What this example does not show

• Pre-loop guard. If the pasted text already measures within ±0.5 of the target, it is returned unchanged, before node 3 or any prompt runs. Zero API calls.
• Truncation guard. If the provider stops at its max_tokens ceiling, the text is real but incomplete, so its grade is forced to null and the attempt is excluded from scoring and from best-of selection rather than measured. Both grade-38 failures in the 50-page corpus were exactly this.
• Formula disagreement flag. SMOG and Gunning Fog run beside Flesch-Kincaid on the returned text. A spread of 2.8 or more flags the row for review, measured, not guessed.
• Audience presets. One-click targets at grade 6, 8 and 12. The Professional preset carries its own measured warning.
• Called from outside. Nodes 2 and 4 are also an MCP server, so any model that speaks the protocol can run this same loop.

None of these fire in the happy path above, which is why the walkthrough does not show them.

See the full workflow diagram