1# WP35 — Literature review and novelty matrix (Phase 1)23Verification status: every reference cited here was checked against Crossref by DOI or bibliographic query on42026-11-01 (`src/wp35/validation/verify_bib.py` → `research/prior_art.csv`; manual corrections and books in5`research/bib_overrides.csv`). The abstracts and key results of the 18 closest works were read from primary or6repository sources by a dedicated verification pass (summary archived in `docs/novelty_search_report.md`). Items marked7[secondary] were known only through reviews or citing works and are cited in the paper only for their general thesis.89## 1. What is already known, by strand1011### 1.1 Scalar stress and information-processing hierarchies (archaeology/anthropology)12The oldest form of the idea is Johnson's *scalar stress* [secondary, chapter not read]: as the number of units that13must be coordinated grows, the communication and decision load grows faster than linearly, and groups respond by14fission or by adding hierarchical levels, with a modal span of about six units per level (Johnson 1982; Alberti 201415tests a related fission threshold on Hutterite colonies). Wright and Johnson (1975) defined the state as a16specialized administrative hierarchy of at least three decision-making levels and rejected population growth and17exchange as single sufficient causes of its emergence in Susiana. Carneiro (1967) documented the log-log regularity18between population and the number of organizational traits; Feinman (2011) reviews the size–complexity relation;19Flannery (1972) framed institutional evolution as changes in information-processing architecture. None of this strand20builds a demand/capacity *ratio* or tests it against instability.2122### 1.2 Declining returns and collapse23Tainter (1988) [secondary] argues that complexity is a problem-solving investment with declining marginal returns, so24that at some point collapse (simplification) becomes rational and the society loses its capacity to absorb shocks.25Middleton (2017), Butzer (2012) and Cline (2014) review the collapse literature and warn against mono-causal accounts.26Turchin's structural-demographic theory (Turchin 2003; Turchin and Nefedov 2009) models secular cycles driven by27population pressure, elite overproduction and state fiscal distress; Goldstone (1991) is the historical origin of that28theory. Sng (2014) is the closest economic precedent: in late-imperial China, the cost of transmitting information over29distance creates a principal–agent problem whose severity rises with territorial size, so that fiscal capacity decays in30the prosperous eighteenth century; this is a demand (territory) versus capacity (monitoring) argument tested within a31single polity. Hoyer et al. (2024, preprint) use the Seshat CrisisDB to show that crisis consequences are heterogeneous,32largely uncorrelated with each other, and not well predicted by territory, religion or administrative complexity taken33one at a time ("no typical societal crisis").3435### 1.3 Seshat-based quantitative history36Turchin et al. (2018) found that nine complexity characteristics (population, territory, capital population,37hierarchical levels, government, infrastructure, writing, texts, money) load on a single principal component explaining38about three quarters of variance across 414 polity-centuries. Shin et al. (2020) decomposed the remaining variance and39identified a second axis opposing *scale* characteristics to *information-processing* characteristics, defined a40"scale threshold" and an "information threshold" along PC1, and wrote — explicitly as a conjecture left for future41research — that "mismatch among aspects of social development should be considered as a generic cause of42instability". Miranda and Freeman (2020) clustered Seshat polities into two super-types with an unstable transitional43cluster in which polities dwell only a few centuries. Turchin et al. (2022) compared more than 100,000 predictor44combinations for the *growth* of complexity and found that agricultural productivity interacting with military45technology (iron, cavalry) was best supported; they did not model instability. Whitehouse et al. (2019) was retracted46after Beheim et al. (2021) showed that recoding missing moralizing-god data as "absent" drove the result; the Seshat47team's "retake" (Whitehouse et al. 2023) reports qualitatively similar conclusions with corrected coding. The48methodological lesson (never treat unknown as absent, never forward-fill, report results on extant data) is adopted in49our pre-registration.5051### 1.4 State capacity and its informational foundations (political economy)52Mann (1984) distinguished despotic from *infrastructural* power; Soifer (2008) and Hanson and Sigman (2021) discuss its53measurement, the latter with a Bayesian latent model over 21 indicators (1960–2015). Scott (1998) describes54*legibility* — standardized names, cadastres, censuses, weights and measures — as the state's way of making society55readable; Lee and Zhang (2017) operationalize legibility through age-heaping in censuses and show it predicts tax56contributions; Brambor et al. (2020) build an "information capacity" index (census, statistical agency, yearbooks,57civil and population registers) for 85 polities since 1789; D'Arcy and Nistotskaya (2017) use cadastral records.58Besley and Persson (2009, 2011), Dincecco (2017), Johnson and Koyama (2017), Mayshar et al. (2022), Borcan et al.59(2018) study the origins and consequences of fiscal and legal capacity; Stasavage (2010) shows that geographic scale60limited representative assemblies in Europe — a direct instance of distance-driven coordination cost. Gennaioli and61Voth (2015) link capacity to military competition. None of these operationalize a *ratio* of coordination demand to62capacity or test it against instability across pre-modern polities.6364### 1.5 Transaction costs, codes and communication in organizations (economics)65Coase (1937), Williamson (1985) and North (1990, 1991) treat institutions as devices that economize on transaction66costs; Hayek (1945) on knowledge dispersion and Simon (1962) on near-decomposable hierarchies are the classic67information-processing views. Arrow (1974) introduced organizational *codes*; Crémer, Garicano and Prat (2007) model68the firm's language as a code of bounded size in which precise words go to frequent events — literally lossy69compression that trades scope against precision. Radner (1993), Bolton and Dewatripont (1994), Van Zandt (1999) and70Garicano (2000) model hierarchies as information-processing networks with explicit communication costs; Dessein71(2002) and Dewatripont and Tirole (2005) study authority and modes of communication. Mechanism-design results bound the72*message-space size* needed to implement allocations (Mount and Reiter 1974; Segal 2007; Nisan and Segal 2006), and73communication complexity (Yao 1979; Kushilevitz and Nisan 1997) gives the computer-science counterpart. These results74are the formal basis for the propositions in Phase 3, but they concern firms, markets or abstract protocols, not75polities.7677### 1.6 Coordination, common knowledge and conventions78Lewis (1969) and Schelling (1960) define conventions and focal points; Young (1993), Kandori, Mailath and Rob (1993)79and Ellison (1993) show how conventions are selected and how local interaction speeds convergence; Morris (2000) on80contagion. Chwe (2000, 2001) argues that coordination requires *common knowledge* and that rituals and public81communication structures exist to manufacture it cheaply; Rubinstein (1989) shows that almost-common knowledge is not82enough. Morris and Shin (2002) and Angeletos and Pavan (2007) quantify the social value of public signals; Dewan and83Myatt (2008) model leaders as coordination devices. Ostrom (1990, 2010) and Olson (1965) are the collective-action84anchors; Boyd and Richerson (1985, 1988), Henrich (2004) and Richerson et al. (2016) supply the cultural-evolution85mechanisms for norm maintenance in large groups; Greif (1993, 2006) and Milgrom, North and Weingast (1990) are the86historical institutional cases of reputation and record-based enforcement.8788### 1.7 Scaling and information theory89Bettencourt (2007, 2013) and Ortman et al. (2014), Lobo et al. (2020) show superlinear scaling of socioeconomic90outputs with settlement population, ancient and modern. Scaling theory implies that interaction demand grows faster than91population, but it assumes infrastructure and institutions co-scale; it predicts benefits, not instability. DeDeo (2017)92is the closest explicit statement that transitions in political order are driven by information-processing innovations93and that coarse-graining (lossy compression) is the central mechanism; Marzen and DeDeo (2017) give a rate-distortion94model of perception with a high-fidelity regime (cost linear in environmental complexity) and a low-fidelity regime;95Flack (2012, 2017), Daniels, Krakauer and Flack (2017) and Krakauer et al. (2020) develop coarse-graining and collective96computation in animal societies. Shannon (1948) and Cover and Thomas (2006) are cited only for the definitions we use97in simulations where messages and distributions are actually defined.9899### 1.8 Instability forecasting and methods100Goldstone et al. (2010) forecast instability onsets 1955–2003 with regime type dominating; Fearon and Laitin (2003) and101Hegre et al. (2001) are the civil-war baselines; Beck, Katz and Tucker (1998) and Carter and Signorino (2010) on binary102time-series–cross-section data; Cox (1972) and Box-Steffensmeier and Jones (2004) for event history; Cameron and Miller103(2015) and Conley (1999) for clustered and spatial inference; Oster (2019) and Cinelli and Hazlett (2020) for104sensitivity to unobservables; Mace and Pagel (1994) and Dow (2007) for Galton's problem; Nosek et al. (2018) for105pre-registration.106107## 2. Novelty matrix108109| Work | Demand/scale measure | Capacity measure | Mismatch index | Instability outcome | Cross-polity test | Communication-cost mechanism | Simulation | Overlap with WP35 |110|---|---|---|---|---|---|---|---|---|111| Johnson 1982 | group size, units per level | hierarchy levels | implicit (span of control) | fission/disputes (ethnographic) | no | verbal (scalar stress) | no | theoretical ancestor |112| Wright & Johnson 1975 | settlement hierarchy | administrative levels | no | no | single region | no | no | definition of K |113| Tainter 1988 | — | complexity investment | no (marginal returns) | collapse (cases) | no | verbal | no | partial |114| Sng 2014 | territory, distance | monitoring/fiscal | implicit (distance) | fiscal decay | single polity | explicit | no | strong partial |115| Turchin et al. 2018 | PolPop, PolTerr, CapPop | government, writing, money | no (single PC) | no | yes (414 polity-centuries) | no | no | data template; collinearity warning |116| Shin et al. 2020 | PC2 scale side | PC2 information side | thresholds, verbal mismatch | conjectured only | yes (descriptive) | verbal | no | **closest precursor, untested** |117| Miranda & Freeman 2020 | cluster features | cluster features | no | transitional dwell time | yes | no | no | consistent evidence |118| Turchin et al. 2022 | complexity (outcome) | — | no | no | yes | no | no | growth model template |119| Hoyer et al. 2023, 2024 | territory (one covariate) | administrative complexity (one covariate) | no | **CrisisDB consequences** | yes (169 crises) | no | no | outcome source; reports weak single-variable prediction |120| Lee & Zhang 2017; Brambor et al. 2020 | — | legibility / information capacity | no | no | modern | no | no | modern K validation |121| Hanson & Sigman 2021 | — | latent state capacity | no | no | modern | no | no | latent method |122| Goldstone et al. 2010 | population threshold | regime type | no | onsets 1955–2003 | modern | no | no | forecasting benchmark |123| Crémer–Garicano–Prat 2007; Arrow 1974 | scope of organization | code size | trade-off | no | no | **explicit** | no | micro-foundation |124| Chwe 2000/2001 | network | common knowledge | no | no | no | explicit (common knowledge) | no | mechanism |125| Bettencourt 2013; Ortman 2014 | settlement population | infrastructure (co-scaling) | no | no | yes | interaction model | no | foil |126| DeDeo 2017; Marzen & DeDeo 2017 | — | coarse-graining | no | no | no | rate-distortion | no | conceptual precursor of "compression" |127| **WP35** | log population, territory, largest settlement, settlement hierarchy | administration, rules/money, communication infrastructure, records (4 disjoint blocks; mean, IRT, PC) | $P = z(\log D) - z(K)$ | CrisisDB severity; polity hazard; modern onsets | Seshat (626 polities), EA (1,291), V-Dem (28k country-years) | propositions on cost per coordinated task | factorial ABM, 5 regimes | — |128129## 3. Precise contributions claimed (and not claimed)1301. **Empirical**: the first pre-registered test of the scale/information mismatch conjecture of Shin et al. (2020) against131 an independently coded instability outcome (CrisisDB), with a transparent and a latent measurement model, uncertainty132 propagation, and out-of-sample validation; replicated in form on a modern country-year panel.1332. **Theoretical**: formal propositions stating when a shared code/record system lowers expected communication cost per134 coordinated task as a function of scale, and when fixed institutional overhead makes it worse; stated as theorems with135 proofs or as conjectures, never as empirical facts.1363. **Simulated**: a factorial agent-based comparison of five institutional regimes with directly measured messages,137 bytes, success rates and overhead, searching for regimes where institutions reduce efficiency or add fragility.1384. **Not claimed**: causal identification in the historical panel; literal Shannon compression of institutions; a139 universal theory if H2 fails (see pre-registration §7). The name "Social Compression Hypothesis" has no prior academic140 use that we could find (search 2026-11-01), but the idea has clear ancestors (§2), which the paper credits.141142## 4. Points reviewers will raise (anticipated, recorded before analysis)143- Collinearity: scale and capacity load on the same PC1 in Seshat; $P$ is by construction a residual. We report the144 correlation of $z(\log D)$ and $z(K)$ and the share of variance $P$ retains, and we test $P$ against $D$-only and145 $K$-only baselines (pre-registration §5.2).146- Crisis-selected outcome (CrisisDB covers crisis cases only) → severity-conditional-on-crisis interpretation, plus147 polity-hazard and modern-onset designs.148- Coding circularity: Seshat coders may infer administrative features from narrative sources that also describe crises;149 we use only the Polaris codes and flag inferred codes (IP/IA) in a sensitivity run.150- Missing data: Beheim et al. (2021) template; missing is never absent.151- Temporal precedence: most Polaris facts are undated → "pre-crisis" is weak; stated as a limitation, with a dated-only152 sensitivity.153