Habits of Verification

Two pages of bad news deserve a page of method. Tessa cannot fact-check every sentence the model writes — she has forty itineraries a month and a job that is not research. If verification means re-doing the work by hand, nobody will do it, and the tool goes back to being a trap.

It does not mean that. In practice, catching most of the damage takes a handful of cheap habits: find the fact the decision rests on, ask for the form of a claim you can look up, and let the stakes decide how hard you look. Verification is a reflex to install, not a research project — and installing it takes one page.

The answer arrivesmostly filler, in the good sense

→

Spot the load-bearing factthis hotel exists and takes bookings

→

Check it — stakes decide how hardown notes · client-facing · contractual

→

Use, fix, or drop itone search, one absence, one hotel gone

Find the Load-Bearing Fact

Most answers are mostly filler, in the good sense. A hotel recommendation is a paragraph of atmosphere resting on one claim that carries weight: this hotel exists and takes bookings. A summary of a refund policy is several sentences resting on one number: fourteen days. A draft email is entirely harmless except for the date it promises. Find that piece — the load-bearing fact, the one that makes the rest either useful or embarrassing — and check it.

Two minutes on the hotel's own website was the entire Harbourview defence. Not a review of the model's reasoning, not a second opinion, not a policy. One search, one absence, one hotel dropped from the itinerary. The skill being learned here is not checking; it is spotting — reading an answer and seeing which two sentences are holding the weight.

Ask for the Checkable Form

The second habit takes one extra message: where could I verify this? or what is this based on? Be clear about what that does and does not do. It does not make the model honest, and it does not produce a guarantee — a source can be invented as fluently as a hotel, complete with an author, a publisher and a year that all look right.

What it buys you is a shape you can act on. A vague paragraph turns into named claims with named places to look, and looking becomes a thirty-second job instead of an open-ended one. There is a bonus effect worth knowing: asked to point at where a claim comes from, an invention sometimes visibly deflates — the answer softens, generalizes, or quietly drops the specific. That is not proof of anything, but it is a useful flinch to watch for.

Route by Stakes

The reason blanket rules fail is that they ignore what an answer is for. Checking everything is impossible; checking nothing is how the Annex reaches a client. So route by consequence, in three lanes.

Thinking for yourself — brainstorming tour names, drafting your own notes, getting unstuck: read it and go. If it is wrong you will find out cheaply, and nobody else is affected. Anything client-facing or money-adjacent — an itinerary, a quote, a policy summary sent to a guest: check the load-bearing facts before it leaves your hands. Anything contractual, legal, medical or regulatory: the model may draft, and a qualified human owns every fact in it, with no exceptions for a busy afternoon.

That is proofreading logic, and you already have it. When a colleague's report crosses your desk before it goes to a client, you do not re-research it. You check the two numbers the decision hangs on and read the rest with one eyebrow raised. Same reflex, same effort, new colleague.

The Second Run as a Smoke Test

One more trick, cheap enough to use constantly. Ask the same specific question again in a fresh chat — not the same conversation, where everything already said is in view and shapes what comes next (Chapter 1), but a clean start.

Chapter 1 also explained why the answers will not be word-for-word identical: there is genuine variability in how the text gets built. So read the result carefully. A stable answer across fresh runs is not proof — a firmly-learned pattern is repeatable whether or not it is true. But wildly different answers are a loud signal that there is no firm pattern underneath the question at all, and that is exactly the territory where inventions live. Cheap test, one-directional result: it can raise the alarm, it cannot clear you.

Four habits, then, and none of them takes more than a couple of minutes: spot the load-bearing fact, ask where it could be confirmed, route by stakes, and re-ask when something feels assembled. Between "never trust it" and "trust it blindly" there is a calibrated middle, and that middle is where all the real productivity turns out to live. Stakes-routing in particular comes back as the decision frame of Chapter 10; you are learning its everyday version now.