human_review()

The gate that makes the rest of this work. Every email passes through a person before it is sendable.

Inputs and outputs

In. Drafted email and its supporting brief.

Out. Approved, edited, or rejected. Rejections include a reason that feeds the eval set.

Current version

Workflow. UI version 3.1. Forkable through the prompt library.

Model defaults

None at the gate itself. Reviewers see the draft, the brief, the signals, the citations, and the model receipts. You can override per stage in your routing config.

How it fails

Reviewer fatigue. We track median review time and warn when it drops below 35 seconds; below that we have evidence the reviewer is rubber-stamping. Reviewing a draft in under ninety seconds is the playbook.

Evals

Reject rate is the headline number. Last quarter we rejected 14.2% of drafts. The breakdown is in the Field Notes piece on rejection categories. The full eval set ships with the prompt and is editable. Eval format.