This bridge paper applies a weak-signal / layered-interpretation architecture to AI safety monitoring, especially runtime monitoring where small deviations may matter before they justify stronger control, persistence, or shutdown consequence. It argues that AI safety systems need a disciplined middle state in which bounded runtime findings can raise attention without immediately becoming durable safety claims, memory changes, capability restrictions, or ignored noise. The paper defines a compact operational transfer contract for early runtime safety anomaly handling: a light state ladder, typed event schema, transition rules, measurable promotion predicates, a minimal governance API, short worked traces, and failure-mode mitigations. The claim is architectural rather than universal: weak runtime anomalies should influence attention before they justify stronger runtime consequence.

This paper is part of a broader architecture project applying the PUTMAN / Spanda weak-signal framework across domains where early signals are partial, noisy, ambiguous, or underdetermined. The larger repository contains related architecture papers, bridge papers, and implementation-oriented notes on stratified interpretation, runtime governance, memory boundaries, deviation handling, and weak-signal promotion control. Repository:https://github.com/putmanmodel/spanda-architectural-framework