> ## Documentation Index
> Fetch the complete documentation index at: https://docs.humalike.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Audit

> Hand over a raw transcript and get back a full audit of how your agent behaved.

The Audit API takes a conversation **in whatever format you already have it** —
a chat export, a log dump, a copied thread — and returns a complete review of
how your agent behaved in it: a reception report, the context your agent was
missing, a per-turn risk read, and rewritten versions of its worst moments.

It differs from [`analyze`](/api-reference/analyze) in two ways. It parses the
transcript for you, so you do not have to build the message list yourself. And
it runs the whole review, which takes minutes, so you start it and collect the
result later rather than holding a request open.

## The flow

Three calls, in this order:

1. [`audit_prepare`](/api-reference/audit/prepare) — send the raw text. You get
   back a `run_id`, the speakers found in it, and a guess at which one is your
   agent. Billable.
2. [`audit_launch`](/api-reference/audit/launch) — tell the API which speaker is
   your agent. This is what starts the audit. It returns immediately.
3. [`audit-run`](/api-reference/audit/run) — poll for the result until every
   section is filled in.

Step 2 exists because the audit reviews **one** speaker, and picking the wrong
one produces a confident review of the wrong participant. `audit_prepare`
guesses so you can pre-select it in your interface, but the guess is never
acted on by itself.

## Nothing runs until you launch

A prepared run is parked. Until `audit_launch` confirms which speaker is your
agent, the run is never executed and never billed beyond the `audit_prepare`
call that created it. A run you prepare and abandon costs you that one call and
nothing more.

## Size limits

Both limits **refuse**. Neither trims your transcript, because a review of a
slice you did not choose is a claim about a conversation you never sent.

| Limit                             | Value               | Status | Charged?                                           |
| --------------------------------- | ------------------- | ------ | -------------------------------------------------- |
| Messages in the parsed transcript | 250                 | `400`  | Yes — reading the text is what counts the messages |
| Size of the raw text              | About 32,768 tokens | `400`  | No — refused before any work                       |
| Length of `raw_text`              | 300,000 characters  | `422`  | No                                                 |

Both `400` responses name the number you sent and the number allowed, so you
can act on them directly. To audit a longer conversation, send the most recent
250 messages — which 250 matter is a judgement only you can make.

## Billing

`audit_prepare` is billable. Launching a run bills the stages of the audit as
they execute. Polling is free. A run that fails a size check after the text was
read is charged for that read only. See
[Credits and billing](/credits-and-billing).

## Authentication

Every call is authenticated and **owner-scoped**: a `run_id` is only ever
visible to the account that created it. A run belonging to another account and
a run that does not exist return the same error, so run ids cannot be used to
discover anything.

Next: [Prepare a run](/api-reference/audit/prepare).