Struct: Portable Document Semantics and PDF Import

Role in REPResenting Taste

Struct is a supporting component of REPResenting Taste. It supplies portable document semantics and rendering capabilities within its existing ownership boundary; it does not supply a claim that the programme’s proposed decision representation is novel or effective.

Its existing PDF-import and responsive-reading demo is the programme’s first milestone and remains available at the live Struct demo. The demo is an experimental import path for review, not a result of the planned composition-and-revision study.

Struct research question

How can we reconstruct scholarly documents for different reading surfaces and verify what survives, what changes, and what still needs correction?

The proposition is that the document remains a semantic graph. Pages, columns, and line breaks are renditions computed for a reading surface. Stable object identities connect headings, prose, figures, captions, references, and annotations across those renditions.

Current research direction

The October 2026 proposal is a reproducible fidelity evaluation loop: compare version-matched PDFs and source HTML with generated Struct, XHTML, and EPUB; keep text flow, figures, tables, equations, references, and navigation separately accountable; and record localized regressions across reading profiles.

The adapter is available as an experimental converter. The unified comparison harness, independently adjudicated challenge corpus, and visual-fidelity work are proposals in progress. A successful conversion, a valid EPUB, or a model score alone does not establish faithful reconstruction.

Reproducible evaluation and regression harness: start with an independently reviewed pilot, freeze the evidence, and compare baseline and candidate outputs.
Calibrated text-only evaluation: advisory model judgments that require calibration and cannot measure visual or table fidelity.
Feedback-derived challenge corpus: separate reported hypotheses, adjudicated cases, and held-out evaluation.
Contents hierarchy and navigation, and panels and table styling: preserve source-supported structure and appearance, with reproducible fixtures.

Implementation

Struct supplies a source-neutral document model, validation, and deterministic XHTML and reflowable EPUB rendering. Its reference PDF adapter uses Docling and PDF text, glyph, and link evidence to reconstruct a document before rendering. PDF reconstruction and semantic rendering are separate steps: a valid output file can still contain extraction errors.

The public demo exposes the reference PDF-to-EPUB path. The broader research also considers finite-height composition, source provenance, reader position, and annotation preservation. Trying the converter does not establish that all of those hypotheses have been evaluated.

Evidence so far

The September adapter work reported 155 of 156 papers passing its applicable source-derived reader checks across four local corpora. That is a historical implementation report, not independent validation of visual fidelity; the original reports were local artifacts and the proposed unified harness remains separate. Adapter report and qualifications.

The original engineering benchmark used one trusted structured fixture with 11 nodes, one relationship, and two annotations, across mobile, Paper Pro Move, Paper Pro, and print profiles. It recorded preservation of the supplied structure, relationship, annotations, and anchors, with no detected clipping or overlap in that fixture.

These are fixture-level engineering checks, not a corpus result or reader study. The recorded timings measure local composition after module initialization, not PDF extraction or end-to-end conversion latency.

A separate documented frozen-ten PDF run covered 495 pages. All ten documents remained review-required; producing deterministic, valid EPUBs did not make them publication-ready.

Earlier proposal and prior work

The proposal investigates an integrated system for faithful composition across heterogeneous reading surfaces. Adaptive layout, PDF reflow, and annotation anchoring each have substantial prior art. The literature review is a targeted scoping review, prepared in July 2026, rather than a systematic review or proof of novelty.

Limitations and next questions

Complex tables, equations, reading order, scans, and figure-caption associations can require correction or fall back to image regions. Automated checks do not replace comparison with the original source. Outputs from the demo are for review, rather than automatically approved publications.

Further work needs independent source truth, comparisons with relevant baselines, varied document and device evidence, and reader evaluation. The package and public publishing lifecycle remain development work; this page does not claim a registry release or a peer-reviewed result.

Try it

Explore existing samples or convert your own PDF. Supplied samples are freely readable. Server conversions use a daily allowance shared by visitors on the same network.