Silent Guard or Assumption Drift

Schema validation, model/token budgeting, or indexing assumptions silently change inputs (e.g., unknown keys stripped, multi-label indexing truncated, dequantization expands memory, “thinking” tokens exceed timeouts), causing guards to run too late or outputs to break under real constraints.