cadrumo.application.ledger.party_colocation module¶

Attribute a party’s address values by the block of the document they sit in.

The structural answer to party attribution, for every document whose layout can carry it. The model’s job does not change and the prompt does not grow: it copies values and quotes role evidence for the two identity fields, exactly as before. CODE attributes, which is where every other attribution in this design already lives.

Measured, no real document’s layout carries it, and that is stated here because this is where a reader meets the resolver. Scored against authored anchors rather than against the reader whose output would otherwise be being graded: of the documents carrying a hand-written reference transcription and a hand-written identity for both parties, NONE could be partitioned by any authored anchor pair, and every one failed for the single cause described under _regions() – a two-column header arriving as one line. The zero is a ceiling rather than a rate, since a perfect reader bounds every real one, so no prompt, model or second pass raises it. Segmenting such a header needs spatial information the extractor discards before this module is reached.

Two consequences, and neither is that this module is unwired. It attributes correctly on a stacked header and its gates prove it, so the finding is about which documents exist, not about the code. And on a prose document the unverified-attribution stamp and its review-gate advisory are the operating control rather than a placeholder – see stamp_unverified_party_attribution(), whose own docstring carries the same correction. The failure direction stays safe: unresolved keeps the warning.

The mechanism is containment, never proximity. A party’s role evidence is a printed heading the reader copied – FACTURAR A, Verkaufer, the label a NIF sits under – and it is already checked to occur in the document. Those headings partition the transcription into regions, and a value belongs to the party whose region contains it. A nearest-neighbour or reading-order-distance rule would be inference dressed as determinism: it would answer confidently on a two-column layout where the nearest heading belongs to the other party.

Segmentation is by line, not by blank line. Measured against the real evidence corpus, not assumed: every text-layer transcription in it carries zero blank lines, because a PDF text extractor emits reading-order lines and the visual gap between two address blocks leaves no character behind. A blank-line-delimited implementation would have been correct-looking and dead – never firing on any real document, which is the same failure as evidence no resolver consumes, in mirror image.

Three outcomes, and the third is the honest one. A value found in exactly its own party’s region is ATTRIBUTED. A value found only in the OTHER party’s region is CONTRADICTED – the transposition this whole apparatus exists for, and the one case where refusing is right. Everything else is UNRESOLVED: no role evidence to segment on, a value in both regions, a value in neither. Unresolved keeps the unverified-attribution stamp and the operator advisory rather than inventing a verdict, because a document that cannot answer the question has not answered it.

The zugferd specimen in the corpus is why the both-regions case is not theoretical: it prints the supplier’s postal code a second time in a remarks block below the customer heading, so that value genuinely occurs on both sides and no containment rule can honestly resolve it.