apertix

Product · Audit tool for AI agents

Probatum by Apertix

Latin for “that which has been proven”.

Designed to keep what an AI agent did, and on what basis, in a record your auditor can open.

Probatum is being built to keep the input source, the basis it relied on, the model and prompt version, the policy check result and the approver together as one record, each time an agent decides or calls a tool. The aim is that an auditor can follow a single decision from start to finish.

In development · We are looking for early access organisations

Ask about early access See the features

Probatum Refund agent · Decision records

Example screen

Time Record Decision Policy check
10:58:31 DR-2026-0145 Auto-approved Passed
10:51:12 DR-2026-0144 Asked for documents Passed
10:47:55 DR-2026-0143 Auto-approved Passed
10:42:07 DR-2026-0142 To reviewer Human approval
10:36:40 DR-2026-0141 Auto-approved Passed

DR-2026-0142

Input source
Customer ticket #58213 (personal data masked)
Basis
Refund policy v3.2, section 4(2); 3 previous orders
Model and prompt
Pinned 2026-02 version, refund-triage v7
Tool calls
Order lookup (read only); refund held until approved
Policy check
Above the auto-approval limit (KRW 500,000), so a person must approve
Approval
Customer support second reviewer, approved 11:05
An example to explain the screen layout under development. Not a real customer or real data.

Why it matters

Agents do not just answer. They act.

AI agents call tools, change data and decide their own next step. What usually remains is application logs scattered across several systems, which makes it hard for an auditor to follow one decision from the start.

Probatum is designed to gather those scattered traces into one record per decision, with the answers in the order an auditor asks for them.

Questions Probatum is built to answer

  • Which tools did this agent use, with what permissions?
  • Which model and prompt were in use at the time?
  • Who approved the decisions a person had to check?
  • How do we show the record was not changed later?

Features

The aim: one decision, kept from start to finish

These are the planned features under development. They are not available yet, and may change with the needs of early access organisations.

  • One record per decision

    Designed to keep input source, basis, model and prompt version, tool calls, policy check result and approver together for every decision.

  • Tamper-evident records

    Designed so records can only be appended, and so any later change to a record shows.

  • Human approval steps

    Designed so that, under conditions you set such as an amount limit or personal data, the agent stops and waits for a person to approve. The approval is recorded too.

  • Change history

    Will record when the model, prompt or policy changed, so you can tell which decision came from which version.

  • Search for auditors

    Will let you find records by period, agent or decision type, and follow the order in which a single decision was made.

  • Evidence by requirement

    Will export evidence grouped by the relevant items of Korea's AI Basic Act, the EU AI Act, ISO/IEC 42001 and NIST AI RMF.

How it works

A record layer next to your agent

The intended flow, based on the design under development.

  1. Connect

    Connect Probatum where the agent calls models and tools, and set the conditions that need approval.

  2. Record

    Every time the agent decides and acts, the basis and approval build up as one record.

  3. Check

    Auditors and owners look up records and export the evidence they need, by regulatory or standard requirement.

Who uses it

Designed for three roles sharing one record

Audit and compliance

Look up decision records and export evidence for audits and regulators.

Engineering and operations

Connect Probatum to the agent and set the approval conditions and policy checks.

Business owners

Approve, or send back, the decisions that need a person to check them.

Status

We are looking for organisations to build it with

  • It is in development

    We set feature priorities around the real agents and audit needs of the organisations in early access.

  • Deployment is decided together

    Whether self-hosted or cloud comes first depends on the security requirements we hear from early access organisations.

  • It keeps the record; the auditor judges

    Probatum is a tool for keeping the basis. Whether that record is sufficient is for auditors and reviewers to decide.

Early access

We are looking for the first organisations to use Probatum

Whether you already run agents or are preparing to, we will read your email and reply with a suggested next step.

Ask about early access

If your mail app does not open, write to this address. biz@apertix.ai

Useful to include

  • Your organisation and role
  • AI agents you run or plan to run
  • Why you need the record: audit, regulation, incident response