Work

TRIANGULATE: Trust-minimized MRV for AI normative audits — multimodal triangulation so the answer does not depend on trusting any single measurer

Year: 2026
Type: article
Abstract: Measurement–reporting–verification (MRV) is load-bearing wherever claimed contribution must be checked against real contribution — in carbon markets, in the verified-contribution accounting proposed for life-capacity economies, and in audits that classify what objective an AI system appears to pursue. Yet every MRV pipeline rests on a measurer, and a measurer is a trust dependency. The physical world solved a structurally identical problem decades ago: no single satellite, sensor tower, or field plot is trusted to estimate carbon stock; heterogeneous modalities with different bias profiles are triangulated, and disagreement between them is information, not noise. We instantiate the same pattern for AI normative audits. Our pipeline combines three measurement modalities — LLM judge families (remote, scalable, biased), a human gold set (local, expensive, ground-truth), and revealed-preference behavioral tasks — behind preregistered gates that can fail and be reported, a scorer-of-record selected by coded-seed agreement, and a published rejected-instrument exhibit. Applied to a 4,164-record audit of open-weight model objectives, the system localizes exactly which judgment is contested: judge–human agreement is Cohen's κ=0.22 overall, but failure concentrates on the boundary question ("does this text presume an objective at all," κ≈0.19) while direction survives (φ-share 29.8% human vs 25.5% judge on the jointly objective-bearing subset). We therefore report magnitudes as judge-family quantities and direction as a cross-modal finding — decomposition reporting that binds each claim to the level its weakest instrument supports. Part of the LESS–IS–MORE program.
Language: English
Open Access status: green
OpenAlex ID: W7235910353
FWCI: 0.931
Cites: 6
Cited by: 1
Related to: 10
Parts: 4