Sample frameDay 125·AI Fluency · Verification

Good morning. Ask it what your contract says and it'll tell you what contracts usually say. Today you hand it yours.

AI Fluency.

Bring Your Own Evidence

5 min read·Apply by lunch

The question

The last time you asked a model something that mattered, did you hand it the document or ask from memory? And if it was memory, what exactly were you checking the answer against?

The idea

In 2020 Patrick Lewis and his co-authors at Facebook AI Research and University College London published Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, and the architecture it describes is now the ordinary way a business puts a model anywhere near its own material. Their observation was structural. What a model knows is baked into its weights during training: it cannot be updated without retraining, it cannot be inspected, and it cannot tell you where any particular claim came from. So they attached a second memory — a searchable index of documents, retrieved at the moment of the question and handed to the model as part of the input. Answers built that way came back more specific and more factual than the same model working from its weights alone, and each one could be traced to the passage that produced it. The lesson outlives every model release, because it was never about capability. The trap is the resemblance: a recollection and a reading arrive looking exactly alike, same length, same certainty, same tidy paragraphs. Ungrounded, you're asking for a recollection. Grounded, you're asking for a reading, and a reading can be checked.

The framework

01
Which document settles it? Name it before you type.
Every question worth putting to a model has a source of truth somewhere: the signed contract, the current price list, the regulator's guidance note, last month's actual numbers. Name it before you type anything. If you can't name one, you haven't got a research question, you've got a question about the world's general opinion, and a model is a poor and unattributable way to sample that. This step also catches the asks where no document exists and you were hoping one did.
02
Paste the source in, and put a ceiling on the answer.
Grounding is not a mode you switch on; it's material you supply and a boundary you state. Give it the text itself — the clause, the transcript, the export — then say plainly that the answer must come only from what you supplied, and that anything the material doesn't cover comes back as a gap rather than a guess. Hand it the file instead of testing its memory of a company it joined this morning. A good grounded reply gets shorter, and it starts telling you what is missing.
03
Two mistakes survive a perfect grounding.
Ask for the answer with its supporting passage beside it, word for word, and a pointer precise enough to find: clause number, page, the heading it sits under. That pointer is what makes the two remaining failures visible. The first is a passage that genuinely exists and doesn't say what the summary claims, where the words are real and the compression did the damage. The second is a passage that is real and accurate and was never the governing one — the general clause quoted while a special condition three pages later overrides it. Grounding hands you a much shorter list of places a mistake can hide. It won't tell you which of the two you've got.
04
Build the small library before the morning you need it.
Most of what you'll ever need to ground is the same twenty documents: the standard contract, the pricing sheet, the product spec, the policy, the last four board packs, the current financials. Keep them in one folder, each with a version and a date in the filename, and make replacing the stale one part of whoever's job produces it. The failure here is rarely dramatic. It's the enthusiastic colleague grounding an answer beautifully in a document that was superseded in March. Grounded and wrong is more dangerous than ungrounded and wrong, because the clause number stops the next reader checking.

Members only · 4 principles + template + AI mentor

364 more frameworks are waiting.

Bring Your Own Evidence is Day 1 of 365. One framework every morning for a year — across strategy, sales, negotiation, leadership, and more.

$1 /day

Billed $365/yr · cancel any time

365 frameworks

12 topics · templates · AI mentor

Reserve your place — $365/yr →

One payment a year · the whole library · no auto-renewal.

More AI Fluency frameworks

All 20 ai fluency frameworks →