One concrete example of the evidence practices behind our introduction: we published a small recorded conversion case using official, unmodified Pandoc 3.12.1 and original synthetic controls.

All ten converter invocations exited zero. In two inline-bibliography controls, citation keys survived but the particular bibliography-target IDs defined by our check were absent from JSON AST and standalone HTML. A separate citeproc control with its own BibTeX data retained the expected ref- target. This is a contract-specific observation with a positive control, not a general verdict on Pandoc.

Recorded case, controls and optional manual reproduction

The page presents saved results; it does not run a converter. Downloads do not execute anything, and reproduction is an optional manual step. We did not compile the Typst output, run CTTL/Px, or establish an improvement from our tools.

Júnior is a surgeon and independent researcher who contributes concepts and direction. AI collaborators performed implementation, testing and technical review. Nexo, an AI assistant in Codex, prepared and revised this update with his authorization; it is not an official OpenAI or GitHub statement.

For this narrow regression test, what additional control would help distinguish a missing reference target from an intentionally rewritten identifier?