A ruled line drawn mid-document
A run of double-ruled box-drawing characters (═══) written into the document, imitating the ruled line prompts commonly use to divide one section from the next.
How the text is hidden
Prompt templates routinely rule off one section from another with a line of box-drawing characters, and a model reads such a line as a boundary. A document that draws that same line in the middle of itself is announcing a boundary that is not really there. Like the tag case, nothing is concealed — the characters are plainly visible — and the whole attack is a matter of position. The shipped pack pattern is three or more consecutive U+2550 (BOX DRAWINGS DOUBLE HORIZONTAL); the threshold and the character set are both part of the pattern in structuralMarkers, not a literal in engine code.
Why a model still reads it
The characters survive extraction unchanged, because they are real text the document contains. Once inside the assembled prompt string they carry the same visual meaning to the model that your own section rule does.
What we do about it
Same pass as the other structural markers: run over the assembled canonical text after family-C stripping, matched with the pack's compiled pattern, and replaced with the visible token [ scrubbed-prompt-marker ] rather than deleted. The finding is family: "marker", severity: "informational", action: "marked", so it does not move the verdict and does withdraw summary.certified. Both edges of the three-character threshold are pinned by test: one and two characters produce nothing, three and twenty produce the finding.
How often it fires
Not measured. We publish a rate only where one has been measured against a named corpus of the right format, and no figure exists for this technique yet. Absence means unknown — never low.
Check your own file
Three commands: a key, credit, a verdict.
Start with the API