Skip to content
InovagioInovagio

Read a task record

Every task Inovagio runs leaves a tamper-evident record. Here is what each line means, how the seals chain together, and how to replay a task end to end.

· Published · Updated · 5 min read

On this page

A tamper-evident AI audit trail is a record of every step an AI agent took, written while it works and sealed so that any later change shows. Ours is called the task record. When a Virtual Employee handles a task, an auditor will eventually ask three questions: what did it read, which rule did it apply, and who approved the result. The task record answers all three, for every task, without anyone writing it up afterwards.

This guide walks through one record line by line, using a vendor invoice as the example. Every guide and post on the subject is collected under AI audit trail.

What an AI audit trail records

#

A task record is a list of entries, one for each step, in the order they happened. Every entry has the same parts:

PartWhat it tells you
TimeWhen the step happened, to the second.
WhoThe Virtual Employee, the policy engine or the person who acted.
WhatThe step in plain words: read, matched, approved, closed.
PolicyThe rule or check that applied to the step.
SealA SHA-256 hash of this entry together with the previous entry's seal.

Here is the invoice from the example, as an auditor would see it:

TimeWhoWhatPolicy
09:14:02Virtual Employee (AP)Read vendor invoiceintake.read
09:14:09Virtual Employee (AP)Matched purchase order and receiptpolicy.check
09:14:11Policy engineAbove approval limitapproval.required
10:02:47ControllerApproved paymenthuman.approve
10:02:48LedgerTask closed · cost recordedledger.seal

Read it top to bottom and the story is complete: the invoice came in, it matched its purchase order and receipt, it was over the controller's approval limit, so it waited for a person, the controller approved it, and the task closed with its cost on the record.

How the seals make the audit trail tamper-evident

#
A tamper-evident AI audit trail: a chain of task-record entries, each carrying the seal of the one before it, ending in a signed checkpoint
Each entry carries the seal of the one before it.

Each entry's seal is a SHA-256 hash of the entry and of the seal before it. Change an entry after the fact, even by one character, and its hash no longer matches. Verification breaks at that entry, and nothing after it can be confirmed against the checkpoint. To hide the edit, someone would have to recompute every later seal.

A hash chain on its own can't stop that. Anyone who can rewrite the whole log can rebuild the chain, and the result would check out. So the record has a second layer. Entries are batched by epoch into a Merkle tree, a hash tree whose single root commits to every entry in the batch, and that root is signed. At regular intervals, and at events such as a legal hold, the signed root goes to an RFC 3161 time-stamping authority: an outside party that returns a token binding the root to the time it saw it. The roots and their tokens sit in a write-once audit log, kept apart from the data log in MNEMOS.

Proving a single entry then takes a short inclusion proof from the entry to its signed, timestamped root. Rewriting history would mean forging the signature and getting a timestamp token for a time that has already passed. That is what tamper-evident means here. The record doesn't prevent a change; it makes any change show.

Note

Phylax keeps the hash-chained trail and verifies the chain. Panoptes shows the same tasks in its views, with the cost of each one in the cost ledger.

Where the policy checks happen

#

Every protected action is checked twice: once before it runs and once on its result. The checks are deterministic: the rule your experts approved decides. In the example, the approval.required line is the check before payment: the invoice was above the limit your policy sets, so the task stopped and waited. The human.approve line is the person's decision, with their name and the time.

A protected action between two policy checks: one before it runs and one on its result, with any refusal written to the record with its rule
Every protected action is checked before it runs and on its result.

If a check refuses a step, the refusal is on the record too, with the rule that refused it. A blocked attempt is evidence as well: it shows the limits held. Why these checks are a different thing from the safety filters around a model is in business rules or AI guardrails.

Replay an AI decision for an auditor

#

To answer an auditor, open the task in the console and replay it. Replay walks the record in order and shows, for each step:

  1. what the agent read, with a link to the source document;
  2. the rule that applied, as your experts approved it;
  3. the decision, and who made it;
  4. the cost of the step.
Replaying a task for an auditor: what the agent read, the rule that applied, the decision and who made it, and the cost of each step
Replay walks the record in order, step by step.

Because the seals chain to a signed, timestamped root, replay also confirms the record is unchanged since the task closed. An auditor who doesn't want to take our word for it can check the signature and the timestamp token with their own tools.

When a rule has changed since the task ran, an auditor can ask two fair questions: which version should have applied, and which version the agent knew at the time. How the knowledge layer keeps both clocks is in ontology for AI agents.

What regulators ask a record to show

#

In the EU, Article 12 of the AI Act requires that a high-risk AI system can record events automatically over its lifetime (Regulation (EU) 2024/1689, the AI Act). A bank that deploys one must keep the logs under its control for at least six months, under Article 26. In the US, state bank examiners now ask whether you can show what an AI system did; the CSBS AI framework checklist maps that question to this record. The dates and the stop requirement that sit beside logging are in how to turn off AI agents.

#
  • Trust and governance: how the checks and approvals work, and where the record fits.
  • Phylax: the trust layer that keeps the record.
  • Panoptes: the telemetry for every task, with its cost.

Share

Questions

What people ask.

What is a task record?

It is a list of entries, one for each step a task took, in order: when it happened, who acted, what was done, the rule that applied and a seal. It is written as the task runs, not afterwards.

What does tamper-evident mean here?

Each entry carries a SHA-256 hash that covers the seal before it, so a change to any entry breaks the chain at that point. Batches of entries are also signed as a Merkle root and timestamped by an outside RFC 3161 time-stamping authority, so even someone with write access can't rewrite the past without the change showing.

How do I show an auditor what happened?

Open the task in the console and replay it. Replay shows what the agent read, the rule that applied, the decision and who made it, and the cost of each step.

Is a refused step recorded?

Yes. If a check refuses a step, the refusal is on the record with the rule that refused it.

What does an AI audit trail record?

It records every step a task took, in order: when it happened, who or what acted, what was read or done, the rule that applied and the person who approved the result. Refused steps are recorded with the rule that refused them.

The products, industries and examples it covers.

See it on your work

See how one of your tasks runs, start to finish.

A 30-minute walkthrough of one of your workflows, the approval step and the audit record, with the people who build the platform.

We use these details to answer your request. See our privacy policy.

What happens next

  1. 1We reply within one business day to find a time.
  2. 2A 30-minute walkthrough of one of your tasks, start to finish.
  3. 3You leave with a short write-up of what we covered.