Notes that survive a chatbot: file your reading so future-you can audit it

A chat transcript is a poor archive. It scrolls, it re-renders, and when you come back weeks later it has the permanence of steam on a window — the words are technically there, and the thread is practically unusable. The third leg of the desk’s working method (after tracing claims and reading past summaries) is the least glamorous and the most valuable: notes that keep themselves honest.

Three layers, never blended

The desk files every piece of research in three layers, and the separation is the whole trick:

• Quote. The verbatim passage from the primary document, with its source link, the page or section reference, and the date it was read. Nothing is reworded in this layer, ever — if the source said “adjusted earnings” the quote says “adjusted earnings”.
• Fact. One plain sentence stating what the quote establishes — only what the words on the page support. If the quote is conditional, the fact carries the condition.
• Interpretation. Your reading of it, clearly marked as provisional and signed with a date, because interpretations age faster than facts and should be allowed to.

When the three layers live apart, re-checking is cheap: open the quote, confirm the fact, revisit the interpretation. When they are blended into a paragraph, every re-check is an archaeology dig.

Models are excellent filing clerks

Here is where AI genuinely earns a place in the workflow. Language models are good at the clerical work of filing: sorting quotes into layers, drafting the one-sentence fact line for you to verify, formatting references, noticing that two quotes probably describe the same event. Hand the model raw text and ask for structure, not truth — then keep the human step of confirming each fact line against the source. The clerical layer is delegable; the responsibility is not.

What a filed note looks like at the desk

One line per claim, kept in a plain list. The format is deliberately dull — dull formats survive software changes, operating systems and decades. Roughly:

[2026-01-15] Q: "adjusted net income rose on a constant-currency basis" — src: FY report, p.47 — F: growth applies only after currency effects are excluded — I: relevant when comparing to peers reporting raw figures (check again vs. French peer's convention)

Five fields, no ceremony. The date matters because documents get revised and pages move; the “read on” date anchors which version you saw.

A working definition of a good research note

The desk’s test: a good note is one where a stranger — or you, six months later, in a different mood — can find any claim, see exactly where it came from, and disagree with the conclusion without re-doing the reading. Notes that pass this test are an asset that compounds. Notes that fail it are a diary at best.