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

# Audit desk

> Exceptions, checks, sweeps, human actions, and an Auditor chat that says what it did not verify.

# Audit desk

Audit at `/audit` is the read desk for both the bookkeeping and accounting layers.
Its tabs are deliberately different views of the same control evidence.

| Tab        | What it answers                                                                          |
| ---------- | ---------------------------------------------------------------------------------------- |
| Exceptions | Which durable exceptions are open, acknowledged, waived, resolved, or recurring.         |
| Checks     | What the generated Close inventory covers, when each check runs, and which lane owns it. |
| Sweeps     | What was scheduled, what last ran, what it found, and what it did not cover.             |
| Actions    | The subset of findings a human can act on, routed to bookkeeping or accounting.          |
| Chat       | A scoped conversation with the read-only Auditor.                                        |

<Frame caption="The current demo publishes an unreadable pagination envelope. The register refuses the empty and all-clear claims that it cannot prove.">
  <img src="https://mintcdn.com/vinmake/56mx-KkSKyEFY2E7/images/greatbook/features/trusted-response-audit-unavailable.png?fit=max&auto=format&n=56mx-KkSKyEFY2E7&q=85&s=2313e7d25b4eac277700f9c6938bf40d" alt="The current GreatBook Audit exception register refusing an unreadable demo envelope" width="1440" height="1080" data-path="images/greatbook/features/trusted-response-audit-unavailable.png" />
</Frame>

Checks is inventory, not a green run.
The current generated catalogue contains 104 definitions: 82 zero or tie rules and 22 report-only balances.
Nothing on that tab is a run-result or cleanliness claim.

<Frame caption="When the demo worklist cannot be read, the current UI makes lane counts and contents unavailable rather than hiding possible tasks.">
  <img src="https://mintcdn.com/vinmake/56mx-KkSKyEFY2E7/images/greatbook/features/trusted-response-human-tasks-unavailable.png?fit=max&auto=format&n=56mx-KkSKyEFY2E7&q=85&s=f94ebc7e331747002d7b2dfa08ad7e3c" alt="The current GreatBook Audit human worklist refusing a missing demo response" width="1440" height="1080" data-path="images/greatbook/features/trusted-response-human-tasks-unavailable.png" />
</Frame>

## What this surface refuses

Closing a human task does not resolve its underlying exception; only a sweep that rereads the book can do that.
A waiver does not make a finding pass, an unchecked family cannot be called clean, and the Auditor's browser run cannot inject job mode or choose a check family outside the server-owned scope.
The Auditor cannot write to the books.

## When the answer is unknown

An invalid or unevaluable close period is inconclusive.
A sweep names every family it was asked to run and lists what it did not verify.
If the register, counts, or selected filter cannot be read, the desk shows an error and withholds the empty or clean claim.

## Verified at source

* The GreatBook web repository's `src/lib/audit-nav.ts` defines Exceptions, Checks, Sweeps, Actions, and the gated Auditor Chat destination.
* `langgraph_chat/devcenter/audit_api.py` and `backend/app/logics/finance/close_inventory.py` own findings, runs, the generated check inventory, and lane attribution.
* `langgraph_chat/agents/auditor/gl_read.py` implements the four-mechanism read-only channel and refuses transaction-pooler sessions.
* Its `src/lib/audit-read.ts` and `src/lib/keyed-load.ts` prevent stale filter results and unreadable responses from becoming current clean evidence.
