| name | reality-audit |
| description | Audit a "done" report or near-final artifact against reality. Use when an AI agent or a person reports work as finished, or when something substantial is about to be sent, published, submitted, or shipped. Separates corroborated structure from self-reported metrics, hunts real defects before release, and returns a trust or ship/hold verdict tied to live checks. |
Reality Audit
A worker cannot certify its own work. When something claims to be done, the claim and the proof often come from the same source. When something is about to ship, the maker reads what they meant, not what is there. This audit creates the separate check.
Use it on any hand-off you are about to trust or artifact you are about to send: an assistant that says it emailed the team, updated the doc, booked the trip, finished the draft, hit a number, or prepared a near-final launch artifact.
Choose the entry mode
- Done-report mode: start from claims. Split what was reported from what is proven.
- Pre-ship mode: start from the artifact. Assume it is broken and look for what would embarrass, mislead, or get rejected.
1. Split the claims into two piles
For done-report mode, read the report and sort every claim:
- Structural, something exists or ran: a file was created, a message was sent, a step executed, a section was added, a setting was changed.
- Self-reported metric, a number or judgment the worker assigned itself: "100% done", "all checks pass", a score, a quality rating, "fully tested".
The two piles get checked differently. Most false confidence hides in the second pile dressed up as the first.
2. Corroborate the structural claims yourself
For each structural claim, look at the actual artifact, not the summary of it:
- "Sent the email" → open the sent folder and read it.
- "Updated the doc" → open the doc and find the change.
- "Added the section" → read the section.
