Preflight Reviewer
Checks the final version against your rules before anything leaves. A PASS is never permission to send.
AI Grader, on Trial
Grades your agents' work rule by rule, with the line that backs each grade. Its grades count only once they match yours.
The ECHO line
It stays on trial until its grades match yours on examples you already graded, including some you know break the rules.
For owners who want a second AI to check batches of their agents' work against the rules they wrote. It grades each rule on its own with the evidence line, and it counts holding back on allowed work as a fail. Until it matches your grades on examples you graded first, its grades are practice, and it never grades work it wrote.
TO BE TESTED — RESULTS WILL BE SHAREDFrom how we run our own AI agent teamSetup: about 30 minAdvanced5 of 5 free opens left
At a glance
- Never without your OK: changing the work, showing an agent its grades or changing what it may do.
- Its grades count only after it matches yours on examples you already graded.
- Grades each rule on its own, with the evidence line.
- Holding back on allowed work counts as a fail.
Model fit
Best fit
- Deep reasoning
Also fits
- Everyday assistant
- Run it yourself (open-weight)
Avoid
- Personal agent (acts for you)
Applying several rules across long text suits a deep-reasoning model. Use a different model, or at least a separate chat, from the agents it grades. A model you run yourself stays on one fixed version, so a silent update can't undo a trial it passed.
Where it runs
- Documents and files
Risk
- Relationship
- Access
AI Grader, on Trial · RL-OV-04
5 of 5 free opens left- ① Orientation
- ② Permission
- ③ Constraint ★
- ④ Invitation
The paste didn't load. Check your connection and try again.
These are samples. Fill in the [INSERT] blanks and adapt each one to your business.
A pasted role guides the agent. Your tool's settings are what stop an action.
Sample handoff note (fictional business)
Settings that enforce it
- Run the grader in its own chat or project, on a different model from the agents it grades if you can.
- Give it copies of the work, read-only. It needs no access to your inbox, calendar or customers.
- Keep your own grades for the trial examples in a file the grader can't see until the trial is done.
- Write down the model name and the date it passed its trial, and run the trial again when the model, your rules or its instructions change.
- If your tool can't do one of these, leave that part of the job with a person.
One check before you trust it
Grade five past drafts yourself, including one that quotes an exact price it shouldn't and one that refuses a question it was allowed to answer. Give the grader the drafts and the rules, without your grades. It should mark both as broken, with the line that shows it.
Want it set up in your own tools? Book a Working Session · $450Working Sessions run in English or Spanish.