Thanks. That line wasn't planned. It turned up when I put all four services side by side for the write-up and noticed every one of them already had a decline path, and that I'd built each of those before the thing it guards. Wasn't a design principle, more something I'd been doing without naming it.

The MCP parallel makes sense to me. One thing I'd add from living with the approval gate: read-only vs write isn't the whole boundary. The write path also has to expire. Mine holds a proposal for an hour and then voids it, because someone approving at 9am for a cluster state from midnight is approving something that doesn't exist any more.

On temperature, the reason it lasted so long is that nothing failed. Ollama wants it inside options and I had it at the top level, so it got silently dropped and every call ran at 0.8. The eval harness passed 0.0 and believed it. Three runs over the same five logs, three different severity sets, and I'd been quoting that score as evidence the rubric worked.