Hardening with Guardrails

Guardrails are autonomic filters that monitor training health in real-time. Unlike simple alerts, Guardrails can be configured to trigger active interventions like learning rate reduction or process termination.

Setting a guardrail involves defining a signal (e.g., 'Loss Divergence'), a threshold (e.g., '> 2.0'), and a reaction protocol. These are managed via the Tactical Command Hub or the local.py config.

The 'Steps' field in a guardrail refers to consecutive occurrences. A guardrail with 'Steps: 3' will only trigger if the failure condition is met for three steps in a row, preventing false positives from stochastic noise.