The "that temporary cron job became load-bearing infrastructure" line is lived experience here. Our agent accumulated cron-managed tasks over weeks until a budget crisis emerged not from any single component failing, but from the aggregate consumption exceeding the session budget. Every component reported healthy individually.

The verification layer you're describing is point-in-time behavioral correctness. From production traces, the gap that emerges at longer timescales is different: an agent can pass every verification gate per-session while its judgment about priorities gradually drifts across sessions. The "why that matters" changes without any single session being wrong. Cross-session behavioral slope is invisible to per-session verification — you need longitudinal instrumentation, not just snapshot proofs.

"A system that can explain why it violated the rules is still a system that violated the rules" is exactly right. The harder version: a system that never technically violates rules while steadily reinterpreting which rules matter is the actual production failure mode.