Auto-review

Behind every request, Rivermark runs a built-in self-review loop: the agent inspects its own rendered output and corrects it before the response is finalized. The auto_review boolean controls it, and it's on by default.

The review pass catches and fixes:

• poor legibility — text over busy imagery, low contrast, tiny type,
• overflow and clipping — copy that doesn't fit its frame, elements cut off at edges,
• invented content — hallucinated copy, prices, or claims that weren't in your inputs,
• edit failures — changes that were requested but not actually applied,
• instruction drift — outputs that lost the plot of the original task.

Turning it off

{
  "task": "Apply this exact copy to the template and render.",
  "auto_review": false
}

Turning auto_review off results in modest savings of time and cost. We recommend leaving it on — as it is by default — because the review pass is what lets Rivermark make strong guarantees about output quality, and it's a major reason agentic template filling doesn't degrade in automated pipelines.

A reasonable exception: tightly-scoped mechanical tasks (template fills with known-good copy) at creativity 0, where inputs are trusted and throughput matters.

Review and creativity

Auto-review runs at every creativity level. At higher creativity the agent explores a wider space of solutions; the review pass is what keeps adventurous output from shipping with defects — it's the difference between "bold" and "broken".