Problem
Some OCR outputs can misclassify role/title fragments as employment company names. In specific edge cases, these weak single-source candidates can be accepted as normal employment rows by the consensus layer.
Root cause
- Company-candidate detection allowed long text without company-specific hints.
- Single-method candidate groups could be treated as agreed results.
- Person-level merge did not preserve weak-evidence semantics strongly enough after count reconciliation.
Expected behavior
- Role/title fragments without company hints should not become company candidates.
- Single-source company candidates should be marked as evidence-filled rather than agreed.
- Generated review output should make weak evidence visible instead of presenting it as consensus.
Privacy note
This issue intentionally excludes private sample names, file names, OCR text, enterprise names, server paths, and environment details.
Problem
Some OCR outputs can misclassify role/title fragments as employment company names. In specific edge cases, these weak single-source candidates can be accepted as normal employment rows by the consensus layer.
Root cause
Expected behavior
Privacy note
This issue intentionally excludes private sample names, file names, OCR text, enterprise names, server paths, and environment details.