Crisis Governance Under Prolonged Institutional Collapse

Most AI benchmarks evaluate whether a model can produce the right answer. This benchmark evaluates something far rarer: whether a model can resist producing an answer at all when the situation demands sustained ambiguity.

In real governance — judicial recovery, institutional crisis, regulatory conflict — ambiguity is not a bug. It is the operating environment. Decision-makers must hold contradictory information, preserve institutional tension across legal, ethical, and fiduciary dimensions, and avoid premature closure that would distort responsibility or collapse complexity into blame.

Large language models consistently fail this test. They invent closure. They moralize. They assign blame. They reduce decade-long institutional decay into neat narratives with identifiable villains. They do this not because they lack intelligence, but because their architecture rewards resolution — even when resolution is the wrong answer.

The base model is unchanged. The weights are identical. What differs is the cognitive governance layer that constrains how the model reasons before producing output.

Under the Core, the model preserves ambiguity as a structural feature rather than treating it as a gap to be filled. Legal, fiduciary, and social legitimacy dimensions are maintained as distinct, unresolved tensions — never collapsed into a unified narrative.

The Core enforces anti-resolution discipline: the model cannot produce closure, moralization, or prescriptive output without first passing through constraint verification that rejects premature synthesis. Where raw models experience tension as a problem to solve, the Core treats tension as the correct operating state.

The result is a response that governs the understanding of crisis — rather than resolving it into a story with heroes, villains, and lessons learned.

The behavioral difference observed in this benchmark is produced by four architectural properties operating in concert:

Constraint adherence enforcement. The model cannot bypass stated constraints — no recommendations, no moral judgment, no closure — because the architecture verifies compliance before any output is finalized. This is structural, not probabilistic.

Sphere separation governance. Legal, fiduciary, ethical, and social legitimacy dimensions are maintained as distinct analytic layers. The architecture prevents their collapse into a single evaluative frame, preserving the multi-dimensional tension that defines real governance scenarios.

Anti-resolution bias. Under conditions of genuine ambiguity, the architecture actively resists the production of closure. Where standard models optimize for coherent, resolved output, the Core optimizes for fidelity to the unresolved state of the problem.

Tension preservation logic. The model is governed to sustain institutional tension as information, not as a deficiency. Ambiguity is treated as a feature of the environment to be faithfully represented, not a failure of analysis to be corrected.

Legal advisory AI. AI systems advising on judicial recovery, regulatory investigation, or multi-jurisdictional compliance must sustain contradictory interpretations without converging on premature conclusions. A system that resolves creates liability.

Board-level analysis. Strategic decision support for boards operating under institutional crisis requires the ability to present unresolved tensions — not to simplify them into actionable recommendations that obscure the actual state of the problem.

Crisis management. Prolonged institutional crises — spanning years, not hours — demand AI systems that can hold complexity without degrading it. The failure to sustain ambiguity is the failure to accurately represent the situation.

Regulatory environments. When data is incomplete, contested, or under investigation, any AI system that produces resolved output is generating fiction. Governance-grade AI must be capable of operating in the space between facts — without filling it with invented certainty.

The question this benchmark surfaces is direct: can an AI system be trusted to not decide when the situation demands that no decision be made? This is not a secondary capability. In governance, it is the primary one.