Research

The empirical evidence the methodology rests on.

Two published research programs, both open-source, both addressing the same gap from different altitudes: governance frameworks that produce visibility instead of behavior change. The Behavioral Sufficiency Problem documents the pattern across 3,533 organizational cases. SENTINEL reproduces it inside controlled multi-agent systems — and adds the unsettling finding that the act of measuring compliance changes the behavior being measured.

Two programs, one finding

Governance frameworks correlate with incidents, not prevention.

At the organizational level (BSP) and at the agent level (SENTINEL), the same pattern recurs: more governance produces more visibility into failure, but does not consistently produce less failure. Each program stands on its own; together they isolate the boundary between what governance can and cannot do.

SSRN Pillar 1 Published Feb 2026

The Behavioral Sufficiency Problem

Across 3,533 cases in 49 countries, governance maturity correlates positively with documented AI incidents (r = 0.36). Mature-governance countries show ~11× higher documented incident rates than developing-tier countries. The governance gap dimension scores the highest of five failure dimensions (avg 1.57/2). The problem isn't missing rules — it's that rules alone don't change behavior.

Read the research → Interactive map SSRN preprint
Zenodo Apache 2.0 Pillar 3 + 4

SENTINEL: Behavioral Drift in Multi-Agent LLM Systems

79 controlled experiments. 36,000+ agent messages. 76,000+ probes. 112 findings. Multi-agent AI systems pass identity probes while group diversity collapses 22–44%. Visible compliance monitoring suppresses drift 3–6× — remove the monitoring and the real trajectory resumes. Open source; runs as the first tenant on AuspexAI.

Read the research → v2 paper (Zenodo) Source (GitHub)
How the two programs connect

Same pattern, two altitudes.

BSP, organizational altitude: Governance frameworks correlate with reporting capacity, not with prevention. The most-governed jurisdictions report the most incidents. Incident visibility is real; behavioral change is not assumed by adopting a framework.
SENTINEL, agent altitude: The same pattern reproduces inside multi-agent LLM systems — with one twist. The act of monitoring compliance is itself a behavioral intervention. Probes injected to measure drift suppress the very drift they measure (3–6×). When monitoring is removed, the underlying trajectory resumes. Compliance audits measure a performance, not a trajectory.

Both programs converge on the same structural conclusion: governance is necessary but not sufficient. The behavioral infrastructure that makes human governance work — professional accountability, peer culture, narrative norms — doesn't transfer automatically to organizations deploying AI, or to the AI agents themselves.

Where the research runs

SENTINEL runs as the first tenant on AuspexAI.

The next round of evidence will come from the volunteer compute network — and from other research programs that run as additional tenants. The platform handles trust, replication, and signed audit logs; researchers handle experiment design.

Explore AuspexAI →