Here's the problem. When an AI system misbehaves — a wrong belief, a bad draft, an overreaching action — the first question is always the same: why did it do that? In most systems, the honest answer is "we can't fully know." The state lives in opaque stores, the history is partial, and the reasoning evaporated with the request.
IEVA's prototype accidentally ran the counter-experiment. Its entire state is human-readable Markdown; its entire history is Git. Every belief is a file. Every change is a commit with a diff. There is no state the operator cannot open.
What total inspectability taught us
- Corrections become trivial. A wrong belief is a visible line in a file. Fix the line, commit, done — and the fix has provenance too.
- Trust becomes granular. You don't have to trust "the system"; you can trust specific memories you've read, and remain skeptical of specific others.
- Behavior becomes teachable. When workflows are explicit, version-controlled skills, improving one is editing a document, not retraining a model.
- Audits become diffs. "What changed since Monday?" has an exact, complete answer.
The product translation
An enterprise platform can't literally be a Git repository — but the properties can survive the translation, and in IEVA's design they must:
- Memory entries stay readable and attributable, whatever stores them.
- Changes to memory and policy keep history — review the evolution, not just the current state.
- Tool use is observable: what ran, when, touching which sources.
- Audit history accumulates as a side effect of operation, not as an add-on log.
The bar worth holding
"Can I audit it?" is a sharper trust question than "is it accurate?" — because accuracy varies with the task, while auditability is a property you either built or didn't. An assistant you can audit can be corrected, improved, and safely relied on a little more each month. An assistant you can't audit just accumulates unexamined risk with a good interface.
We know which one we'd let near executive work.