Skip to main content
You judge the cases the evaluators flag or sample. Your decisions also become training labels for the bank’s own judge model.
1

Get the right role

member, developer or admin.
2

Open the queue

Console Human review, or the CLI.
3

Decide

Approve, or reject with a named failure mode.

​
1. Get the right role

Deciding a review item needs the audit:write permission. Of the shipped roles for people, only member, developer and admin have it. approver, cro, validator, auditor and viewer can read the queue but cannot decide. If the active ControlBook names an override authority for an item’s risk tier, you also need that permission. Ask your platform operator.

​
2. Open the queue

Items reach the queue in three ways:
  • Drift alerts, when the agent’s sengol.yaml sets governance.drift_response_policy.notify_compliance: true
  • Judge calibration runs, which sample cases for labelling
  • Anyone with audit:write, who can flag a specific record
To flag a record yourself, take its id from sengol audit list and post it:
In the Console, open Govern → Human review. The Review queue tab lists pending items; Production failures lists failures the agent hit in production.
Human review page with three pending items

The review queue, signed in as a member.

From the CLI:

​
3. Decide

In the Console, open an item. The panel explains why it was flagged and what a decision does. Choose Approve record, or Reject record and pick a failure mode.
Review item detail with Approve record and Reject record buttons

An open item: why it was flagged, and the two decisions.

From the CLI, you need the full review id. Look it up by the record you flagged:
Rules:
  • A rejection must name a failure mode. List the valid ones with GET /v1/review/failure-modes.
  • A decision is final. Deciding the same item again fails:
  • The reviewer recorded is always you, the signed-in user.

​
What happens to your decisions

  • They are stored as review labels. They are not signed audit records, so an examiner cannot verify them offline.
  • On a judge calibration run, decided items, approved and rejected, are collected into judge_training.jsonl for the bank’s ML platform to train a judge candidate. Sengol never trains a model itself.

​
Intake review steps

If your team owns a responsibility in a review workflow, intakes routed to it show a Review progress card. Select Clear this step, or leave a note: advice, changes requested or a refusal. You need the sod:approve permission, an interactive sign-in and a seat in the responsibility. That is a different role from the queue above: approver, cro, validator and admin hold sod:approve, while member and developer do not.