Stop unsafe AI actions
before they execute
REMORA is an assurance layer for tool-calling AI agents. Before any action runs, it is gated to one of four outcomes — with multi-oracle consensus, policy-as-code that overrides a confident-but-wrong majority, and an immutable audit trail. Human authority at every safety threshold.
ACCEPTassurance conditions met
VERIFYvalidation required
ABSTAINtoo uncertain to decide
ESCALATEhuman review
0%unsafe execution700 adversarial tool-call tasks · Wilson CI [0, 0.55%] · baselines 10–20%
88%selective accuracy, held-outlocked threshold, out-of-sample · p = 1.45×10⁻⁵
99.9%ordered-phase conformal coverage0 of 20 calibration seeds failed
SHA-256tamper-evident audit chainevery decision hash-linked to the last
🛡 Open the Control Room — watch a live decision →Read the paper (PDF)Enterprise white paper (PDF)View on GitHub
⏱Enter →View →
Live Operations
Procedure-linked AI assurance · 4 agent proposals
AROMER Live Status
Real-time safety scorecard · the system that learns, running now
Agent proposes action → Multi-oracle consensus → Policy gate → One of four outcomes → Immutable audit · Human-on-the-loop