The workbench reports observed structural differences between two files. It does not certify that an import will work, that data is correct, or that anything is safe to release. This page lists the boundary explicitly so nobody has to infer it.

No compatibility certificate

A review with zero blocking findings means no encoded structural finding was raised, and nothing more. Rare values outside the sampled rows, locale and formatting conventions, business meaning, and transformations applied after export all remain invisible to a structural profile.

Inference is observational

Types, blank rates, distinct counts, and widths describe the file in front of the parser. In the bundled sample quantity is reported as integer in August purely because all four sampled rows held whole numbers. Had the producer emitted 2.5 one month earlier, the same column would have profiled as number from the start and the September finding would never have appeared.

No privacy or regulatory classification

The workbench does not detect personal, health, financial, or otherwise regulated data, and it does not tell you whether a column may be shared. Handling obligations are yours and unchanged.

No migration generator

A committed mapping records reviewer intent. It does not rewrite a file, transform a column, deploy code, or configure a downstream system.

No execution of consumers

Consumer policies are descriptions written by people. Nothing opens the workbook, runs the import, or checks the description against reality. A wrong policy produces a confidently wrong review.

Bounded by design

File size, row count, column count, cell count, and text length are all capped, and files past a cap are refused rather than truncated. Distinct counts stop at a bound and report "at least". These limits keep an in-browser review predictable, and they mean very large exports are outside the tool's scope.

What it is good for

Given those boundaries, the workbench is useful for exactly one job: making a recurring handoff's structural change visible, attributable, and reviewable before somebody's spreadsheet silently changes its answer.

Do this in the app

Open the workbench, and read the limitations page bundled inside any exported packet: the same boundary travels with the evidence.

Related reading