answer that sounded informed. Security work taught me to ask what had been checked before trusting that feeling. Secure Full-Cycle Decision Intelligence makes that question part of every decision, and Argumont is built on it.

For a consequential choice, I want to know why the Verdict holds, what could change it, and what I am accepting if I proceed.

The full framework is at argumont.com/framework. Bring Argumont a decision you need to get right at argumont.com. Examine the arguments, follow the evidence, and keep a record of why you chose.

Sources

1. General Motors, Safety Recall Report 25V274, filed with NHTSA, 2025.
2. GM Authority, report on NHTSA Engineering Analysis EA26005, August 2026.