Table Workbench

A table calculator, not a spreadsheet

Watch anyone use a spreadsheet for an ad-hoc question and you will almost never see them type =SUM(B2:B400). They select the column and read the number off the status bar at the bottom of the window. That strip — Sum, Average, Count — is the most used feature in Excel, and it is the only part of a spreadsheet most people need most of the time. This is that strip on its own: a table calculator that is all answer bar and no formula bar, with the two-minute wait for an application to open removed and a few things it gets wrong put right.

Select a column heading, or drag a rectangle across cells, and you get the total, the average, the median, the minimum, the maximum, how many values there are, how many of them are distinct, and how many cells are empty. Every one of those is a button: click it and the plain value is on your clipboard, ungrouped and without a currency symbol, ready to paste into whatever comes next.

Why the sum is exact

Computers store decimals in binary, and most decimals have no exact binary form — which is why 0.1 + 0.2 famously gives 0.30000000000000004. Add four hundred currency values one after another and that error accumulates into something you can see. Spreadsheets hide it by rounding the display, so the number shown and the number stored quietly disagree.

Here, a column whose values have at most nine decimal places is added as whole numbers — scaled up, summed, scaled back down — which is exact by construction rather than approximately right. Anything outside that range falls back to compensated summation, which tracks the low-order bits ordinary addition discards and folds them back at the end. You will not find a total here that is off in the fifteenth digit.

Identifiers are not measures

The famous spreadsheet failures are not arithmetic errors. They are category errors: something spelled with digits gets treated as a quantity. A zip code loses its leading zero. A twenty-digit order number becomes 1.23457E+19 and the real digits are gone. A gene called SEPT2 becomes a date — a problem so persistent that in 2020 the body that names human genes renamed the genes rather than keep fighting the software.

So every column here is classified as a measure or an identifier, and two things settle it: padding that a number cannot carry, and a magnitude larger than 9,007,199,254,740,991, which is the biggest whole number that can be held exactly. Identifier columns show counts and distinct values but not a total, because adding order numbers together produces a number that answers no question. Whatever the classification, the characters you pasted are the characters you see and the characters you export.

Decisions are made per column, and explained

Every judgement is made by looking at a whole column at once, never a cell on its own. That is what makes it safe to recognise $1,234.56, 1 234,56, 45% and (890) as notation: the column as a whole says which reading applies, so no single row can be read differently from its neighbours. Where a column contradicts itself, it stays text and names the row that broke it rather than picking a winner.

The same rule governs dates. 01/02/2026 could be February the 1st or January the 2nd, and nothing in the value decides it. If any row in the column holds something like 13/02/2026, the order is proven — there is no thirteenth month — and the whole column is read that way. If nothing proves it, the column stays text and asks you, instead of sorting your rows into an order that looks perfectly correct and is not. Anything the tool concluded appears in a short list above the grid, so there is never a decision made about your data that you cannot see. If you need the format string for those dates rather than the dates themselves, the date format generator works backwards from an example and applies the same refusal to guess.

Getting the answer somewhere useful

Whatever you have selected can leave as a Markdown table, CSV, TSV, JSON, SQL INSERT statements or an HTML table. Markdown is first for a reason: a summary pasted into Slack, a pull request or a Jira comment renders as a real table inline, where a link would make the reader click through to a site they have never seen. Every format except JSON carries your original characters through untouched, and JSON only differs because it has to commit to a type — where it does, a value becomes a number only when its own digits are already a valid JSON number, so 1.50 stays 1.50.

Everything above happens inside this tab. The parsing, the type detection, the arithmetic and every export are JavaScript running on your machine, so a result grid from a production database can be pasted here without it going anywhere. There is no upload, no account and no server-side step to disclose — open your browser's Network tab and watch nothing happen.