Safe-start guide

Run one accounting case with a complete evidence trail.

Start with made-up data. Set the company rules. Then test one small question before you add more work.

01

Before you start

Use Python 3.11 or 3.12. Choose one company workspace and one named person who can review accounting recommendations.

Do not enter a secret in onboarding.

Keep passwords, tokens, bank credentials, payroll details, and private files out of chats and configuration files. Start with the included made-up data and file exports.

02

Install the toolkit

  1. Download the ZIP and verify its SHA-256 value.
  2. Unzip it into a new local folder for one company.
  3. Open START-HERE.md and run the setup command.
  4. Choose the default 55-skill accounting install. Use the 81-skill install only when you also need the CFO overlay.

The ZIP includes platform instructions for macOS, Windows, and Linux.

03

Run the safe demo and self-test

The demo uses made-up entities, policies, source records, and decisions. It proves that the case ledger, checksums, replay, and review controls work on your computer.

acctctl self-test
acctctl demo

Do not connect QuickBooks, Xero, a bank, payroll, or another live system during this step.

04

Onboard the company

Run acctctl onboard. Review the generated company, policy, calendar, mapping, owner, and unresolved-question files.

  • Confirm each legal entity and ownership boundary.
  • Record the accounting basis, currency, fiscal calendar, and closed periods.
  • List systems and file exports without entering credentials.
  • Confirm materiality, preparer, reviewer, approval, and escalation rules.
  • Leave an accounting policy unresolved when no authorized person has chosen it.
05

Prepare one narrow case

acctctl judgment prepare
acctctl judgment show CASE_ID
acctctl judgment run CASE_ID
acctctl judgment verify CASE_ID

The case must have:

  • One entity, period, currency, subject, and question
  • No more than five allowed verdicts
  • A sealed evidence and policy snapshot
  • Deterministic checks before agent recommendations
  • Independent role passes and preserved disagreement
Do not ask the agents to “review everything.”

A long list becomes a queue of small cases. A new question becomes a new case after scope validation.

06

Review the evidence and record the human decision

  1. Read the question, allowed verdicts, authority, and stop rules.
  2. Open every cited claim and source identity.
  3. Check missing evidence, assumptions, counterevidence, and dissent.
  4. Confirm the policy snapshot and deterministic tool results.
  5. Have the named reviewer accept, reject, or return the case.
  6. Build the audit bundle and keep its checksum with the review record.

The toolkit cannot mark human review complete. It also cannot post a journal, make a payment, process payroll, file a return, close a period, or decide a final accounting policy.