Essays
Three arguments
The reasoning behind the checkpoint, written to stand on its own: where the control actually belongs, why the popular fix does not give you a bound, and where a bound has always come from.
Gate the enabling change
Approvals tend to land on the actions people fear, and the expensive failures enter earlier, through administrative changes nobody feared. Finance learned this decades ago. Other domains grew their own versions as human process — and agents act through APIs, where human process does not reach.
Read it →An LLM reviewing an LLM is not a safety bound
A second AI catching the first one's mistakes makes errors rarer, but it never makes them impossible: the two fail on the same inputs. Software engineers measured this in 1986, with human programmers.
Read it →A familiar procedure, applied to a new kind of worker
Companies have always handed real work to people who make mistakes. The machinery that makes that safe is centuries old, and AI agents are being deployed without it.
Read it →