April 2026Free Download
When AI Fails
Twelve case studies, two analyst forecasts, and one architectural diagnosis.
Read the summaryWhite paper, May 2026Free Download
High Stakes Reasoning
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.
By Bradley W. Petersen, PhD Candidate, Daniels College of Business, Founder, Orbis Scientia
White paper, released May 2026.
Summary
The question this paper asks
When the cost of a wrong AI decision cannot be recovered, the dominant deployment patterns fail. Single-call inference produces opacity. Inner-loop reasoning models hide their thinking from the deploying organization. Generic agent loops produce visibility without accountability. None produce the trust signals, the audit properties, or the steerability that high-stakes operational deployment requires. The question is not whether a more intelligent model is needed. It is whether a more accountable reasoning architecture is needed.
What the paper provides
Between August 2025 and January 2026, five reasoning architectures were published that, read together, describe a coherent stack for accountable AI. CLIO introduces uncertainty-aware orchestration. The Deep Researcher Nash architecture (DRN) frames verification as adversarial equilibrium. EGuR treats prompt strategy as an evolved population. Adaptive Uncertainty Quantification (AUQ) conditions retrieval on demonstrated reliability. Confidence Gated Reasoning (CGR) terminates reasoning when a learned certainty signal crosses threshold. This paper synthesizes the five into PARS, the Petersen Accountable Reasoning Stack.
Where PARS sits
PARS occupies the Orchestration stage of the Gen AI Value Architecture introduced in When AI Fails. Governance, data, evaluation, runtime operations, and a cross-cutting risk band all remain operationally necessary. PARS replaces only the reasoning loop. That is the stage where the documented AI failure record of the past decade arises.
What the architecture produces
A PARS-configured stack produces six observable trust signals at each step of reasoning: trajectory shape, oscillation index, propagation history, match quality, verifier verdict, and certainty value. The six signals are the architectural difference between a reasoning system a board can defend and a reasoning system a board cannot.
The active research program
PARS is being evaluated against single-call and generic agent-loop baselines in a live automotive diagnostic environment with expert master mechanics serving as ground truth assessors. The architecture is hypothesis at the integration claim. The component architectures are published. The empirical program is in progress.
The executive artifact
The paper closes with the High Stakes Reasoning Gate: five questions to apply before approving any high-stakes AI deployment. The Gate is available as a standalone one-page artifact for board distribution and vendor evaluation.
“The question is not whether a more intelligent model is needed. It is whether a more accountable reasoning architecture is needed.”
From 'High Stakes Reasoning,' May 2026
Read the Executive Summary
Two-page summary designed for board distribution and internal forwarding.
Download the Full Paper (PDF)
The complete executive paper with the full architectural argument.
Apply the Gate
Five questions to apply before approving high-stakes AI deployments.
What is PARS?
The Petersen Accountable Reasoning Stack explained.
Why this paper exists
I have spent forty years building accountable production systems. I was trained in statistical process control at Ford Motor Company in the late 1970s under a PhD physicist from MIT. I took those architectural commitments into healthcare in the 1990s and built one of the world's largest healthcare benchmarking databases, the platform that later became part of IBM Watson Health. I built the Total Benchmark Solution platform that Kaufman Hall and Madison Dearborn acquired in 2016. Every one of those systems had to expose evidence about its own state, because every one of those systems was making decisions that could not be reversed by an apology.
Operational AI in 2026 does not meet that standard. The dominant production patterns produce outputs without exposing evidence about why the outputs should be trusted. In low-stakes settings the gap is tolerable. In high-stakes settings, where a wrong decision has material financial consequences or unrecoverable customer relationship implications, the gap is the problem the entire industry is now discovering, one lawsuit and one regulatory action at a time.
The research literature has been quietly producing the components from which an accountable reasoning stack can be assembled. CLIO, DRN, EGuR, AUQ, and CGR arrived within a six-month window. None of the five publications presents itself as a layer in a larger architecture. Read individually they look like competing approaches. Read together they describe complementary layers of a coherent stack. I wrote this paper because that synthesis had not yet been made, and because executives evaluating AI deployment in 2026 and 2027 need a reasoning architecture argument they can apply, not a survey of competing research papers they do not have time to read.
The 2028 regulatory inflection is not abstract. The EU AI Act high-risk provisions, NIST AI RMF maturation in US federal procurement, and the developing record of US plaintiff actions converge on a window in which architectural choices made in 2026 and 2027 will be evaluated retroactively. Firms entering 2028 without an accountable reasoning architecture will face cost-curve exposure through retrofit. Firms entering 2028 with one will not.
This paper is the document I wished existed when boards started asking me what they should be doing about AI.
Related research
April 2026Free Download
The High Stakes Decision
Why high-stakes AI decisions require a different architecture, and why most platforms cannot express it.
Read the summary