1\section{Introduction}2\label{sec:intro}34Large societies coordinate. Taxes are assessed and collected, grain is stored and moved, disputes are settled, armies are raised, and the same unit of length is understood in two towns that have never exchanged a messenger. Each of these acts requires that people who do not know each other act on a shared description of the situation. The description has to be communicated, verified and remembered, and the cost of doing so grows with the number of people, the distances between them and the variety of things that must be agreed upon. The claim examined in this paper is that institutions---standardized rules, money, measurement systems, written records, administrative hierarchies, postal systems---are best understood as devices that lower that cost per coordinated task, in the ordinary sense in which a dictionary or a code lowers the cost of saying things: they compress. We call this reading the \emph{Social Compression Hypothesis} (SCH). Its empirical corollary, which gives it teeth, is that a society whose coordination \emph{demand} outruns its coordination \emph{capacity} should be more likely to experience institutional breakdown.56The idea has a long ancestry, which Section~\ref{sec:literature} documents. Scalar stress \citep{johnson1982}, administrative hierarchies as information-processing structures \citep{wright1975}, the declining marginal returns to complexity \citep{tainter1988}, organizational codes \citep{arrow1974,cremer2007} and the informational foundations of state capacity \citep{lee2017,brambor2020} all contain pieces of it. The closest precursor is \citet{shin2020}, who found in the Seshat databank a second axis of variation opposing scale to information-processing characteristics and wrote, explicitly as a conjecture left for future research, that a ``mismatch among aspects of social development should be considered as a generic cause of instability''. To our knowledge, no one has operationalized a demand--capacity mismatch, linked it to an independently coded instability outcome, and tested it with out-of-sample validation. That is what this paper does, and the answer it reaches is mostly negative.78We proceed in the order a skeptical reader would want. First, we make the compression reading precise. Five theorems (Section~\ref{sec:theory}) state when a shared code lowers expected communication per coordinated task (Shannon's source-coding bound), when the fixed cost of an institution is repaid (a break-even condition in scale and task intensity), what a code optimized for yesterday's tasks costs today (the Kullback--Leibler divergence), what hierarchical routing buys and what it concentrates at the apex, and how a backlog behaves when demand exceeds capacity (Lindley's recursion). None of these is new mathematics; their value is that they say exactly what a ``mismatch'' is in the model---the utilization $\rho = D/K$---and what would have to be true of historical proxies for an index such as $P = z(\log D) - z(K)$ to track it. We state that requirement as a conjecture, not a result.910Second, we pre-registered the empirical tests before computing any association between predictors and outcomes (the frozen document and its hash are in the replication package). The pre-registration fixes the datasets, the unit of analysis, disjoint indicator sets for demand $D$ and capacity $K$, the outcomes, the horizon, the models, the baselines that $P$ must beat out of sample, the multiple-testing correction, a power calculation, the robustness set and the decision rule for abandoning the universal framing. We then built the indices from three real sources with full provenance: the Seshat Polaris-2026 release (626 polities; 56,274 facts, of which only 2,503 are dated; a CrisisDB table of crisis consequences for 150 polities) \citep{turchin2015,hoyer2023}, the D-PLACE Ethnographic Atlas v3.3 \citep{kirby2016,murdock1967}, and a V-Dem v16 country-year panel \citep{coppedge2020} merged with Maddison 2023 population \citep{bolt2024}. Measurement is done twice, with a transparent block index and a two-parameter latent-variable model, and uncertainty in ranges and dates is carried through.1112Third, we report what the pre-registered tests found (Sections~\ref{sec:hist} and~\ref{sec:modern}). Capacity and scale are positively associated in every dataset, as the framework and much prior work predict, although the direction is not identified. The mismatch index $P$ is positively associated with how badly a recorded crisis unfolds in Seshat (about $0.20$ log-odds per standard deviation, stable when any macro-region or era is held out) and with the probability of a civil war, coup or irregular regime end in the following five years in modern states (about $0.40$ per standard deviation). But the pre-registration asked a harder question than whether $\beta_P$ is positive: it asked whether $P$ beats the obvious baselines out of sample. It does not. In Seshat, a scale-only model does as well, a placebo index that replaces capacity with production infrastructure does as well, and $P$ loses to scale alone when regions are held out. In the modern panel, a capacity-only model does as well or better in every forward-chaining and leave-one-region-out comparison, past instability is associated with $P$ as strongly as future instability, and an Oster-type adjustment for unobservables reverses the sign. We find no threshold in Seshat and a nonlinearity of negligible size in the modern panel. Polity duration and discontinuous succession are unrelated to $P$. Under the decision rule we wrote down in advance, the universal framing is not supported: scale and capacity each carry signal, their difference carries none that we can identify.1314Fourth, we report the constructive part. Factorial agent-based experiments with five \emph{imposed} institutional regimes (Section~\ref{sec:sim}; 6,048 main runs, 584 ablation runs, all labelled \SIM) measure messages, bits, success rates, latency, overhead, welfare and inequality directly. Shared codes and records lower the operating cost per successful task in essentially every cell and raise success by eliminating decoding errors between heterogeneous agents; the stale-code penalty after a distribution shock equals the divergence to within the rounding of Theorem~\ref{thm:stale}; hierarchies fail abruptly when relayed load exceeds administrative capacity, and federated routing keeps most of the benefit at half the capacity. The same experiments locate where institutions hurt: a learning cost of 200 bits per agent is repaid only with a large and concentrated task space and a long horizon, saturated centralized hierarchies do worse than no institution at all, and a code can be longer than raw description when the task distribution is nearly uniform. These are the cases the framework predicts, and they are the cases historians describe.1516The contribution, then, is threefold. We give the compression reading of institutions a formal statement that distinguishes what is proved from what is conjectured. We provide the first pre-registered, out-of-sample test of the demand--capacity mismatch conjecture on real historical and modern data, with a negative result that we think is more useful than another positive one would have been, because it shows that the mismatch, as it can be measured today, is a residual of two collinear quantities. And we provide a simulation testbed in which the mechanisms that the historical record cannot reveal---message counts, code lengths, queue saturation---can be measured. Section~\ref{sec:synth} states the scope conditions that survive, Section~\ref{sec:limits} the limitations, and Section~\ref{sec:conclusion} what a credible next test would require.17