Proposal Versus Verification¶
Status: Current table Detection contract
Authority: Current contract (2026-08-12)
Last reviewed: 2026-08-14
Question¶
Why is a model proposal not a verified table?
Short answer¶
A provider only proposes a possible table region. The shared verifier checks finite rules, OCR clues, and negative evidence. Human review then accepts, adjusts, rejects, or adds missing truth. A model score never skips those gates.
Proposal¶
A proposal is a source-labelled box from one provider. Examples include PP-DocLayoutV3, Geometry, PicoDet, Docling Heron, and Microsoft Table Transformer. PP-DocLayoutV3 receives the whole-page Detection raster and returns table-labelled boxes with a source score. Providers remain independently visible. Providers are never automatically fused.
Proposal confidence¶
Provider confidence is a diagnostic only. It cannot amplify its own verification. A high model score does not equal a verified table.
Shared verifier evidence¶
The verifier is provider-neutral. Every whole-sheet proposal uses the same evidence types.
| Evidence | Meaning |
|---|---|
| Finite rule segments | Short horizontal and vertical separators found in ink |
| Measured intersections | Real crossings of horizontal and vertical rules |
| Provisional cells | Early cell candidates before full confirmation |
| Cached OCR compatibility | Existing readings whose centres fall inside the proposal |
| Negative evidence | Drawing-like residual ink, arrowheads, arcs, or an oversized box |
The verifier does not rerun OCR. Sparse OCR support can block automatic verify. Drawing-like or oversized proposals can also fail automatic verify.
Why weak verification stays reviewable¶
Verifier outcomes include verified, uncertain, and rejected. Uncertain proposals remain in the review rail. Rejected proposals can still be inspected before a human decision. The review rail exists to retain reviewable table regions. Reject weak grids, but do not hide the proposal from human truth.
Why the split exists¶
Original morphology heuristics found grids and cells, but false positives remained. External detectors helped proposal quality. The lab still needed one honest verification and review path. That shared path keeps provider comparison fair.
What comes after verification¶
Reviewed regions can enter structure recovery and cell OCR. Understand later assigns type, headers, values, and title-block fields. See Table Review and Later Understanding.
Open Questions¶
- Which negative-evidence rules should become permanent verifier policy across all providers?
Sources¶
- Current contract:
table-dimension-detection-architecture.md(2026-08-12) - Geometry provider role retained from the external drawing-table-detector integration