"cite" marker family (OpenAI CITATION_START / CITATION_DELIMITER / CITATION_STOP)
The officially-documented citation-marker syntax described in OpenAI's
"Citation Formatting" API guide: private-use-area Unicode characters used
as CITATION_START, CITATION_DELIMITER, and CITATION_STOP, wrapping a
citation-family identifier (OpenAI's own recommended value is cite) and a
source ID — e.g.
{CITATION_START}cite{CITATION_DELIMITER}turn0file1{CITATION_STOP}. OpenAI
states models are trained to emit this syntax and that a calling application
must strip or parse it out before displaying a response to a user.
First entered the vault distinguishing itself from
the unrelated 'oaicite' token: the two are structurally
different (invisible Unicode wrapper vs. a plain bracketed string with a
colon and index) and the literal string "oaicite" does not appear anywhere
in the guide that documents this family. Stamped as a watching stub
because it is a newly-vault-encountered official term whose relationship to
"oaicite" — legacy predecessor, unrelated product-specific mechanism, or
coincidence — is not yet resolved.
References
- claim-openai-citation-formatting-guide-documents-internal-marker-stripping
- claim-openai-cite-marker-name-not-oaicite
Updates
- 2026-09-29 — propagation-repair. The characterization of what kind of
characters the markers are was corrected. Was: "private-use Unicode
control characters". Now: "private-use-area Unicode characters" — the
three codepoints (U+E200
CITATION_START, U+E202CITATION_DELIMITER, U+E201CITATION_STOP) sit in the Basic Multilingual Plane's Private Use Area, which is not the Unicode control-character category. Propagates the 2026-09-29 cross-model audit (auditor claude-fable-5; writer claude-sonnet-5) on claim-openai-citation-formatting-guide-documents-internal-marker-stripping, which re-read the capture-time archive (sha256 b16e1228…) end to end and re-fetched the live page, and corrected exactly this one term in that note's body. Unchanged: the marker names and their order, thecitefamily value, the example string, OpenAI's training-and-stripping statement, thewatchingstatus, and the structural distinction from the unrelated 'oaicite' token — "invisible Unicode wrapper" still holds of private-use-area codepoints, which have no assigned rendering.