The evidence so far.

Method

• Five wholly synthetic table images, rendered with Pillow and Arial. No personal document was used.
• Original image bytes were submitted in filename order to the Cup's ordinary local HTTP worker, with non-loopback networking denied at the operating system.
• The worker's final grid was scored against the authored expected cells by literal string at each row and column position: no trimming, number conversion, fuzzy matching or row realignment.
• A whole table is exact only when every expected cell matches and there are no missing or extra rows or cells.
• Repeated runs of one image are not new cases, and cells from one image are not independent samples. No percentage or interval is reported.

Not yet tested

• Strong tilt, shadow and blur across a wider range of phone photographs; direct original-JPEG browser upload was not exercised in the photograph check.
• Handwriting, small type, merged or spanning headings, and larger or more complex layouts.
• Any document that is not a clear, upright, tightly cropped printed table in English.
• How reliably people spot and correct errors in everyday use, and which mistakes make a draft difficult to use.

Both photographed tables were exact in both runs.

• Photographed ruled table: 12 of 12 cells exact; no missing or extra rows or cells in either run.
• Photographed unruled table: 12 of 12 cells exact; no missing or extra rows or cells in either run.

Two phone photographs of printed synthetic tables were cropped with assistance and converted to metadata-free PNGs, without resizing, enhancement or perspective correction. Each PNG was submitted twice to the real local HTTP reader using the unchanged model lock shown below. The results were compared with the original authored cells.

This checks these two prepared photograph inputs. It does not establish accuracy on arbitrary photos or independent human correction. Direct upload of the original JPEGs through the browser was not tested, and no new operating-system network audit was performed. The five rendered examples were also repeated: the unruled render still added an empty row.

The model files these results belong to

Results are only evidence for the exact files they were recorded with. A changed model, preprocessing step or runtime needs fresh inference; re-scoring an old recording is not a new result.

Lock

RapidAI/SLANet-plus-PP-OCRv4-det-PP-OCRv6-rec
SLANet-plus-d57a942a_PP-OCRv4-det-1.4.4_PP-OCRv6-rec-3.9.2

Runtime

RapidTable 3.0.2 and RapidOCR 3.9.2 on CPU, in the Cup's pinned Python 3.13 environment

Size

34,323,737 bytes across four files

Licence

Apache-2.0 upstream (PaddleOCR and RapidAI). That identifies the release terms of the software and models, not an audit of their training sources.

Each sentence Tables shows a user is listed with the check behind it and the boundary of that check. “Unchecked” is a visible limit, not a promise. As of 2026-09-11 the deterministic suite is 50 deterministic Python checks, 3 style checks and 32 controlled browser checks. These establish what the software does with recognised text under its documented rules. They say nothing about how often the reader recognises text correctly.

Human use

14 September 2026

I’ve used Tables on 12 everyday tasks and used two exported CSVs afterwards. In a separate seven-task trial, both ordinary Tables exports were exact, but mistakes remained in two prepared correction tasks.

That’s enough for me to keep using and improving it. It doesn’t yet show how well it works for other people, or how reliably they can spot mistakes.