14 Attribution

Stage 2: which layer caused a defect class

14.1 Why attribution comes first

A repair removes a defect class only if it changes the layer that produced it. A normalizer cannot restore text that the engine never transcribed; a new OCR pass cannot correct a DTD that forbids a structure the source prints. The attribution assigns each class one layer before any remedy is chosen.

14.2 Layers

Layer	The defect is	The remedy changes
`reocr`	text missing from the page	nothing in code; the next engine of the chain transcribes the page
`prompt-overlay`	wrong structure produced by the engine	the engine’s overlay (or the shared base template)
`normalizer`	a deterministic rewrite of the XML repairs it	the collection’s normalizer module
`css`	the XML is right, the rendering is wrong	`page.css`
`dtd`	the schema lacks a structure the source prints	`page.dtd` via `schema-derive`
`engine`	one engine fails where another succeeds	the order of `engine_chain`
`source`	the scan is defective	a new scan

14.3 Sources of the assignment

The collection’s config/audit_rules.yaml: the field layer of a structure rule or a cue rule. It takes precedence.
pdai/skills/pipeline/pdai-cure/defect_layers.yaml assigns the text-validity codes, which have no per-collection rule.
The fault dossier of pdai-audit: the fix type of its cases (prompt_overlay, normalizer, normalizer_map, dtd_schema). A code whose cases name one layer takes it; the compound type prompt_and_normalizer names two layers and leaves the code unattributed.

14.4 Defect codes and their layers

Structure codes (layer from audit_rules.yaml of copernicus-prim):

Code	Defect	Layer
`LOOP-REP`	generation loop: padding or one unit repeated	prompt-overlay
`TRUNC`	output stops before `</page>`	prompt-overlay
`NORM-REOCR`	the normalizer cut a loop or closed a truncated page	reocr
`EMPTY-MAIN`	`<main>` nearly empty although the scan carries ink	reocr
`TABLE-FLAT`	a number grid transcribed as pipe-separated `<p>` rows	prompt-overlay
`DUP-BLOCK`	an apparatus entry or note transcribed twice	prompt-overlay
`FN-BIJECTION`	a marker without its note	prompt-overlay
`FN-MARKER`	a note without a marker	normalizer
`TAG-VARIANT`	HTML tags or legacy footnote forms	normalizer

Text-validity codes (layer from defect_layers.yaml):

Code	Defect	Layer
`TXT-PAGE-LOST`	most of the page text is missing	reocr
`TXT-GAP`	two or more consecutive scan lines missing	reocr
`FN-MISSING`	a printed note number without `<footnote>`	reocr
`FN-PARTIAL`	a footnote shorter than its scan lines	reocr
`TXT-EXCESS`	XML text without a counterpart in the scan	prompt-overlay
`BODY-IN-FN`	body lines filed as a note	prompt-overlay
`FN-IN-BODY`	note lines filed as body text	prompt-overlay
`SEG-ORDER`	zones out of the scan’s reading order	prompt-overlay

Codes of the collection’s cue rules (SCH-*) carry no layer until the owner assigns one. On 2026-09-26 the baseline found SCH-APPARATUS (copernicus-prim, GW-II and GW-IV, 7 pages) and SCH-REGEST-MISSING, SCH-SALUTATION-MISSING (pirckheimer, 3 pages) unattributed. LN-DRIFT (marginal line numbers out of step with the lines) is a candidate code of the line-numbered editions without a rule.

14.5 Output and gate

The op pdai:cure_attribute collects the flags of fail and warn pages, every code of the pdai-audit defect register (quality/<doc>_defects.json) and the codes of the fault dossier, and produces one class per code: page count, layer, anchor, the source of the layer and the first five pages as specimen candidates. A code without a layer from any source is listed as unattributed and fails the stage. The role attribute-uncoded (kimi-k2.7-code:cloud, fallback gemini-3.8-flash) may propose a layer; the owner confirms it before it is written into the rule file.