Teaching while you ask

Take any teach-back line on your sheet and read it aloud to someone who does not do data work. Any word that makes them say "sorry, what do you mean?" is a jargon cost, and session 3's bench prices it: every jargon question in the bank takes five minutes, because you have to explain, and costs a point of trust, because she is a little less sure you understand her shop. The line was meant to earn trust. With one wrong word it spends it.

The scientist's usual offenders, with the plain word that replaces each. Model: the thing that makes the guess, or simply "the forecast". Feature: one of the facts the guess looks at, like last month's sales or the supplier's lead time. Label: the answer we already know for the past, here a record that a shelf was empty on a given day. Precision: when it says a shelf will run out, how often it is right. Recall: of the shelves that really ran out, how many it warned you about. Threshold: how sure the guess must be before the van moves stock. Class imbalance: the thing we are hunting for is rare, most shelves are fine most weeks. Backtest: pretending it is last March and checking what the guess would have said. None of the plain versions is longer than a breath, and every one is a thing she already knows exists.

The test is not "would a smart person understand this". Adeline is a smart person who runs 38 outlets. The test is whether the word is one she uses at work, and the honest way to find out is her face, which is why the reading is done aloud and never on paper.

Notice where the line falls. "Predict" is her word: she used it in her first sentence, so a teach-back line can say "prediction" freely. "Model" is yours. Six of the twelve teach-back lines in the build-along say "model", and they were written by someone trying to be plain. It is the scientist's hardest word to hear, because it is the word of the whole job. To her it is either a thing on a catwalk or a thing a consultant sells, and neither is the idea you mean: a set of rules learned from the past that makes a guess about the next week.

There is also a professional reason plain is the norm and not a courtesy. Douglas Hubbard's definition of measurement, a quantitatively expressed reduction of uncertainty based on one or more observations, is itself a plain sentence with no term of art in it. The person paying for the work has to be able to check it. A teach-back line she cannot decode is a claim she cannot test, and a claim she cannot test is one she either accepts on faith or quietly ignores. Faith runs out on the first empty shelf the forecast missed; ignoring means you spent a sentence and taught nothing. Plain words are how a scientist makes their reasoning inspectable by the person who has to sign it off.

Look at the five false beliefs the bench records on the fast run, because each one is a concept that was never explained. "The target is sales" attaches to what exactly is being predicted. "Daily predictions" attaches to the day the decision is made, and the van runs on Thursday. "Stock history is in the POS" attaches to what has been recorded, and the stock sheet is overwritten every Friday. "Optimise for accuracy" attaches to the two wrong guesses costing different amounts. "The model replaces the pharmacist's check" attaches to the baseline, the person looking at a shelf that any forecast has to beat. Pick the belief most likely to form in this meeting with this person, and the concept behind it is your one.

For Harbourline there are two honest candidates. The two wrong guesses cost different amounts, because "accuracy is what matters most, yes?" is the belief waiting to happen: the fast run records it at minute 5, it is the number every scientist is trained to report, and only later in a good meeting does she say she does not care what the accuracy number is. Or nothing has been recorded, because she assumes the history is in the POS and there is no stock history at all. Either is right; session 2 spent a whole evening on the second. Choose one, not both; two concepts in ninety seconds each is a lecture with an interval.

Three sentences, in a fixed order: what it is, why it matters, what it means for her Thursday. Written in full, as it would be said after "when it happens, what does it cost you, and what does it cost when an outlet orders too much?" and her answer that a missed stock-out sends a patient to a competitor and a complaint to head office, while an over-order becomes expiry write-off, which head office measures her on:

What it is. "Any forecast of empty shelves will be wrong in two ways: sometimes it warns about a shelf that would have been fine, and sometimes it stays quiet about one that goes empty, and those are two different mistakes, not one." Why it matters. "You have just told me they cost different things, a patient and a complaint on one side and write-off on the other, so a single accuracy score hides the only thing you care about, because a forecast that never warns anyone is right almost every week and catches nothing." Her Thursday. "So before I build anything, you and I decide how many false alarms a missed empty shelf is worth, and that decides which outlets get stock on Thursday's van."

Ninety seconds, then back to the next question. The tell that it worked is that later in the meeting she talks about "a missed one" and "a false alarm" as two things, rather than asking how accurate it will be.

Session 1 did the arithmetic on the constructed sheet: about 40 chronic-medication products in 38 outlets is 1,520 outlet-product pairs, and roughly 30 stock-outs a month is about seven a week. In any week, well over 99 percent of pairs do not run out, so a forecast that says "no" to all of them scores above 99 percent accuracy and prevents no complaint calls. That is the reason "accuracy is what matters most, yes?" is dangerous rather than merely vague: it is a leading question she will agree to, and the model built on her agreement is the one that looks best and does nothing.

The ratio between the two costs is hers to set, not yours. The bank gives no number for it, on purpose: a scientist who invents "a missed shelf is worth ten false alarms" has made a business decision in a notebook. Writing the costs down as a table and deciding with them is a discipline in its own right, and the Decision Intelligence course owns it. Tonight's job is smaller: make sure she knows there are two mistakes before anyone asks her to agree to one score.

The board meets in six weeks and wants one slide. Somebody will ask Adeline how accurate the AI is. If the one concept landed, she can say something like "we are not chasing an accuracy score; it is tuned so Thursday's van reaches the empty shelves that turn into complaints." That sentence is the whole test: it is hers, it is short, it carries the reason, and it would survive a sceptical follow-up. If instead she says "they talked about precision and recall", the concept was heard and not learned.

Write the sentence you hope she says before the meeting. It forces the ninety seconds to be built backwards from a line a non-data person could carry, and it is the cheapest rehearsal there is. It also tests the choice itself: if you cannot write a one-line version a board member would repeat, you have picked an idea that matters to you and not to her.

Failing, before "when it happens, what does it cost you, and what does it cost when an outlet orders too much?": "Because the positive class is rare, raw accuracy is a misleading metric, so I would want to set the decision threshold on an asymmetric loss, weighting recall above precision, and validate it on a backtest before we talk about deployment." True, and it fails all three tests at once: six words she would ask about, three claims joined into one, and a lecture on evaluation hiding inside a reason for asking.

Passing, same question: "Guessing wrong one way costs you a complaint and wrong the other way costs you write-off, so I build the guess around which one you would rather have, not around one score." Plain, because every word is one she uses; true, because it is the principle the failing version was reaching for; one sentence, so she is answering again within three seconds. The failing version would have earned a nod. The passing version earns her answer about head office and the write-off it measures her on, which is the fact the whole model has to be built to.

There is only one proof that a line taught something, and it is a follow-up question from her. "So a forecast could be right nearly every week and still miss the ones that matter?" means the two costs landed. "Wait, so it cannot learn stock-outs if nobody wrote them down?" means the recording concept landed. A nod proves nothing, and a polite "makes sense" proves slightly less than nothing, because it closes the topic. Silence after a teach-back line is a reason to ask your next question, not to add a sentence.

On the bench this is the mechanism, not a metaphor. Every open question earns a point of trust, and six facts on her sheet sit behind trust gates she will not cross unprompted; the most sensitive, that the only IT person holds the POS login and is leaving in January, needs trust 3. The teach-back line is where that point is earned: she answers the next question more fully because the last one came with a reason. A lecture spends minutes and earns nothing, which is the same arithmetic as the technical opener in session 3 with a friendlier face.