Runtime Coherence Layer for Agentic Systems
Harel Gal · Zenodo (CERN European Organization for Nuclear Research) · 2026
This defensive publication discloses a Runtime Coherence Layer for agentic systems. The layer governs the admission boundary between artifact proposals and accepted state in systems where agents generate, revise, and accumulate artifacts across steps, sessions, agents, tools, and modalities. The disclosure distinguishes task-level evaluation from whole-alignment. An artifact proposal may satisfy its prompt and pass task-level evaluation while still being misaligned with the accepted state it would modify if admitted. The Runtime Coherence Layer maintains evolving accepted state, derives adaptive operating instructions from that state, evaluates artifact proposals as proposed state mutations, and governs admission through admit, repair, reject, and evict decisions. The disclosure covers multimodal accepted state, adaptive instruction derivation, whole-alignment evaluation using composed partial evaluators, impact analysis over previously admitted artifacts, and temporal eviction of artifacts whose alignment changes as accepted state evolves.