The honest science underneath.
A discipline about provable decisions must itself be provable. So the central claim was pre-registered before the data came in — and then it was tested honestly, even when the result was inconvenient.
The hypothesis was written down first.
The central question — does an assembly of artificial-intelligence agents make a consequential decision better than its best single agent acting alone? — was registered on the Open Science Framework (OSF) before the results were known. Pre-registration is the discipline that separates a finding from a story told after the fact.
Open Science Framework pre-registration. Digital Object Identifier (DOI): 10.17605/OSF.IO/ZWM3S. The hypothesis, method, and analysis plan were filed in advance; the results were published against that plan.
Published preprint. The seven-regime disconfirmation is written up and openly available on the Social Science Research Network (SSRN) (June 2026).
Confirmatory study, pre-registered. The designated confirmatory test is now itself registered in advance (OSF, DOI 10.17605/OSF.IO/2YWPS), to be run on real, independently-labeled enterprise cases. It is registered, not yet run.
We tested whether the machine decides better. It does not.
The assembly-bonus hypothesis — that an artificial-intelligence board would out-decide its best single agent — was disconfirmed. The board did not make the call better than that single agent, and it tended to over-reject. We report that plainly, because a discipline built on proof cannot bury its own.
The machine does not out-decide its best single agent. So let it assure the decision, and keep the judgment human.
A disconfirmed result is not a dead end. It is a design input.
The finding is the reason the discipline takes the assurance-not-control position — and it is written directly into the standard rather than left as a footnote.
Assurance, not control
If the machine does not decide better, it should not decide. Its job is to make the human decision provably sound — complete, consistent, governed, grounded, and on the record.
Caps written as policy
The Decision Assurance Levels cap autonomy by consequence. The higher the stakes, the more the standard requires a human to decide — because the evidence says so.
Revisable by design
The caps are stated as evidence-based and revisable. As confirmatory research arrives, the standard adjusts. The framework improves as the science does.
Most of this market is sold on a claim that has not been tested in the open.
Agentic artificial intelligence is entering governance, often sold as a smarter brain that decides for you. We pre-registered the opposite question, published the disconfirming result, and built the standard around it. That is the kind of evidence the discipline asks of every consequential decision — so it asks the same of itself.
Evidence-based is a founding principle, not a slogan. Guidance from the Council is grounded in research and outcomes, not assertion. Where the science is exploratory, we say so; where a confirmatory study can change a recommendation, the standard is versioned to absorb it.
The record is open.
The full record is public — the published preprint, the parent pre-registration, and the registered confirmatory plan. We encourage practitioners, analysts, and researchers to read the plan and the results directly, and to hold the standard to them.
Citation: Guitarte, A. (2026), "No Assembly Bonus: A Pre-Registered, Seven-Regime Disconfirmation of Super-Additivity in Multi-Agent Large-Language-Model Decision Boards," Social Science Research Network, ssrn.com/abstract=7009758. Parent pre-registration: Open Science Framework, DOI 10.17605/OSF.IO/ZWM3S. The disconfirmation is reported against the registered analysis plan. A confirmatory study is now pre-registered (OSF, DOI 10.17605/OSF.IO/2YWPS); findings to date are exploratory and the standard's autonomy caps are stated as revisable accordingly.
We track the standards forming around autonomous agents.
A discipline stewarded in the open watches the public record beyond its own. In the United States, the National Institute of Standards and Technology (NIST), through its Center for AI Standards and Innovation (CAISI), opened a request for comment on AI agent security (2026), summarized the responses in NIST AI 800-5 (2026), and launched an AI Agent Standards Initiative. That work concentrates on securing, identifying, and connecting agents.
Decision Assurance addresses the adjacent, still-open question: once an agent is secure, was the decision it made authorized, grounded, within policy, and reconstructable? The Council tracks these processes, maps the Decision Assurance Levels to their audit-and-accountability themes, and participates in open comment where it can serve the discipline.
The Decision Assurance Council is independent. It is not affiliated with, sponsored by, or endorsed by NIST or any government agency; references to external standards are for orientation, not a claim of alignment by those bodies.
Healthcare is already writing the decision into law.
Where a wrong decision can cost a life, the requirement arrives early. In the United States, regulators now insist that a licensed human — not the model — own an Artificial Intelligence (AI)-influenced coverage or care decision, and that the basis stay open to audit: the Centers for Medicare & Medicaid Services (2024), California’s Physicians Make Decisions Act (2025), Texas (2026), and the federal records rule that requires a certified system to show how a predictive model was built (2025). The Joint Commission, with the Coalition for Health AI, now offers a responsible-use certification covering governance, transparency, and safety reporting (2025–2026).
Read against the Decision Assurance Levels, these requirements describe the upper rungs: a care or coverage decision that is hard to reverse demands the fuller trail — a named human, the reasoning recorded at the time, and a record durable enough to survive an appeal or a review. The law names the human; the Levels name how much assurance the decision’s consequence deserves.
The Decision Assurance Council is independent. It is not affiliated with, sponsored by, or endorsed by the Centers for Medicare & Medicaid Services, the Office of the National Coordinator for Health Information Technology, the Joint Commission, the Coalition for Health AI, or any agency; references to external requirements are for orientation, not a claim of alignment by those bodies.
Hold the discipline to its evidence.
Founding endorsers help keep the standard honest — grounded in research, revised in the open, and never oversold.