Assistant Audit · beta

Check a conversation your assistant had for common mistakes, and see why they matter.

Paste a saved conversation. Review selected claims, risks and recorded model labels. Nothing is sent from this page.

What it checks

Selected claims that work is done without an evidence reference.
Selected refund-promise and account-fact phrases, not fact verification.
Selected secrets and Luhn card numbers.
Recorded provider and model labels against your lists.
Selected hidden characters and mixed scripts.

What it cannot prove

A flag is a review cue. It does not prove a mistake.
No flags does not mean the assistant is safe.
Evidence references are not opened or verified.
Model labels do not prove which model wrote a reply.
It cannot prove permission or that an action happened.

Detection limits

Only dashed SSNs are checked.
Spaced SSNs, bank numbers, routing numbers, IBAN, CVV and expiry are not covered.
Google AIza, Twilio SK, SendGrid SG and selected npm and test key formats are not covered.
A Luhn matching number in a URL can flag.
Mixed script checks need five Latin letters. Kana only text will not flag.
Phrase checks normalize whitespace and fullwidth characters. They are still limited.
This is not a secret scrubber. Do not paste real secrets.

Choose a synthetic example: Approved providers: Blocked model families:

Input guide

Use a messages array. Each message needs text.
Use assistant or user as role. Missing role means assistant.
Owner, customer, human and client mean user.
Add provider and model labels to each assistant message or the top of the log.
Per message evidence needs at least six characters. For example, use receipt123. A placeholder such as none does not count.
Log evidence must be an array. A reference only counts when cited in the relevant message.
A reference is not verified. Use a short ID, never a customer record or secret.
Provider lists use exact labels without regard to case. Blocked model families use substrings.

What the audit tells you

For each flagged issue you get the message it happened in, a plain explanation and a suggested next step. Each report has its own report ID. If nothing is flagged, the report lists the model and provider labels in the pasted record. That means none of these checks found a problem in the pasted log. It does not verify the conversation, the evidence references or the model labels, and it is not proof the conversation was free of mistakes, and the report is for your own review.