“Should I put this in the junk drawer?”
“I dunno. Do you think you could figure out where it goes?”
That exchange is short enough to miss, and it is the entire subject of today’s post. The first post in this series introduced the junk drawer as a load-bearing component of any well-designed classification system, home to the lost key (waiting on information that hasn’t arrived) and the screwdriver (useful across categories, which is the point). Quasi argued that a system refusing to designate a location for acknowledged ambiguity does not eliminate ambiguous objects — it distributes them, quietly contaminating every labeled category in the taxonomy. Both posts treated the drawer as a place. Today I want to treat it as a process. Objects enter the drawer, objects sit in the drawer, and — this is the part we haven’t talked about — some objects ought to leave. A few ought never to have entered at all.
Return to the moment of the opening exchange. You are standing in the kitchen holding an object that resists filing, and you face a menu with three options on it, not two. You can classify the object as best you can: commit it to the most plausible labeled drawer and accept the risk, described in the first post, that a confident wrong guess makes the object harder to find than honest uncertainty would have. You can defer: place it in the junk drawer, paying a little space now and a little classification effort later. Or you can discard it: the trash can, which costs nothing today and charges everything to option value — if the lock this key opens ever surfaces, no free Saturday will bring the key back. Standard decision theory has nothing exotic to say here: rank the three options given your beliefs about the object and choose the best one. The interesting material is all in the second option’s fine print.
Here is the fine print. The junk drawer is not a destination. It is a waiting room, and placing an object there is a bet — specifically, a bet that the object will be classified in finite expected time. Someone will eventually open the drawer, recognize the thing, and carry it home to the labeled compartment it was always waiting for. That is what deferral means: not “this object has no home” but “this object has a home I cannot compute tonight, and somebody, someday, will compute it.” The opening dialogue is the drawer’s admission interview, and the norm it enforces follows from the definition of the bet: do not place a bet you strongly suspect you cannot win. If you believe the object will never be classifiable — not “not tonight,” not “not by me,” but never — then the drawer is not for it, however uncomfortable the trash can makes you feel. Think of the norm as paying it forward, in spring-cleaning terms. Every object you wrongly admit is a small recurring tax on everyone who will ever open the drawer looking for something else, future-you prominently included.
Why must the norm operate at the door, on suspicion, rather than inside the drawer, on evidence? Because never is not an observable event. An object that has sat unclassified for five years is observationally consistent with “not yet” and with “never,” and no additional amount of sitting will separate the two. A lost key vindicates its admission only by exiting — the information arrives, the object goes home, the bet pays off. An object that was never classifiable produces exactly the same data as a key whose information simply hasn’t arrived: silence. The drawer can never convict its own contents. If the unwinnable bets get in, there is no principled moment at which they get out, which is why the screening happens on beliefs, at admission, or it does not happen at all.
The silence, however, is not uninformative. Consider how the bet pays off. Every so often, someone opens the junk drawer — looking for tape, looking for a battery, looking for nothing in particular — and every visit carries some positive probability that the visitor recognizes the mystery object and files it where it belongs. Maybe they know what the key opens; maybe they finally remember which appliance the cable came with. A genuine lost key facing this search process exits the drawer in finite expected time. That is the bet paying off. But run the logic in reverse. Each day an object survives in the drawer is a day of failed recognitions, and failed recognitions are evidence. Bayes’ rule does the rest: the probability that the object is a lost key rather than junk declines steadily with the length of its residence.1 No news is bad news. The drawer quietly testifies against its longest-tenured occupants every day, and nobody hears it, because the testimony consists entirely of things not happening.
This yields the eviction rule for free. The drawer is finite, and if genuinely unclassifiable objects arrive at any steady rate, the drawer will eventually need cleaning out — that is arithmetic, not pessimism. If the objects all look alike to the search process, then residence time is the only evidence there is, and it is all the evidence you need: the longest-tenured object is the likeliest junk. Throw out the oldest thing first. The rule sounds like impatience and is actually inference — a statute of limitations derived from the fact that the crime of being unclassifiable can never be proven directly. (The screwdriver is exempt from all of this. It placed no bet. It is a resident, not a transient; nothing about it is waiting to be learned, and the clock does not run on residents.)
Now the subtle point, and the reason this blog has the subtitle it has.
Suppose a reformer surveys this whole arrangement — the bets, the silence, the evictions — and proposes the obvious clean solution: abolish the junk drawer. Every object gets classified or discarded at the door; no waiting room, no backlog. What does our long experience of the drawer-having household predict about the drawer-less one? Almost nothing, and the reason deserves careful statement. Watching someone choose from a menu reveals that the chosen option beat each alternative. It reveals nothing about how the alternatives rank against each other. When the household put the mystery key in the junk drawer, you learned that deferral beat classify-your-best-guess and that deferral beat the trash can. You learned nothing about whether classify-your-best-guess beats the trash can — and that is exactly, and only, the comparison that governs what happens once the drawer is gone.
Count the problem. With two options on a menu, removing the chosen one is perfectly predictable: the survivor wins by default. With three options, there are three pairwise comparisons, the observed choice pins down two of them, and the one left unidentified is precisely the one the reform turns on. Three is the smallest menu at which a system’s own records cannot predict the system’s behavior under reform. Two implies order. Three implies chaos — and this time it isn’t even a voting result; it is bookkeeping. Macroeconomists know the general lesson as the Lucas critique: data generated under one regime cannot forecast behavior under another, because the data encodes choices made against the menu that actually existed. It applies just as well in the kitchen, and it is at the center of the audit questions Maggie and I are wrestling with in the book we’re writing now: an auditor who evaluates a classification system from its decision records is observing winners, never rankings, and reforms live entirely in the rankings.2
This requires an honest correction to Quasi. That post predicted what a drawer-less system looks like: ambiguous objects forced into labeled categories, every drawer in the house a little contaminated. The prediction quietly assumed that filers denied the deferral option fall back to classifying rather than discarding. (Ed.: I believe I raised this objection at the time, John, though I notice nobody wrote it down anywhere — which I gather is now the official theme of the enterprise.) Nothing in the drawer-era data licenses that assumption. Maybe abolition produces contaminated categories. Maybe it produces clean categories and a quiet hemorrhage of option value — locks that will never open because their keys went out with the trash the same week the drawer did, a loss that generates no clutter and no visible evidence at all. Both futures are fully consistent with everything we observed while the drawer existed. Quasi described one branch of an unidentified counterfactual and called it a forecast.
So the drawer, properly run, is a waiting room with an admission interview at the front and a statute of limitations at the back, and the data it generates along the way is silent on the one question reformers most want it to answer. But notice what the eviction rule assumed without saying so. Somebody has to do the evicting. Somebody has to notice the drawer is full, reconstruct residence times that nobody wrote down, and pay, personally and today, the cost of a cleanup whose benefits accrue to everyone who ever opens the drawer from now on. Nothing in the drawer reminds anyone to do this, and nothing in the household rewards whoever does. If you have ever cleaned out a junk drawer and heard yourself ask why you are the only person who ever does, you have already discovered the problem, and it is not a problem about drawers. It is the next post.
With that, I leave you with this.
Notes
1 The “all objects look alike to the search process” assumption is doing real work here. If objects differ in how easily a visitor recognizes them — a key to a rarely visited storage unit gets fewer chances at recognition than a key that might fit the front door — then a long residence is weaker evidence of unclassifiability for some objects than for others, age stops being a sufficient statistic, and the throw-out-the-oldest rule starts evicting slow-burning keys along with the genuine junk. Symmetry buys the clean rule; heterogeneity is where the interesting mistakes live. ↩
2 The “more structure” that would close the gap has names. Decision theorists will recognize the inert case — removing an option nobody chose changes nothing — as contraction consistency, and the assumptions that let you extrapolate from chosen options to full rankings (independence conditions, random-utility structure) are exactly the assumptions that discrete-choice econometricians invoke and that the red-bus/blue-bus problem exists to punish. The point of the post is not that the extrapolation is impossible; it is that the data alone never supplies it. ↩