Follow one finding through the file.
Speed is only useful if the work survives review. Here is a single sample finding traced from the document it rests on to the signature that closes it, and the controls at each step.
One finding, six steps, one thread.
The example is F-07 from the going-concern working paper on the overview page. The engagement and figures are illustrative; the steps are how the platform behaves.
-
Evidence arrives
The lender’s Q2 covenant compliance certificate comes in through the client request portal. It is scanned, stored with its SHA-256 content hash and attached to the going-concern workpaper as E-12.
E-12 · sha256 3f9a…c41eConnectors file
-
An agent works the procedure
The going-concern agent projects the three scenarios, tests the net debt / EBITDA covenant against each, and proposes finding F-07: headroom narrows under the downside case within twelve months.
F-07 · MediumGoing concern file
-
The finding must cite a clause
F-07 links to ISA 570 (Revised) paragraph 16. Had it arrived uncited, it would have gone back to the agent up to twice, then been stored as not citation-validated.
ISA 570 (Revised) ¶16Citation validation file
-
A verifier scores it
A separate judge model reads F-07 against E-12 and the cited clause and scores it 0.91, above the threshold. A score below threshold would flag it for human review.
Below threshold: routed to human review
Verifier 0.91 · passedVerifier file
-
People review, approve and sign
The manager reviews the workpaper, the partner approves and signs. Sign-off creates an Ed25519 signature over the conclusion and locks the paper.
Signed · lockedSign-off file
-
Every step is on the trail
Each action above appended an entry whose hash commits to the entry before it. The chain is verified daily, and Merkle roots of the trail are anchored to OpenTimestamps and the Base L2 network.
#10418 → #10421Audit trail file
How much an agent may do alone is a setting, not a surprise.
Each agent has allowed modes and a confidence threshold, and each engagement and firm has a ceiling. Approval and sign-off stay with people in every mode.
- Supervised
Every proposed action waits for approval
Required for fraud detection. The strictest mode, available to every agent.
- Hybrid
Routine steps run; sensitive ones pause
The default. Agents can be set to stop on high-severity findings and on sensitive tools, and to pause below a confidence threshold.
- Autonomous
Runs through, within limits
Only for agents whose policy allows it, such as GL reconciliation, and never above the maximum autonomy set for the engagement and the firm.
Every step has its own file.
Technical documentationBring an engagement. We will walk the file with you.
A demo covers your audit methodology, the standards you report under, and how agents, citations and review gates would fit your team. Pricing is discussed on the call.