acceptodds
Desk rejected at ICLR 2027

Mantle: A Typed Safety IR for Tool Agents Under Interface Drift

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Tool-agent safety logic is commonly maintained independently in adversarial discovery, trace repair, risk prediction, and runtime enforcement. A changed interface can therefore leave four consumers attached to different meanings of the same hazardous action. We introduce Mantle, a typed safety intermediate representation (Safety IR) that assigns a stable, versioned identity to each hazard predicate and compiles that identity into the four lifecycle consumers. Mantle makes a semantic update an auditable systems operation: a validated predicate revision relabels prediction targets, refreshes repair constraints, updates exploration rewards, and changes the runtime action rule together. We also formalize a maintenance frontier that compares shared and stage-local designs under fixed model, data, compute, and human-maintenance budgets. The formulation separates candidate selection from held-out paired assessment and states the matching, denominator, and uncertainty contracts needed to identify cross-stage propagation. This yields a concrete design criterion for tool-augmented agents: safety maintenance should be evaluated by the complete effect of a semantic edit across the lifecycle, not by the behavior of one isolated guard.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
alignment safety and oversight, Mantle

Related papers

Discussion (0)

Sign in to comment.