Research and Policy Analyst · EI MAR Automation
Diagnostic, then release
Every table is audited automatically, cell by cell and in both official languages, before it is released. A diagnostic run lists what it finds and releases nothing. A release run goes ahead only when every check passes.
A page, drawn without data
Three errors to catch
A table shows three kinds of error the audit is built to catch. It is drawn without data.
The audit reports · 1
Two digits swapped
The page and the data disagree in one cell.
The audit reports · 2
A total that does not add up
The cells of a column do not add up to its total.
The audit reports · 3
English and French differ
The same cell reads differently on the two pages. Every other cell matches what the data gives.
Diagnostic run
Lists every difference
Nothing is released. Each difference comes to me to examine.
Release run
All checks pass, or nothing is written
When it passes: the documents, and a receipt of every check.
Every cell, both languages
The program rebuilds the table from the raw data and compares each cell with the page, in English and in French. It also checks that totals add up and that tables sharing a figure agree.
Each difference gets a ruling
I examine every difference and record its cause, so the same question is not asked twice.
All checks pass, or nothing is written
A release run writes the documents only when every check passes. If one fails, it writes nothing, so a page that has not passed cannot be mistaken for one that has.
In more detail. The gate fails closed. A release needs every named condition to hold, among them: the inputs are present, no difference is left without a ruling, totals and the figures shared between tables agree, and the English matches the French. A refused release leaves the output folder exactly as it was. A passing one writes the documents with a receipt that binds each file by its SHA-256 fingerprint.
An explanatory reconstruction of the automated audit, drawn without data.