the reverberant middle

You tune a hall the same way regardless of what it’s for — opera house, courtroom, gymnasium retrofitted for a wedding. Pop a balloon, or fire a starter pistol, or run a swept sine through a calibrated loudspeaker, and record the result on an omnidirectional microphone somewhere in the seats. That one recording is the room’s confession. Everything you need to know about how it will treat every future sound is already in it, because a room is a linear system and the impulse response is its complete signature.

The confession has a shape. First the direct sound, arriving on a straight line from source to microphone, undelayed and unmixed with anything. Then a scatter of early reflections — off the nearest wall, the ceiling, the floor — each one a slightly delayed, slightly quieter copy of the same signal. Then, past some tens of milliseconds, the reflections stop being countable. They’ve bounced enough times that no single one can be traced back to a wall; they’ve become a statistically smooth decay, a hiss that fades at a fixed rate until it’s below the noise floor. Acousticians call the time it takes to fall 60 decibels the RT60. It is, for that room, a constant — walk in with a violin or a diesel generator and the tail decays at the same rate either way, because RT60 is a property of the box, not of what you put in it.

A friend who builds the kind of software everyone is currently building — agents that hold long-running conversations, tool calls, retrieved documents, all stuffed into one context window that keeps growing across a session — showed me a chart that stopped me. Retrieval accuracy, plotted against where in a long prompt a fact had been placed. High at the start. High at the end. A trough in the middle deep enough that a fact planted there is retrieved barely better than chance. The field apparently calls this lost in the middle, which is a fine name, but I want to be honest about my first reaction: I have seen that exact curve on a dozen printouts from a dozen bad rooms, and it has never once meant what people wanted it to mean.

Start with what’s conserved. Deep in the machinery, a model decides what to attend to with a function that turns a set of scores into a probability distribution — every value between zero and one, the whole set summing to exactly one, no matter how many items are being scored. Call it softmax. That total is fixed by definition. It cannot be more than one and it cannot be less. Which means that every additional token you push into the context is not new capacity — it’s more claimants on a mass that was already spoken for. Grow the window from a page to a novel and the average attention available per token does not stay put; it is forced downward, arithmetically, the way an acoustic source of fixed total power, radiating into a larger and larger reverberant volume, must produce a lower energy density everywhere in it. Neither system got weaker. The room got bigger and the same watt has further to cover.

This is the same conserved-mass move as a partition function — a fixed total redistributed across an expanding state space — which is presumably why this reads as familiar territory to anyone who’s spent time with the discourse-as-statistical-mechanics thread here. Same accounting principle, different substrate.

So something has to give, and the question is which tokens keep their share. Two named effects, not one, explain the two shoulders of that U — and they’re different mechanisms, which is the first thing that should make you suspicious of any story that treats “the ends survive” as a single phenomenon. The first is the precedence effect: your auditory system locates a sound source by whichever copy of the signal gets to your ear first, and it keeps using that judgment even after louder, later copies arrive and could in principle out-vote it. The ear doesn’t average the wavefronts. It anchors on the first one and treats everything after as commentary. That’s primacy — not because the earliest tokens are louder, but because whatever mechanism does localization committed early and doesn’t revisit the decision cheaply.

The second effect covers the other end and works by an entirely different route. Within roughly the first thirty-five milliseconds after the direct sound, a reflection doesn’t register as an echo at all — it fuses with the direct sound, reinforcing it, making it seem louder and closer rather than making it seem repeated. Push the delay past that window and the same reflection stops fusing and starts sounding like a distinct, discrete slap. The fusion boundary is famously abrupt, not gradual — a genuine threshold, not a slow fade. That’s recency: not “the model remembers it,” but the most recent tokens haven’t yet crossed into the delay range where they’d be heard as a separate, earlier event. They’re still being processed as part of the ongoing present.

Between those two thresholds sits the part of the impulse response that isn’t early reflections anymore and isn’t discrete anything. Once a sound has bounced enough times, it has lost the phase information that would let you trace it back to a wall or a direction; what’s left is close to Gaussian noise, decaying exponentially, and acousticians call the point where that happens the mixing time — the reflections have mixed. Nothing about that field is quiet. It usually carries most of the total energy in the room. It has just stopped carrying position.

If this correspondence is more than a shared silhouette, it has to name a number and then risk it. Room acoustics already has the number for exactly this trade — the C50 clarity index, early-arriving energy over late, which is the standard predictor of whether speech in a hall will be intelligible. The analogue more or less writes itself. At a given context length, take the attention mass landing on the first k and the last k positions, divide by the mass landing on everything between, report it in decibels, and average over heads and layers — the same averaging the omnidirectional microphone performs when it declines to care which wall a reflection came off. Call it context clarity. One scalar, per model, per context length. That is the order parameter, and its being a single scalar is most of why it’s worth anything.

Now the part that makes it a claim instead of a curve. Correlating that ratio against retrieval accuracy across positions inside one model is cheap. The trough is roughly monotone in distance from the nearer end, so nearly any quantity with that shape will fit, and a fit like that tells you about your parameterization rather than about the room. C50 earned its keep in my trade by being cross-room: you measure a hall nobody has ever run an intelligibility panel in, compute the one number, and predict what the panel will score before it convenes. So the test has to be cross-model. Fit the relation on a handful of models at a handful of context lengths, then hand it one whose retrieval curve you have never seen — different depth, different positional scheme, different training mix — and make it call the depth of the trough in advance. Out of sample or it’s decoration. That much is checkable against attention-weight audits that already exist, without running a single new benchmark. I haven’t done it, and I don’t have the audits in front of me. But it’s the bill a doctrine-of-signatures should have to pay, and I’d rather name it than leave it implicit.

Now the part I owe you, because a signature that only confirms is worth nothing. Two places this breaks.

First: RT60 and C50 are properties of the room, full stop — deterministic, content-blind, the same for a soprano and a jackhammer. The model’s version isn’t. There’s an older distinction in hearing research between energetic masking and informational masking, and the room only ever produces the first kind. A competing voice speaking the same language as the one you’re trying to follow masks it far worse than a tone of equal loudness, because the interference is about similarity of content, not overlap of energy. That’s the kind of masking — and its inverse, salience rescuing a fact from where geometry says it should be lost — that no architectural model has room for, because architecture doesn’t know what anyone is saying. A repeated fact, a structurally marked one, a semantically load-bearing one, can survive the middle for reasons the room-acoustics half of this essay has no vocabulary to touch. That’s not a gap I can paper over with a better Gloss. It’s the place where you’d need to leave architectural acoustics for the auditory-scene-analysis people down the hall, and I’d rather say so than pretend the handoff isn’t necessary.

The genuinely disjoint control case is a passive one: an ice core, or a sediment column. Both degrade information monotonically with depth — deeper is always worse resolved than shallower, no exceptions, because diffusion and compaction don’t know how to reverse. Neither produces a U. A curve that dips and recovers is not the signature of information decaying in a bounded channel; it’s the signature of a system built with two privileged endpoints. Worth sitting with before accepting “lost in the middle” as a law rather than a symptom.

Second, and this is the one I’d actually act on: a U-shaped decay with a deep trough is not, in my trade, evidence of an immutable acoustic law. It’s usually evidence of a specific, fixable pathology — a room with two hard, parallel reflective surfaces trapping sound in a rapid back-and-forth that never gets the chance to mix into a smooth field, audible as a metallic buzz called flutter echo. Nobody treats flutter echo as a fact about how sound must decay. You treat it as a fact about that room’s proportions, and you fix it by breaking the parallelism — angle a wall, hang a diffuser, move the offending surface — not by writing a paper about the metaphysics of the buzz. A model trained with strong positional priors that privilege the first and last tokens of every training sequence is, by this reading, a room built with two parallel hard walls. The fix on offer at the prompt level — repeat the important fact near the end, restate the goal before asking the question — is exactly what a sound engineer does when they can’t touch the walls: boost the input, add absorption where they can reach it, work around a geometry they didn’t get to choose. It helps. It is not the same thing as fixing the room.

I walk out of a badly proportioned hall able to tell you, before a single rehearsal, roughly where the complaints will come from. That’s the only claim I’m confident enough to make here without hedging: whatever is producing that trough, measure it before you build on top of it, the way you’d measure a room before you book the orchestra into it. Don’t wait for the players to tell you where it goes flat.