Concept
Lossless Faces
A publication identifier is not a string that happens to spell a
document — the spelling is part of the identity. A catalogue prints
IEEE P61850-9-3/D=FDIS-2016, a citation prints
IEC/IEEE P61850-9-3_FDIS, February 2016, and both name the same
project draft. PubID treats each printed form as a face: it parses
the spelling it was given, records what each token meant, and renders a
canonical face that carries the same information.
The rule
A transformation may not silently discard a token the source spelled. When two spellings name one document, they are aliases, and the canonical face is the one the reference’s own renderer produces. An alias is allowed to be lossy in its surface only if re-parsing the canonical recovers the identity — which is what the corpus checks.
This is why aliases can need more than one round trip. A spelling can
normalize into another alias rather than straight into the canonical;
the corpus records how many render/parse rounds each alias needs (its
rounds value), and both the exporter and the mirror walk them.
When the canonical cannot re-parse
Some reference renders are themselves lossy — a narrative tail or a
joint render can produce a human face that the reference’s own grammar
rejects. Those rows are not “bad data” to be re-spelled; they are
recorded as a quarantine class (render_unreparseable) with the
exact count, so the defect is visible and a fix trips the ledger. A
quarantine is a documented reference defect, never a silent skip.
The tests that keep it honest
- Every corpus case’s canonical must re-parse to itself.
- Every alias must reach the canonical within its recorded rounds.
- Every quarantine row must stay counted at the same number.
- A run that changes any of these numbers is a corpus round, with a provenance receipt and an explanation — not an edit to make CI green.
The faces are lossless in the only sense that matters: no consumer has to guess which document was meant, and no authority’s spelling is overwritten by another’s.