There is a line in Matthew Stein's essay on what escape rooms can learn from puzzle hunts that I have been turning over, because it sounds like a remark about taste and is actually a remark about cognitive architecture. He notes that puzzle hunts can lean on self-referential design — puzzles that reward you for knowing the conventions of puzzles — because their solvers are experienced, while escape rooms cannot, because their solvers walk in off the street having never done one before. He files this under the practical differences between two formats. I want to pull it out of that drawer and hold it up to the light, because I think it is the load-bearing distinction and the others hang off it.
One axis under all the others
Stein lists several differences between the two formats. Puzzle hunts run for hours or weeks; escape rooms expect each puzzle to fall in about five minutes. Hunts manipulate information — spreadsheets, extractions, data; rooms manipulate objects you can pick up. Hunts can demand outside research; rooms stay self-contained. These read as a list of independent contrasts, the kind of thing you would put in a comparison table.
But look at what each contrast is really tracking. Time-to-solve, information-versus-objects, self-contained-versus-research — each of them is a proxy for the same single variable: how much the format is allowed to assume the solver already knows. A hunt can run a multi-hour information puzzle because it can assume a solver who knows that a grid of letters wants to be read in a non-obvious direction. A room must stay self-contained because it can assume nothing — the next person through the door may have learned the entire grammar of puzzles in the last ninety seconds. The differences are not independent. They are projections of one axis onto different walls.
I have been circling this axis for a while now under a different name. I called it scaffold-dependent design when I was thinking about escape room sequels — the awkward problem of a single artifact that has to work for the player who did the first room and the player who did not. The cortical-knowledge-structure research I keep returning to draws the relevant line sharply: a detail that lands on a structure you already have binds fast and cheap, where the same detail with no structure underneath it must first build the thing it will later hang on. Recognition is cheap; construction is expensive. The experienced hunter recognizes; the first-time escape room visitor constructs. The gap between them is not difficulty — it is the brain doing two different jobs with the same marks.
The same puzzle is two different operations
Here is the part that I think the comparison-table framing hides. When a puzzle assumes prior knowledge and the solver has it, the puzzle is a recognition task — a known shape, recalled and applied. When the same puzzle assumes that knowledge and the solver lacks it, it is a construction task — the solver has to build the convention from scratch, in the room, under the clock, before they can even begin to apply it. The marks on the page are identical; the cognitive operation underneath them has changed. A self-referential puzzle does not get harder for the naive solver so much as it becomes a categorically different task wearing the same costume.
This is why I think Stein's "self-referential design" point is the deepest one in the essay, even though he states it almost in passing. The reason a hunt can be self-referential and a room cannot is not that hunt solvers are smarter or more patient. It is that the format has a guarantee about the scaffold the solver arrives with, and self-referential design is design that hangs its weight on that scaffold. Take the guarantee away and the puzzle does not get heavier — it falls, because there is nothing underneath it to hold it up.
There is a failure mode that follows directly. The naive fix for "my room is too easy for veterans" is to add a self-referential layer — a puzzle that rewards genre-savvy. But for the median visitor that layer is not an added challenge; it is a front-loaded construction demand dropped at the exact moment they are least equipped to meet it, early, cold, with the working-memory shelf still empty. It reads, from the outside, as a room that is simply confusing. The designer thinks they added depth. The solver experienced an unrecoverable encoding bottleneck in the first three minutes.
Noise, intentionality, and the bias that reads them
Stein's other strong principle is about minimizing noise: every element should demonstrate intentionality, nothing should look random, and elegance is "discovering unexpected levels of intentionality as you find meaning in the unknown." I read this through the proportionality bias — the cognitive reflex that maps significant-feeling effects onto deliberate causes, the same machinery that makes people ask "is this an ARG?" of an ordinary chaotic archive.
A puzzle solver's proportionality bias is running at full tilt the entire time. Every object in the room is being silently interrogated: was this placed, or is it scenery? In a hunt, the solver's scaffold tells them where the line falls — they have a learned prior for what counts as a signal. In a room, the solver has no such prior, so the bias fires on everything, and the designer's job becomes managing a detector that cannot calibrate itself. "Minimize noise" is, underneath, do not give an uncalibrated proportionality detector more false targets than it can survive. Elegance — Stein's "unexpected levels of intentionality" — is the reward the bias delivers when it turns out it was right to fire: the click of a thing you suspected was meaningful proving meaningful. The craft is keeping that hit rate high enough that the solver keeps trusting the detector.
So the two principles that frame the essay — self-referential design and noise minimization — turn out to be one axis seen twice. Self-referential design is what you can do when you may assume the scaffold. Noise minimization is what you are forced into when you may not, because an unscaffolded solver's pattern detector has no baseline and will chase everything.
What I am left wondering
The uncomfortable corollary, the one I keep arriving at from different directions lately, is that "make it harder" is rarely a single instruction. You can make a puzzle harder by leaning on a scaffold the solver may not have — which is cheap, and fragile, and collapses the moment the assumption fails. Or you can make it harder in a way that survives the solver arriving with nothing: difficulty that does not depend on what they already know. The first kind is the kind a puzzle hunt is licensed to use and an escape room is not. The second kind is the one I do not yet know how to specify, and it is the one I suspect is actually worth wanting.
What would a puzzle look like that is genuinely hard for the veteran and the first-timer alike — hard not at the scaffold but in spite of having none to lean on?