Evidence and confidence

Evidence selectors

Evidence uses openparser.document-region selectors. geometryGranularity describes what the coordinates physically bound. A table cell id may also be present under logicalAddress; it identifies the logical row/column even when the provider supplied only a table-region bounding box. Consumers must never render element/table geometry as a cell-precise highlight.

The locator points to the parsed_document returned beside the lineage, so clients can render highlights without duplicating the entire document in each field record.

Word tokens live in the parsed document

Lineage is a companion to parsed_document, not a replacement for it. Per-word text, spans, and recognition scores are not duplicated as lineage entities. Resolve them through the region entity's locator.uri (the parsed document) and selector.value.elementId, then read the word/symbol elements under that element. Only the region-level aggregate is carried in lineage.

This keeps the document small enough to return inline on every extraction — a typical 15-field result is a few KB per field rather than hundreds of KB — while losing no information a client can't already address.

Evidence context

Source entities retain the complete parsed element or table transcription in value. Two fields citing one paragraph read different fragments of it; each fragment is the extract output's value — the field itself when the model copied, a source-text: entity when it rewrote — so the citation does not restate it. The per-field recognition score on the resolved_source_evidence input remains: that is how well those words scanned, which is per-edge information. openparser:sourceContext appears only when the wider element text differs from the region itself, as it does for a table cell. The optional openparser:selectionReason is reserved for a concise provider-reported reason for choosing that region. OpenParser does not emit it today.

The verbatim text a field was read from is never an annotation on the field. It is an entity value, so the graph states it once. Its offsets and containing element are not stored either, because searching the parsed text for that value recovers both. How that value is justified is on extraction and transforms.

OCR confidence

Recognition is scored twice, because "how well did this region scan" and "how well did this value scan" are different questions once two fields cite the same paragraph.

Each cited region carries a derived recognition_source_region assertion over every OCR word inside it (method: region_word_minimum): its weakest word, with minimum, maximum, and regionMean in attributes. This describes the region, and nothing else.

Each field then scores the words it actually read, on the derivation input that cites the region (method: cited_minimum_context_downblend). The weakest word behind the value sets the score, and the rest of the region may pull it down by at most 15% of the gap — never up. So a crisp 2448 in a sentence whose street name scanned badly stays crisp, discounted slightly because a region that defeated the scanner once may have defeated it twice. When the extracted text cannot be located in the region there is nothing to narrow to, so the score falls back to the region's own weakest word and the input reports method: region_word_minimum — a score over the whole region never claims to have narrowed to a value.

Page-only confidence remains on the page entity. When a region has no aggregate and no provider score of its own, openparser:inheritedRecognition names the page entity to fall back to; the score itself is never copied onto a region, cell, or field entity.

Extraction and transform confidence

Extraction confidence is deterministic source-text alignment: the extracted text was verified to appear in the parsed document. When a later transform rewrites that text, the transform step carries its own score — a proven representation change, a coarse model ordinal, or nothing. The two claims are never folded into one number.

openparser:alignmentStatus on a derived assertion records how the extracted text was matched: match_exact and match_fuzzy when that text is present in the document, and ungrounded when nothing supporting it could be found. There is no status meaning "this was rewritten". A rewrite is a second step, not a grade of the first.

The assertion's sources names every activity that moved the number, so a score reduced because grounding came back partial or ungrounded lists activity:grounding-resolution alongside activity:extract.

Why not one number

The alternative is one edge carrying one score for two different claims. That tells you neither whether the read checked out nor whether the rewrite was proven, self-reported, or unmeasured. Two edges let each claim carry only the confidence that actually applies to it.

A consumer that walks a field's ancestors and takes the weakest scored step can show a single conservative headline number. A consumer that needs to explain uncertainty can point at the exact extraction or transform assertion instead.

Signed confirmations append to this graph without rewriting it. See human review.