Transfer — and Where the Model Lies

Four lessons in, you own a small arsenal. You know the 37% rule and why the number 1/e falls out of it. You know the deeper shape underneath it — explore-then-commit, a look phase that builds a ruler and a leap phase that uses it. You know the expected-value / descending-threshold cousin, where the bar drops as your runway shrinks. And you know the real-world patches from Lesson 4: recall, reservation values when you know the distribution, search costs, unknown N, and offers you can choose to reject.

Now we do the two most important things you can do with any model: carry it somewhere new, and find out where it lies to you. A model you can’t transfer is trivia. A model whose limits you can’t name is a superstition. This lesson makes the secretary problem into a genuine thinking tool — and then, with equal energy, tells you every place it quietly breaks.

Hiring — calibrate the bar, then take the first who clears it

You’re filling one role. Candidates arrive in a stream, you interview them one at a time, and — in the idealised world — each one wants an answer before the next walks in. The naive move is to hire the first person who seems “good.” But good is meaningless on the first candidate: good compared to what? You have no scale yet.

So run explore-then-commit. Spend the opening slice of your pipeline — call it the first third — interviewing hard and hiring no one. These candidates are not applicants; they’re instruments. Their job is to teach you what “strong for this role, this budget, this market” actually looks like. Note the best one you see. That becomes your bar. Then switch to commit mode: hire the first later candidate who clears the bar.

Worked version: you’ll see roughly 12 candidates before the req expires. Interview the first 4 and hire none — but log that candidate #3 was the standout. Now from #5 onward, hire the first person who is clearly better than #3. If #7 beats them, #7 gets the offer, and you stop.

But hiring is where you must not run the pure rule, and it’s worth being honest about why:

• The objective is wrong. The classic rule maximises the chance of landing the single global best candidate and gives zero credit for hiring the second-best. In real hiring you don’t need the platonic best engineer alive; you need someone great. That’s the expected-value version from Lesson 3 — a superb #2 is a win, not a failure.
• Recall usually exists. People rarely vanish the instant you interview the next one. You can very often go back and make an offer to a strong earlier candidate. The moment recall is on the table, the strict “reject the first third forever” rule softens — a fantastic candidate #2 can be revisited if the rest of the field disappoints.

So in practice: use the first third to calibrate, keep a shortlist you could recall, and take the first strong-enough clearer — while staying willing to loop back if the pipeline craters. That’s the model with its real-world guardrails on.

Apartment / house hunting — the viewings are reconnaissance

Flat-hunting is the secretary problem wearing a coat. Listings appear over time, you tour them one by one, and in a hot market a place you love is gone by tomorrow — a near-perfect model of sequential, irreversible choice.

The rookie error is treating the first viewing as a real candidate: you love the light, you make an offer, and you’ve committed with a sample size of one. You have no idea whether this is a steal or the worst place you’ll see all month, because you’ve never seen “the rest.”

Instead: the opening viewings are reconnaissance, not choices. Decide roughly how long you’ll search or how many places you can realistically tour, spend the first third of that touring and bidding on nothing, and let those viewings set your bar — layout, light, commute, price-per-square-metre. Then flip to commit mode and pounce on the first place that clears the bar.

Worked version: you give yourself six weekends and expect to tour about 15 places. Weekends 1–2 (roughly the first third, about 5 flats) are pure calibration — you bid on none but you learn that “good” in your budget means an eat-in kitchen and under 30 minutes to work. From weekend 3, you make an offer on the first flat that beats everything in that opening sample, and you do it fast.

Why the pure rule fits house-hunting better than most domains: in a genuinely fast market, recall is near zero. The place you dithered on is under offer before you’ve parked the car. When recall really is ~0, the strict “don’t go back” assumption is close to true, so the clean explore-then-commit rule earns its keep instead of being an approximation.

Selling a house — a reservation value that DROPS

Flip to the other side of the table. Now you’re selling, offers arrive over time, and each buyer wants a yes-or-no before they walk. This is the reservation-value flavour from Lessons 3 and 4, and it behaves differently from pure calibration because here you often do know the distribution — the market told you what comparable homes fetched.

Because you know roughly what to expect, you don’t need to burn a third of your buyers just to learn the scale. You set a threshold — a reservation price — from the comps, and you take the first offer above it. No global-best hunting; a good-enough offer that clears the line is a sale.

The twist that makes this its own case: the threshold should descend. Every week the house sits unsold you pay a mortgage, taxes, insurance, and the slow ache of a listing going stale — carrying costs. As those mount and your runway shortens, the offer you’d accept should drop. That’s the descending threshold from Lesson 3/4 made concrete.

Worked version: you list hoping for $500,000 and set your reservation there. Week 1 you’d turn down $480,000. But each month of carrying costs you roughly $4,000, the listing is cooling, and buyers smell it — so you lower the bar to $485,000, then $470,000. The offer you’d have scorned in week 1 becomes a rational yes in week 8. Holding out rigidly for the top number while the meter runs is how sellers lose money to their own stubbornness.

Dating and the pop-science “37% rule”

Here’s the version that made the secretary problem famous — and the version most abused. The pop-science pitch: date around until you’re 37% through your “dating window,” then marry the next person who’s better than everyone you’ve dated so far. It’s catchy, it’s mathematically descended from a real theorem, and taken literally it’s terrible life advice. Let’s be brutally honest about why, because the honesty is the lesson.

The model quietly assumes five things that dating flagrantly violates:

• People aren’t a fixed pool in random order. The theorem needs a set, known number of options presented in a uniformly random sequence. Who you meet is filtered by your city, job, apps, and friends — anything but random, and correlated with your past choices.
• You don’t know N. Your “dating window” isn’t a number handed to you at 18. Is it 10 people? 40? The whole rule pivots on N, and you genuinely don’t have it.
• Recall exists — vividly. The model forbids going back. Reality has exes, second chances, “we reconnected years later.” The no-recall assumption is simply false here.
• It’s two-sided. In the secretary problem you choose. In dating, the other person chooses too — and can say no. The math models a one-sided pick; romance is a matching market.
• The objective is wrong. The pure rule optimises the probability of the single global best partner and treats an incredible near-match as a total failure. That is a genuinely insane way to pick a life partner. You want expected happiness, not a lottery ticket on the one theoretical maximum you’ll never verify anyway.

So what’s it actually good for? As a nudge, not an algorithm. It pushes back on two real failure modes at once: under-searching (marrying the first person you date because you’ve never seen the distribution) and over-searching (the perpetual optimiser who won’t commit to anyone because someone better might exist). The honest takeaway is directional: date enough to calibrate what you value, then be willing to actually commit to someone excellent — and stop hunting for a mythical perfect stranger. That’s wisdom. “Compute 37% of your dating window” is numerology.

When to stop researching and DECIDE

The quietest, most useful application isn’t romantic or financial — it’s knowing when the look phase ends. Every explore-then-commit decision has a failure mode called analysis paralysis: you keep gathering information, keep “just checking one more option,” and never cross into commit mode. The model’s deepest advice is almost aggressive: the look phase must END. Exploring forever isn’t diligence; it’s a decision to let the options expire while you watch.

This shape is everywhere once you see it:

• Parking near a venue. You can’t circle back — pass a spot and it’s gone (no recall). So drive the first stretch noting how full it is (calibrate), then grab the first spot within an acceptable distance. Circling for the theoretical closest space costs you the time you were trying to save.
• Choosing a restaurant while travelling. You’re hungry, walking a strip, and every place you pass is behind you. Scan a few blocks to learn the range (calibrate the bar), then commit to the first one that clears it. The alternative — walking the whole strip twice — means eating late and cranky, having “optimised” yourself out of dinner.
• When to stop A/B testing. Running a test forever to chase statistical perfection has a real cost: the search itself is expensive (Lesson 4’s search costs). At some point the expected value of more data is less than the cost of the delay — so you stop and ship the winner. Endless testing is analysis paralysis with a dashboard.

Where the model lies

Time for the uncomfortable part. The secretary problem is a masterpiece of idealisation — which is exactly why you must know which idealisations it smuggles in. Every clean assumption is a place the model can lie to you if you apply it straight. Here they are, named honestly.

Three of these deserve a closer look.

The objective is usually wrong

This is the big one. The pure 37% rule maximises the probability of catching the single best option and assigns zero value to a near-perfect second place. For almost every real decision — hiring, housing, partners, products — that objective is simply not what you want. You would be overjoyed with the second-best apartment; the model would call that a total loss. Whenever a wonderful runner-up would make you happy, you’re in expected-value territory, not best-only, and you should optimise for how good the outcome is, not the odds of a theoretical maximum you can never even confirm you hit.

Recall changes everything

The whole drama of the secretary problem — the tension, the regret, the elegance — comes from you can never go back. Remove that and much of the rule evaporates. If you can recall a past option, there’s little penalty for exploring longer, because a great earlier choice isn’t lost. Before you apply the strict rule, ask the most important diagnostic question in this entire course: can I actually go back? If yes, relax. The rule is at its sharpest and most trustworthy precisely when recall is genuinely near zero — a fast housing market, a one-shot auction, a train you literally cannot re-board.

Don’t fetishise the number

37% is a gorgeous, exact answer to a fictional problem — a clean world with known N, random order, no recall, and a best-only goal. The real world satisfies none of those exactly. And here’s the mercy: the success curve is a broad plateau, as the lab above shows. Looking at 25% or 45% instead of 37% costs you almost nothing. So the transferable rule was never the decimal. It’s the shape — calibrate on the opening stretch, then commit — with “about a third” as a rule of thumb, not a commandment carved in e.

Sort the decisions: does the rule even apply?

The single most valuable skill from this whole course isn’t computing 1/e — it’s diagnosing whether you’re even in secretary-problem territory. The rule applies cleanly only when a choice is genuinely sequential, irreversible, and no-recall. The moment you can wait, compare side-by-side, or go back, the clean rule stops fitting. Sort these.

The practical checklist

Strip away the elegance and here’s the tool you actually carry out of this course. Six steps, in order.