Cross-format offset properties¶

OpenMed adapters normalize PDF, OCR, RTF, ODT, and presentation content into a text stream with half-open character offsets. The shared property harness in openmed.structured.offset_properties checks the part of that contract that is independent of a parser:

• offsets use Python code-point indexes and satisfy 0 <= start <= end <= len(text);
• source spans are ordered and non-overlapping;
• detector spans project to every overlapping source span;
• zero-width detector spans are valid and do not create a fake source range; and
• failures are reported by a stable category rather than a parser message.

The harness is dependency-free and never opens a path or makes a network call. It accepts any adapter result exposing text and spans, so format-specific parsers can be tested with the same invariant suite without bundling parser dependencies.

Running the shared suite¶

from openmed.structured.offset_properties import (
    build_synthetic_offset_cases,
    run_offset_property_suite,
)

reports = run_offset_property_suite(
    {
        "pdf": pdf_adapter,
        "ocr": ocr_adapter,
        "rtf": rtf_adapter,
        "odt": odt_adapter,
        "presentation": presentation_adapter,
    },
    cases=build_synthetic_offset_cases(),
)

Each adapter callable receives a SyntheticOffsetCase and returns an object with text and spans. The committed test suite uses in-memory adapter-shaped results only; real parser tests remain responsible for their own optional dependencies and source fixtures.

OffsetProjectionReport.to_dict() contains counts, offsets, and a SHA-256 text fingerprint. It intentionally does not include the normalized text or any surface value from a span, making it suitable for deterministic local evidence.

The property suite is structural support tooling. Passing it is not a parser accuracy guarantee, compliance certification, or clinical decision guarantee.