The Brittle
Network

The collapse of the Late Bronze Age is the best-known systemic failure in the ancient record and one of the worst-served by its own evidence. This document specifies how a multidisciplinary team would test the cascading-failure hypothesis rigorously enough to be wrong.

It is a protocol, not a study. It contains no results, and it is written so that its conclusions could fail: §5.1 lists seven conditions under which the central hypothesis is rejected, and the abstract to be published if it fails is drafted in advance.

Colour is used as a register throughout this document. It encodes which evidence stream a claim belongs to, and nothing else.

“Prove or disprove” is not an available inferential mode

The collapse is a single, unrepeatable realization with n = 1 at the system level. No dataset can prove a causal architecture for it. The protocol replaces proof with severity of test: each hypothesis must generate predictions that would very probably have failed had the hypothesis been false, and the analysis reports posterior model probabilities and Bayes factors rather than verdicts. A hypothesis that cannot be made to generate a differential prediction is declared empirically inert for this study and excluded rather than argued about.

Given the observed spatiotemporal pattern of destruction, abandonment, contraction, and survival between 1250 and 1100 BCE, which class of generative model (exogenous-shock-only, network-topology-only, or shock-propagating-through-brittle-topology) assigns the highest likelihood to the observed pattern, and how large is the margin relative to the uncertainty in the evidence?

“1177 BC” is a heuristic anchor, not a datum

1177 BCE corresponds to Year 8 of Ramesses III in the Medinet Habu inscriptions, the best-known Egyptian account of conflict with the coalition conventionally labelled the Sea Peoples. Cline deploys the date as a narrative anchor for a multi-decadal process, not as the year of a discrete event. The protocol treats the collapse as a transition interval, provisionally 1225–1130 BCE, whose boundaries are parameters estimated from a Bayesian chronological model rather than assumed.

The genuine chronological hazards at 1250–1100 BCE are different in kind, and §2.3 is built around them:

1. Calibration-curve structure. IntCal20 is comparatively well-behaved here, but twelfth–eleventh century wiggles produce multi-modal calibrated distributions routinely spanning 60–120 years at 95.4% for single determinationscoarser than the entire causal sequence under investigation.
2. Regional and growing-season offsets. Demonstrated offsets between Northern Hemisphere calibration data and Eastern Mediterranean samples run to a few decades: negligible for most prehistory, decisive for ordering events inside a century.
3. Dating-target ambiguity. A radiocarbon date dates a plant's death, not a destruction. Old wood, curated heirlooms, and residual charcoal all displace the date from the event.
4. The Aegean–Levantine interlock. Mycenaean LH IIIB/IIIC, Cypriot LC IIC/IIIA, Levantine LB IIB/Iron IA, and Egyptian regnal chronology are cross-linked by ceramic synchronisms. An error in one link propagates through the grid, so synchronisms are modelled as uncertain priors, never fixed pegs.

The epistemic contract

• Non-fabrication. No site, text, tablet, core, ingot, or study is cited unless verifiable in the published record. Where an evidence class is required but no specific dataset is confirmed, the document specifies the evidence type, acceptance criteria, and sampling requirement instead, marked [DATA REQUIREMENT].
• Separation of layers. Evidence, model, and interpretation stay in separate registers. A modelled quantity is never reported alongside an observed one without an explicit marker.
• Pre-registration. The analysis plan, falsification criteria, and stopping rules are fixed before data compilation. Deviations go in an amendment register, not silently into the method.
• Adversarial review. Each phase is reviewed by a named scholar on record as sceptical of the hypothesis.

The historiographical arc

The literature falls into five broad, partially overlapping generations. The protocol characterizes each by its causal topology, because that is what determines whether the position can be tested by network methods at all.

Generation III states that stressors interacted; it does not specify the medium of interaction, nor predict which polities fail. Naming that medium as the exchange network, and making the claim quantitative, is where this study begins.

The translation: from historiographical claim to mathematical object

This is the pivotal deliverable of Phase 1. Each narrative claim maps to a formal operator so Phase 3 can implement it, and, crucially, so that rival claims produce different observable signatures.

Foundational positions on the environmental and economic drivers

Five verified anchors, chosen because each supplies a methodologically independent proxy stream rather than a restatement of the same evidence.

A published response contesting Powell et al. 2022 exists and is cited alongside it. Tin provenance is treated as contested, and the contestation is propagated as a prior over source configurations rather than settled by fiat.

Competing hypotheses and their discriminating predictions

The study is a model comparison, so the rivals must be stated in advance with predictions that differ. A design in which every hypothesis predicts the same observations is not a test.

Phase 1 deliverables

• D1.1: Annotated historiographical review (~12,000 words), organized by causal topology rather than chronology.
• D1.2: The claim→operator table in machine-readable form, each row bound to the Phase 3 module implementing it.
• D1.3: Pre-registered hypothesis set with prior model probabilities elicited by structured expert elicitation (Cooke's classical method, with calibration questions weighting each expert).
• D1.4: A register of inert claims: assertions in the literature that cannot generate a differential prediction, with reasons. Published, not suppressed.

The historical universe

The universe is defined by participation in the exchange system, not by modern geography. A site enters if it satisfies at least one of: documented possession of imports traceable to another region within the universe; attestation in an inter-polity textual corpus; or production or transhipment of a commodity attested in circulation elsewhere.

Core spheres are the Mycenaean Aegean, Crete, the Hittite empire and its Anatolian and North Syrian dependencies, Cyprus (Alashiya), the Levantine coast and interior, and Egypt. Three extensions are mandatory, without them the tin question cannot be posed at all:

• Mesopotamia and the Middle Euphrates, as the overland conduit for eastern tin.
• Central Asian and Anatolian tin sources implicated by tin-isotope provenancing.
• Central Mediterranean nodes (Sardinia, Sicily, southern Italy) where oxhide ingots and Aegean-type material occur.

Evidence tiers

Every datum carries a tier label, and tier is a covariate in every downstream model.

Data streams and acceptance criteria

◆ Stream P: Palynological and terrestrial hydroclimate

A core is admitted only if it has a published age–depth model built from ≥ 5 dated horizons within or bracketing 1600–900 BCE with a stated method and uncertainty envelope; sampling resolution ≤ 50 years per sample across the transition; published raw counts by taxon, not only an interpretive curve; and an explicit strategy for discriminating aridity from clearance or abandoned orchard husbandry.

◆ Stream S: Speleothem, marine, and independent physical proxies

Included to break the circularity in Stream P, since pollen responds to human land use as well as to climate. Requirements: U–Th chronology with published errors, δ¹⁸O and δ¹³C series, and a stated hydrological interpretation for the specific cave system. Agreement between streams is a finding; disagreement is data, not noise to be smoothed.

◆ Stream E: Archaeoseismological

The weakest-evidenced and most over-interpreted stream in the collapse literature, and treated with corresponding severity. Every candidate event is scored on a formal quality index; admission requires ≥ 6 of 10, with a non-zero score on damage typology and on alternative-cause exclusion.

Events below threshold enter a shadow register used only in sensitivity analysis. The earthquake-storm hypothesis is then tested as a point-process hypothesis, a conditional-intensity test against a homogeneous Poisson null at the region's long-term seismicity rate, not as a narrative assessment.

◆ Stream M: Archaeometallurgical and provenance

The purpose is to reconstruct flow, not merely presence. Lead-isotope analysis is adopted with its standard cautions promoted to first-class model components rather than footnotes:

• ore-field isotopic fields overlap, so provenance is a posterior distribution over sources, never a point assignment;
• recycling and mixing shift composition along mixing lines, so mixing is modelled explicitly as a two- or three-endmember problem with unknown proportions;
• reference-database coverage is incomplete, so an “unmatched” result is evidence of a gap in the reference set as much as of an exotic source;
• trace-element and, where available, copper-isotope data are required as an independent check before any provenance claim is admitted.

◆ Stream T: Textual

Three corpora supply the political-dependency layer and the only direct evidence of contemporaneous perception: the Amarna correspondence (EA 1–382, Moran 1992), the Hittite state archives from Boğazköy/Ḫattuša, and the Ugaritic and Akkadian correspondence from Ras Shamra.

Chronological normalization

The hardest problem in the design. The approach is to abandon absolute placement: do not attempt to put events on a calendar; estimate the joint posterior over event orderings and intervals.

All determinations are modelled in OxCal against IntCal20 with Sequence/Phase/Boundary structures encoding stratigraphy; general and charcoal outlier models on every determination; a jointly estimated regional offset parameter Δ_R with a prior informed by the demonstrated Mediterranean offsets; ceramic synchronisms as informative priors with explicit uncertainty; and KDE_Plot so that apparent clustering can be distinguished from the sum of individually wide calibrated distributions.

The ordering test: the actual inferential workhorse

Causal claims here are claims about order and lag. For every pair of events the protocol computes, directly from the joint chronology, the posterior probability P(t_A < t_B | D) and the posterior distribution of the lag Δ_AB.

Coping with irreducible imprecision

1. Coarsening. Aggregate to 25-year bins and accept the loss of power rather than manufacture precision.
2. Wiggle-matching. Where sequences of dated samples exist, wiggle-matching can reach decadal resolution, so sampling programmes that would yield such sequences are prioritized (§5.4).
3. Order-free tests. Some predictions need no ordering at all. That failure severity correlates with centrality after controlling for exposure is a cross-sectional test, and survives total chronological failure. The design deliberately front-loads such tests so the study still discriminates H₁, H₃ and H₄ in the worst chronological case.

The destruction-horizon re-audit

Before any site-level failure enters the model, its destruction claim is re-derived from primary excavation documentation: is there a published section or plan; is burning architectural or localized to a kiln or hearth; is the assemblage consistent with sudden abandonment or planned departure; are trauma indicators published or merely asserted; is the deposit dated independently or by assumed correlation with the collapse horizon; and is there immediate reoccupation?

Outcome is a categorical response variable with measurement error, not a binary. Circular dating, a layer dated to 1200 BCE because it is a destruction layer, is flagged, and the site is excluded from chronological tests while remaining in cross-sectional ones.

Survivorship, selection, and the causal firewall

Four distinct biases, each with a distinct correction. Conflating them is the standard error in this literature.

Data infrastructure

A PostgreSQL + PostGIS schema in which no uncertain quantity is stored as a scalar: dates are posterior distributions, provenance is a distribution over sources, edge weights are distributions. Every row references a source and carries a tier. Git-tracked ETL, every figure regenerable by one command, releases DOI-minted, CIDOC CRM alignment for interoperability, coordinates deliberately fuzzed for sites at looting risk, and code and derived data released under open licence at publication.

Formal specification

The system is a temporal, weighted, directed multiplex:

Separating commodities into layers is not cosmetic. Copper and tin have different source geographies, different substitutability, and different route structures; collapsing them into one “trade” graph destroys precisely the asymmetry the study is testing.

Nodes, edges, weights

Node attributes are carried as distributions, never scalars: location, settlement extent, administrative complexity c_i (archives, sealing practice, standardized weights, storage architecture, craft specialization), agricultural capacity π_i, storage buffer σ_i converted to person-months of supply, metallurgical dependency μ_i, and political dependency ρ_i.

An edge requires positive evidence of directional flow, from three independent classes combined by evidence synthesis rather than unionmaterial provenance (discounted by the provenance posterior and the number of plausible intermediaries), textual attestation (full strength on the political layer, weak on commodity layers, because gift-exchange is a poor proxy for bulk volume), and route feasibility (least-cost paths over sea accounting for winds, currents, seasonality and the coastal-tramping character of Bronze Age sailing, and over land across terrain cost surfaces).

d is what matters for cascade dynamics. A node may carry small absolute volume and still be structurally critical if it is a sole supplier.

Metrics and cascade dynamics

Definitions are stated explicitly so results are reproducible.

Interdependent cascade. Where a node's function in one layer depends on its state in another (a palace cannot administer grain redistribution once its metallurgical and prestige economy has failed, and cannot sustain that economy without agricultural surplus) the iterative failure process between coupled layers produces a first-order, discontinuous percolation transition, in contrast to the continuous transition of a single network. Interdependence converts graceful degradation into abrupt collapse.

This is the sharpest signature available. Under H₁ decline should be roughly proportional to shock magnitude. Under H₄ with interdependence there is a threshold below which the system absorbs the shock and above which it disintegrates.

The tin bottleneck

The metals system is modelled explicitly because it is where the argument is most testable. Tin bronze at the standard ~10% Sn requires inputs in near-fixed proportion, a Leontief technology with negligible short-run substitution:

Copper was available within the Eastern Mediterranean, Cypriot deposits above all. Tin had no abundant, securely exploited regional source at the scale of demand, small Anatolian deposits notwithstanding, and travelled great distances from Central Asian, Anatolian, and possibly European sources. The binding term is overwhelmingly sn_i / θ.

Substitution. The ferrous transition is a substitution response, not merely a technological succession. Substitution capacity r_j becomes time-varying, and the aftermath window tests whether the network reconfigured around a commodity with ubiquitous sources, structurally, a shift from a long-chain, high-criticality input to a short-chain, low-criticality one.

Perturbation operators and the experimental matrix

Ensemble ≥ 10⁴ realizations per cell, or until Monte Carlo standard error falls below 1% of the outcome range. Latin hypercube sampling over continuous parameters; full seed and configuration logging; every figure regenerable.

Locating the tipping point

The central question is answered as a critical-phenomena measurement, not a narrative judgement. Rather than the bare largest-component fraction, the order parameter is a functional outcome, whether the system is doing its job:

• Order of the transition is discriminated by a discontinuity in Ψ under infinite-size extrapolation, bimodality of Ψ across ensemble members near f_c, and hysteresis under a reverse sweep restoring capacity.
• Finite-size analysis is mandatory. With N ≈ 120–250, apparent sharpness may be an artefact. Any claim of a sharp transition unaccompanied by finite-size analysis is inadmissible.
• Early-warning signals (rising autocorrelation, variance and skewness before a bifurcation) are directly measurable in the model. The empirical counterpart (settlement-size variance, import-diversity variance, hoarding frequency, assemblage heterogeneity) is deliberately weaker and is presented as exploratory: archaeological resolution is close to the limit this test needs, and a negative result would be uninformative, which must be said when it is reported.
• Attribution decomposition. A Shapley-value decomposition over the factorial design splits the systemic outcome into shock magnitude, topology, and their interaction. H₄ predicts the interaction term carries the largest share, which converts “the shocks interacted with a brittle network” from a metaphor into a number with a confidence interval.

Validation: a model that fits the collapse is worthless unless it could have failed to

1. Out-of-sample temporal validation. Fit on 1400–1250 BCE baseline dynamics only; predict 1250–1100. No collapse-window data touches the fitting stage.
2. Spatial hold-out. Withhold entire regions (Cyprus, or the southern Levant) fit on the remainder, predict the withheld pattern of failure and survival.
3. Independent-event benchmark. Run the identical pipeline on the Theran eruption horizon, where the Knappett–Rivers–Evans results provide a published comparator. If it cannot reproduce known results there, it is not trusted here.
4. Survival prediction scored by balanced accuracy and Matthews correlation coefficient, because the classes are imbalanced and raw accuracy would mislead.
5. Adversarial fitting. An independent team fits the same observations with H₁ and H₂ maximally favoured, on an equal computational budget. Bayes factors are computed between the best H₄ model and the best adversarial model.
6. Sensitivity. Global Sobol indices over all parameters; results reported for the full posterior, never a maximum-likelihood point.
7. Prior sensitivity. Every substantive conclusion re-run under the advisory panel's most sceptical elicited prior. Conclusions that do not survive are reported as prior-dependent.

Operationalizing brittleness

Brittleness is not a mood. It is a measurable property of a response function, defined here as a five-component vector so that no single number can hide a trade-off.

The composite index B = 1 − 2R is the headline scalar; the vector is what gets interpreted, because two systems can share a B for entirely different reasons and will fail differently.

The mechanism, stated so it can be checked

1. Centralized redistribution suppresses redundancy. When a palace stops functioning, the redistribution it performed does not degrade, it stops. Testable: the sharpest discontinuities occur at sites with the highest administrative-complexity index, and dependent settlements fail at a short, measurable lag behind their palace.
2. Elite prestige demand rewards long chains. Value accrues to the exotic, so legitimacy expenditure commits the system to long, thin supply chains, the configuration with the lowest per-hop survival probability.
3. Specialization raises efficiency and lowers substitutability.
4. Buffers are politically expensive. Stored surplus that is neither consumed nor displayed returns no legitimacy, so under competitive display pressure α is driven down. Testable, and genuinely surprising if true: storage capacity relative to estimated catchment yield should decline through the thirteenth century at high-complexity sites.
5. Coupling is a by-product of integration. The same institutions manage metals, grain and diplomacy, and strong coupling makes the transition discontinuous.

Point 4 is the sharpest empirical hook in Phase 4 and should be prioritized: it is a claim about excavated storage volumes over time, it is measurable, and it could easily come out the other way.

The homology test for modern parallels

No parallel is asserted unless this gate is passed, and the modern comparator must be named and characterized in advance, not left as “globalization.”

Scoring: a parallel is asserted only where at least four of six pass and none fails catastrophically. Where fewer pass, the finding is reported as structural difference, equally informative, and considerably rarer in the literature.

The disanalogy register: published, not buried

What actually transfers, if the gate is passed

The transferable content is structural and conditional, never predictive.

1. Efficiency and robustness trade off, and the trade-off is invisible in normal operation. A network optimized for throughput emits no distress signal while conditions stay in-distribution. The absence of failure is not evidence of robustness.
2. Robust-yet-fragile is a topological property, not a management failure. Systems that have survived many random shocks may have learned precisely the wrong lesson about their own resilience.
3. Interdependence converts gradual degradation into discontinuous collapse. Coupling logistics, finance, energy and information moves the transition from second-order to first-order, the warning interval between “stressed” and “failed” shrinks toward zero.
4. The binding constraint is the long-chain, low-substitutability input, whatever its share of value. Criticality is a function of substitution elasticity and chain length, not cost share, exactly the quantity conventional procurement accounting does not measure.
5. Correlated shocks defeat diversification. When drought, seismicity and unrest are spatially correlated, apparently independent routes fail together and measured diversification overstates real protection.
6. The visible agent of collapse may be an output of the system's own stress. If the endogenous-mobility result holds, attributing collapse to whoever appears at the moment of failure is a category error. This is the study's contribution to general theory and the claim most at risk of being over-stated; it must travel with the model's limitations attached.

Conclusion architecture

The concluding argument is written to a fixed structure that makes the strength of each claim visible: (1) what the evidence shows, with tiers and uncertainties, independent of any model; (2) what the model shows, always paired with the null comparison; (3) posterior model probabilities for H₀–H₄ with prior sensitivity; (4) what remains unresolved and the specific evidence that would resolve it; (5) what generalizes, gated by the homology test; (6) what does not.

If the hypothesis is supported, the concluding claim is bounded in this form and no stronger:

The observed pattern of failure and survival between 1250 and 1100 BCE is better explained by exogenous climatic and seismic shocks propagating through a highly efficient, low-redundancy, strongly coupled exchange network than by those shocks alone, by network structure alone, or by exogenous migration. The mobile groups recorded in Egyptian sources are, on this account, consistent with an output of the failing system rather than its initiating cause, though the textual record is too thin, and its survival too strongly conditioned on the collapse itself, to exclude an independent migratory contribution.

If the hypothesis is not supported, the corresponding statement is written in advance, so the framing cannot drift after the results are seen.

Falsification criteria

Programme, team, and risk

Team. A network scientist with percolation and multilayer expertise; a computational archaeologist; four regional specialists (Aegean, Anatolian, Levantine/Cypriot, Egyptian); a philologist covering Akkadian, Ugaritic and Hittite; a palaeoclimatologist; an archaeometallurgist; a Bayesian chronologist; a data engineer; and a heritage-liaison officer.

Data acquisition priorities

Ranked by expected reduction in posterior uncertainty:

1. High-resolution, well-dated hydroclimate records for the Anatolian plateau and northern Levant, the largest spatial gaps in the drought field.
2. Dendrochronologically anchored or wiggle-matched sequences from destruction contexts, the only realistic route to decadal ordering.
3. Systematic lead- and tin-isotope programmes on stratified assemblages spanning 1300–1050 BCE, converting provenance snapshots into a time series.
4. Re-excavation or archival re-study of destruction contexts flagged as insufficiently documented.
5. Systematic quantification of storage capacity by phase at palatial sites, the weakest link in the α estimate, on which the sharpest Phase 4 prediction depends.

Ethics

Coordinates for sites at looting risk are fuzzed in public releases. Sampling requires national permits and is minimized in favour of legacy-collection reanalysis. Regional scholars and institutions in the countries where the evidence originates are collaborators with authorship, not data providers. Published language avoids framing collapse as civilizational judgement, and avoids deploying ancient migration as commentary on modern migration, a use of this material that is both historically indefensible and politically instrumentalized.

Deliverables