Peer-citable research establishing why AI decisions require settlement.
These publications trace a single, unavoidable conclusion: without per-decision closure, accountability in agentic AI is structurally impossible.
This work does not propose compliance checklists or ethical guidelines. It demonstrates—step by step—why responsibility must be closed, not merely observed.
Chang, YC · OIA Lab · 2026
Defines cognitive leakage operationally and introduces Δ1 as a binary condition for responsibility finality.
Δ1 specifies when an AI decision can be considered complete, rather than perpetually open. It formalizes per-decision closure by requiring that all responsibility conditions are met at a specific moment in time.
DOI: 10.5281/zenodo.18525055 ↗Chang, YC · OIA Lab · 2026
Defines Conceptual Sovereignty—the requirement that each agent’s accountability boundary remain independently verifiable—and frames Responsibility Completion as a prerequisite for decision settlement in multi-agent systems.
This work establishes why accountability cannot be inferred post-hoc and must instead be engineered as a completion state.
DOI: 10.5281/zenodo.18524766 ↗Chang, YC · OIA Lab · 2026
Proposes a five-level leakage taxonomy (Data, Information, Cognitive, Intent, Conjunctive) and introduces Decision Reconstruction as an adversarial primitive.
Demonstrates why leakage-based risk models and continuous scoring cannot produce responsibility closure, motivating the need for settlement-based governance.
DOI: 10.5281/zenodo.18524974 ↗Chang, YC · OIA Lab · 2026
Identifies protocol-level behaviors in MCP, tool-augmented AI, and agent-to-agent systems that prevent responsibility closure by design.
Shows how decision sequences and behavioral patterns expose intent without any mechanism to define when a decision is finished—creating persistent liability gaps at the protocol layer.
DOI: 10.5281/zenodo.18526555 ↗Why settlement is unavoidable:
Risk taxonomy → Incomplete governance primitives → No decision finality → Liability gaps at the protocol layer
Each publication addresses a different layer of the same structural failure: AI decisions accumulate responsibility, but nothing closes it.
• Accountability cannot rely on observation alone
• Logging and scoring cannot define completion
• Governance without closure produces unbounded liability
• Responsibility finality must be engineered
These findings motivate a single engineering conclusion:
AI decisions require settlement.
This research does not certify systems or assign blame. It establishes the conditions under which responsibility can be considered complete.
Those conditions are implemented in the Decision Settlement Layer.