Seven failures in twelve rows. Reading finds two.

Twelve rows of a property register.

Nothing is marked. Every column is populated, every cell was typed by somebody, and the sheet has the shape of a finished thing. Read it and find the failures.

the same sheet appears three times on this page and is never edited

Six rules, written as a list.

This is Sentinel’s shipped ruleset in the order it runs. It lives in one block at the top of a single file, and changing what gets flagged means editing this list. A rule names a column, a test and a severity.

Seven rule types are available: required, range, allowedValues, unique, dateNotStale, regex and custom. Given no rules at all, Sentinel profiles the file instead and flags every blank cell in every column as a warning.

Each rule sees a different sheet.

The same twelve-row grid, drawn once per rule, with only what that rule catches marked. Six rules, six sets of cells, and no cell is marked twice. A rule that goes unwritten finds nothing, and no other rule covers for it.

rows run top to bottom, columns left to right · swipe to see all six

Five clean rows out of twelve.

The same sheet, with every rule run over it. Sentinel reports per cell: the row, the column, the severity, the rule and the reason.

Of the seven, two are blank cells. A person scrolling this sheet finds those and stops. The duplicate identifier needs every row compared against every other, the stale review needs today’s date, the unexpected status needs the agreed vocabulary, and the area above the ceiling reads as a perfectly ordinary number.

Severity is an opinion, and it has to be written down.

Four of the six rules stop the file. Two let it through with a note. A blank identifier is an error, a status outside the agreed vocabulary is a warning, and a review 755 days old is a warning. Somebody decided that, and the list is where the decision is recorded.

Two people can agree on the rule and disagree on whether it stops the run. That argument is the governance, and it is worth having in front of a six-line table rather than after a quarter of reporting.

A model cannot tell you its input was wrong.

Every other tool in this lab reads a file like this one. A scenario engine will rank options on a portfolio holding a duplicated asset. A simulator will run a population with rows missing. The output arrives with the same confidence either way, and the only place to catch it is in front.

Those four are sourced and each is recorded with its status in FACTS.md beside this page. Everything drawn above them is illustrative: the specimen is written to be validated, and the counts come from running Sentinel’s own six rules over it.

Run it on your own file.

Sentinel is one self-contained file. Drop an xlsx, xls or csv on it and the report appears in the browser, or connect a folder once and it re-checks on a timer. It makes no network call with the file. The rules are the part worth arguing about, and there are six of them.

Open the validator →