REFERENCE 09 / 14 · WORKING METHOD · 6 MIN

Loading the window

The model reads signal, not decoration — the sentence is 37 bytes; the Word file around it is 36,607.

PLATFORMWhich product’s names this page uses — BOTH shows every name.
TOY — LOADING…
TOY — LOADING…

The autopsy above is a real measurement, run live on the bundled files: the same 37-character sentence saved as Markdown and as Word. The Markdown file is 37 bytes — 100% signal. The Word file is 36,607 bytes that unzip into a pile of XML, and the sentence is about 0.1% of it. Almost everything else is packaging — styles, themes, namespaces — plus the property fields below.

Two consequences, one practical and one that should raise an eyebrow:

  • Paste for control. Pasting is the only way to see exactly what you hand it — no extractor deciding for you, no property fields riding along. Convert to Markdown for anything you'll reuse. When you do attach — structure, length, a format worth keeping — check what came through before you build on it. Finance translation: hand over the trial balance itself, not the deck it was printed in.
  • Know what leaves the building. Open the metadata panel in the autopsy — a Word file quietly ships property fields (creator, last-modified-by, revision count, timestamps), filled in unless someone stripped them. Attach a client file to anything and whatever they hold rides along. That isn't trivia; it's a leak surface.

The tokenizer toy is where the budget lesson actually lives: the identical table costs wildly different amounts as pixels versus text — that difference is measured in tokens, the window's own currency — and the text version is the one the machine can actually reason over cheaply.

Say it back in one line: what does a Word attachment carry that the sentence inside it doesn't?