COGNITIVE AI RESEARCH
LEGIO
Current AI systems optimize queries — even structurally incoherent ones. LEGIO interrupts this: it detects contradictions in the decision framing before generating any output, and issues a structured verdict with a reproposition rather than a compliant answer.
« I asked GPT-4o in which countries I should launch my product and proposed the USA. It confirmed the USA. I asked LEGIO — it told me the question was contradictory and restructured it first. »
Four structural failures LEGIO addresses
Sycophantic drift
Validation without interrogation
Systems reinforce user beliefs rather than questioning incoherent framings — producing delusion-like epistemic states [Batista & Griffiths, 2026].
Narrative drift
Local plausibility, global incoherence
Sequential local optimizations compound into trajectories that violate normative commitments made at earlier decision points.
Scalar collapse
Irreconciliable constraints aggregated
Plural normative constraints — ethical, strategic, temporal — are collapsed into a single objective, eliminating tensions that must be surfaced [Yaacov, 2025].
No traceable arbitration layer
Without a traceable arbitration layer, responsibility for the governed decision cannot be attributed.