# SRT POC limitations log

These limitations are part of the result, not deferred polish. An open item means the current benchmark cannot support the corresponding broader claim.

| ID     | Status | Limitation                                                                                                                                                                         | Consequence / next evidence                                                                                                                                                                         |
| ------ | ------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| LIM-01 | Open   | The engineering sample is one trusted structured fixture with eleven nodes, one figure-caption relationship, and two text annotations.                                             | Results do not estimate corpus-level accuracy. Run the dependency issue's local scholarly-PDF corpus benchmark before making ingestion claims.                                                      |
| LIM-02 | Open   | The fixture is schema-valid source, not an arbitrary or scanned PDF.                                                                                                               | PDF reconstruction, OCR, reading-order ambiguity, tables, equations, multi-panel figures, and malformed inputs are outside this result. Keep their completeness gates and corpus evidence separate. |
| LIM-03 | Open   | The fixed PDF is generated from the same canonical source and is not an independent ingestion baseline.                                                                            | It proves that a fixed-page artifact is reproducible, not that SRT outperforms a conventional PDF workflow. Add a licensed, same-content external baseline before comparative claims.               |
| LIM-04 | Open   | No same-content geometric-reflow implementation or validated output is available.                                                                                                  | Geometric reflow is marked unavailable and receives no inferred score. Implement and validate a baseline before a three-way comparison.                                                             |
| LIM-05 | Open   | Clipping and overlap are DOM-rectangle checks in one pinned Chromium configuration with fixed fonts, viewport, locale, and tolerances.                                             | Zero detections do not cover other browser engines, font substitutions, zoom levels, long documents, or device firmware. Expand the browser/device matrix and retain screenshots.                   |
| LIM-06 | Open   | Finite pagination uses deterministic estimated text metrics; the PDF exporter uses built-in font metrics.                                                                          | Page breaks and low-millisecond composition timings are POC diagnostics, not production typography or end-to-end latency. Benchmark shaped text and actual device rendering before claims.          |
| LIM-07 | Open   | Annotation evidence covers semantic text highlights and notes whose source text is unchanged. Freehand ink attachment remains explicitly deferred.                                 | The result does not establish robustness to edits, OCR corrections, ambiguous repeated quotes, handwriting, or collaborative conflicts. Add mutation and adjudication corpora.                      |
| LIM-08 | Open   | Structural and relationship preservation measure faithful transport of the supplied canonical graph, not whether extraction created the correct graph.                             | Use human-adjudicated source truth to measure semantic reconstruction precision and recall.                                                                                                         |
| LIM-09 | Open   | No comprehension, navigation, fatigue, accessibility, or user-preference study was run.                                                                                            | Do not infer reader benefit from engineering invariants. Human-subject work requires a reviewed protocol and authorization and is outside this issue.                                               |
| LIM-10 | Open   | Composition timing is a local microbenchmark after module initialization, with one first-call sample and 50 same-process warm samples per target. Targets share process/JIT state. | Treat cold/warm figures as reference-machine diagnostics only. Add isolated processes, randomized order, confidence intervals, realistic documents, and end-to-end timing for performance claims.   |
| LIM-11 | Open   | The literature review is a targeted scoping review completed on 12 July 2026, not a PRISMA or systematic review.                                                                   | Novelty language must remain “to our knowledge.” Before publication, search bibliographic databases systematically, document screening, and perform backward and forward citation chaining.         |
| LIM-12 | Open   | The evaluator records host class and runtime but not thermal state, CPU governor, background load, or e-ink hardware behavior.                                                     | Repeated performance runs may vary. Use a controlled reference machine and device-side traces if timing becomes a decision metric.                                                                  |