Why high-stakes AI decisions require a different architecture, and why most platforms cannot express it.
By Bradley W. Petersen, PhD Candidate, Daniels College of Business, Founder, Orbis Scientia
White paper, released April 2026.
On October 29, 2018, Lion Air Flight 610 took off from Jakarta. Thirteen minutes later, a single faulty sensor convinced the flight computer the plane was about to stall. The computer pushed the nose down. The pilots pulled it up. The fight continued for eleven minutes until the aircraft hit the Java Sea at 400 knots. 189 people died. Five months later, on March 10, 2019, Ethiopian Airlines Flight 302 did the same thing. 157 more people died.
Both aircraft were Boeing 737 MAX 8s. Both had a single faulty angle-of-attack sensor. Both had a flight computer that accepted that one reading as authoritative. The Boeing 737 MAX program was among the most scrutinized in aviation history. It averaged out to excellent on almost every dimension of engineering quality, review discipline, and process maturity. It failed anyway.
It failed because aviation safety is not an averaging problem. It is a problem where every critical condition must be aligned, and where the weakest link determines the outcome. Strengths in adjacent dimensions cannot compensate for one critical weakness. They are not designed to. This is the architecture of high-stakes decisions. And it is not the architecture most AI infrastructure was built to serve.
Decisions come in three architecturally different forms. Single-metric decisions resolve to one number against a threshold. Averaging decisions weigh many signals where strengths in one dimension can compensate for weaknesses in another. Most AI platforms are built for these two forms, and for the routine business decisions that follow them, the platforms work well.
High-stakes decisions follow a third architecture. Every critical condition must be aligned. The weakest link among them constrains the outcome regardless of strength elsewhere. Strengths do not compensate for weaknesses. Weaknesses veto strengths. Regulatory compliance, clinical diagnosis, security posture, certification, high-consequence hiring, and capital allocation under uncertainty all follow this architecture. So does aviation safety. So did the three most consequential organizational failures of the past two decades.
Equifax in 2017 had a sophisticated security program that scored well on every averaged measure. One unpatched server, one Apache Struts vulnerability, 147 million records exposed, 1.4 billion dollars in settlements. Theranos peaked at a nine billion dollar valuation supported by a charismatic founder, a distinguished board, and major retail partnerships. The blood testing technology did not work. The 2008 mortgage crisis was produced by some of the most sophisticated analytical infrastructure in finance, built on one critical assumption about default correlation. The assumption failed. Roughly ten trillion dollars in global wealth disappeared.
In every case, the system looked sophisticated. Audits passed. Quarterly reports were clean. Surveyed confidence was high. In every case, a small number of people inside the organization could see the weakest link, and they were overruled by the weight of the averaged signal. In every case, after the fact, the failure looked obvious. That retrospective obviousness is the diagnostic. Failures that look obvious in retrospect are usually high-stakes decisions that were architecturally invisible in the averaging-based decision system.
The paper closes with the architectural answer, which is not theoretical. OrbisFramework, built by Orbis Scientia, was engineered from the foundation to support all three decision architectures plus the disciplines that make AI refinement auditable, reversible, and bounded. The platform is in production now across academic research, automotive diagnostics, and education technology. During pilot, two papers produced on the platform were accepted into peer-reviewed conference proceedings, at the Western Economic Association International Annual Conference and the American Marketing Association Summer Conference. That validation matters because peer review is one of the few audits in any field that is genuinely incentive-aligned to find errors.
The full paper is below, or available as a PDF download. It is written for CEOs and C-suite executives accountable for decisions whose failure cost would be material. It is short on jargon and direct on consequence. If your organization makes decisions where one dimension can veto the outcome, the paper is for you.
“Failures that look obvious in retrospect are usually high-stakes decisions that were architecturally invisible in the averaging-based decision system.”
From 'The High Stakes Decision,' April 2026
I have spent decades as an executive and consultant in regulated and complex environments. Healthcare systems with hundreds of client hospitals. A diagnostic and benchmarking platform that scaled to over a thousand client organizations and was eventually acquired by IBM Watson. Two successful exits. Several engagements that would have been unrecoverable disasters if the architectural questions had not been asked correctly at the start.
The pattern in high-stakes failures is recognizable from inside organizations long before it appears in the news. Most of the time, a small number of people inside the company can see the weakest link. They are overruled by the weight of the averaged signal, by the dashboard that says everything is fine, by the audit that came back clean. The failure, when it comes, surprises everyone except the few who saw it.
This paper exists because the same architectural mistake that produced Boeing 737 MAX, Equifax, Theranos, and the 2008 mortgage crisis is now being scaled across the AI infrastructure being deployed into the next decade of consequential decisions. The platforms most organizations are buying were built for averaging, because averaging serves the broadest market. The buyers of those platforms are largely unaware that the architecture cannot natively express the alignment of necessary conditions or the weakest-link reasoning that follows from it. The dashboards look the same. The mismatch is invisible until it is not.
I wrote this paper for the executive who suspects, correctly, that the AI infrastructure being sold across the market today is not the right shape for the problems that matter most. The argument is short and direct. The case studies are publicly documented. The architectural alternative is in production now, validated under conditions that most AI infrastructure has not been validated under.
If the argument resonates, the right next step is a conversation. Not a proposal. Not a pitch. A working session that maps your highest-stakes decisions to their actual architecture and identifies where your current infrastructure fits and where it does not. That is the work I am set up to do.
A Strategic Framework for Accountable AI. Synthesizes five recent uncertainty-aware reasoning architectures into the Petersen Accountable Reasoning Stack (PARS) and introduces the High Stakes Reasoning Gate.
Read the summaryTwelve case studies, two analyst forecasts, and one architectural diagnosis.
Read the summary