Relay-Gradient + Alarm Channel: A Locality-Preserving Invariant for Propagation Under Adversarial Failures (v1.0)

This technical note introduces a locality-preserving propagation invariant for distributed agents operating under adversarial failure.
We implement a minimal relay-gradient channel for path formation and a fast alarm channel for monotonic negative event broadcast. The architecture requires only local state and bounded-cost signaling. No global map, centralized controller, or deep search is used.
Across increasing stochastic failure regimes and a hard poisoned-corridor condition, the system exhibits graceful degradation rather than cascade collapse. Successful routes remain stable and task completion persists even when the primary path becomes permanently invalid.
This demonstrates that reliability can emerge as a structural property of the propagation substrate itself.
A fully reproducible reference implementation is provided as a public Google Colab. All reported results can be regenerated by running the notebook with default parameters.
This work defines a locality-preserving compute primitive for edge robotics, capability operating systems, distributed agents, and adversarial tool-use environments.