EMET 1.3.0 is public on PyPI and GitHub. It is a byte witness: it can re-check whether a file or evaluation packet still matches recorded bytes and report MATCH, DRIFT, or UNVERIFIABLE. I am looking for one nonsensitive AI-workflow or eval-reporting failure case that would make this useless in a real review. Good examples: a generated summary that is cleaner than the source, an evaluation packet whose final score hides changed score history, or a receipt that is intact while the underlying task failed. Install the public package or inspect its adapter tests: pip install emet==1.3.0 or inspect the source/test: https://github.com/HarperZ9/emet... https://github.com/HarperZ9/emet... Limits: byte integrity is not semantic truth, model safety, evaluator independence, or regulatory compliance. If this misses the failure case that matters to you, I want to test that.
EMET is a small MPL-2.0 CLI that anchors file bytes and re-computes SHA-256 later to report MATCH, DRIFT, or UNVERIFIABLE. It helps reviewers check whether an AI-touched file, prompt, or presented view still matches the source bytes. It does not decide semantic truth or model safety; it keeps byte evidence outside the system being checked.
I run Claude Code, Cursor, and Codex most of the day, and I keep flip-flopping between two setups that both feel wrong.
Setup 1, babysit: I watch every command and approve everything myself. Nothing slips past me, but I'm stuck at the terminal and the whole point of running an agent is gone. I might as well be typing it myself.
