Here's the problem. In low-stakes work, an AI error costs a retry. In executive work, an error can cost a quarter, a relationship, or a reputation — and the decision it corrupted may be irreversible. Systems that feed high-stakes decisions need different design rules than systems that autocomplete email.
Rule one: optimize for defensibility, not fluency
A high-stakes recommendation must survive the question "why should we believe this?" That means sources attached, uncertainty stated, and reasoning available for inspection. A fluent answer without provenance is optimized for exactly the wrong thing: being convincing rather than being checkable.
Rule two: make reversibility explicit
Before any action, the system should know — and show — whether it can be undone. Drafts are reversible; sent emails aren't. Reversible operations can run with light review; irreversible ones earn approval gates by definition. Sorting the world this way is more useful than a single "autonomy level" slider, because it matches how experienced operators already think.
Rule three: degrade toward the human
When context is thin, signals conflict, or confidence drops, a high-stakes system should narrow its claims and surface the ambiguity — not paper over it. The failure mode to design out is silent confidence. The behavior to design in: "here's what's known, here's what's contested, here's the decision that needs you."
Rule four: keep the record
Every consequential path through the system — what was recommended, on what basis, who approved, what happened — should leave an audit trail as a side effect of normal operation. Not for compliance theater; because post-decision review is how operating teams actually improve, and you can't review what wasn't recorded.
Why IEVA starts here
IEVA is aimed at executive work on purpose — it's where these rules are non-negotiable, and where a system built on reviewable memory, provenance, and human authority earns its keep. Design for the highest-stakes user and the discipline benefits everyone downstream.