Redundancy Is Not Resilience: regional automation and correlated failure in supply chains
Cite this paper
SomaSoft. (2026-10-07). "Redundancy Is Not Resilience: regional automation and correlated failure in supply chains". SomaSoft Research. https://somasoft.ai/papers/redundancy-is-not-resilience. Licensed under SAGL-1.0.
SomaSoft is a research project, not a product: no users, no deployments. Nothing here is medical, legal or financial advice. Limits and corrections
Reading conventions. Where this paper describes ecology, epidemiology, institutional economics or classical philosophy, it reports standard results in those fields, cited as such. [I] marks our own inference or argument — not sourced, offered to be disagreed with. [measured] marks a result from our own system, with the experiment named. Several of our ecology citations are marked not re-read in our own knowledge packs, meaning the source was recorded but not re-verified at curation time; we carry that flag forward rather than quietly dropping it.
This paper is the applied case of a general argument we made in The Correlation Problem, which established that diversity protects a system only in proportion to the independence of component failures. Here we take one domain where the decision is live and the engineering concrete, and ask what that condition implies.
Two different insurances, routinely confused
Ecology distinguishes two properties that supply-chain language collapses into the one word "resilience".
Functional redundancy is the presence of multiple members able to perform the same role, so that losing one does not remove the function from the system.
Response diversity is the variation among members in how they react to the same disturbance. It is response diversity, rather than the number of members, that buffers a system against fluctuation — components that fail together provide no buffering however numerous they are (Loreau & de Mazancourt 2013).
The distinction is not a nuance. It identifies two different insurances against two different events:
| insures against | bought by | |
|---|---|---|
| functional redundancy | losing a component | second supplier, safety stock, dual sourcing |
| response diversity | components failing together | different decision rules, different exposures, different timing |
Almost all supply-chain resilience spending buys the first. Safety stock, dual sourcing, nearshoring a second plant, qualifying a backup vendor — these are redundancy purchases. They are genuinely useful against the loss of a supplier.
[I] They do very little against the failure that actually characterises an automated supply chain, which is not that one node went down but that every node moved the same way at the same time. A second supplier who reads the same demand signal, runs the same forecasting package and reorders on the same trigger is not a second response. It is the first response with extra paperwork.
The mechanism: shared decision rules, not shared geography
The clearest statement of the mechanism comes from outside logistics. Algorithmic monoculture is the condition in which the same or very similar automated system is used for a decision across an entire domain, so that an individual rejected by the system is rejected everywhere rather than by one evaluator (Kleinberg & Raghavan 2021, not re-read). The original setting is hiring and lending. The structure transfers without modification.
Correlated failure is failure of many components at once because they share a vulnerability or a decision rule. It converts what would have been independent random error into systematic error affecting the whole system (same source).
Note the "or a decision rule". A shared vulnerability is a physical fact — one port, one strait, one fab. A shared decision rule is a software fact, and it is the one that has been installed deliberately, at scale, over the last decade, under the name of best-practice standardisation.
[I] The supply-chain form runs like this. Several regional operators each adopt the same demand-forecasting and replenishment software, because it is the best available and because interoperability is valuable. A signal arrives — a weather event, a price move, a news story. Every operator's model reads the same signal and reaches the same conclusion within the same few hours. They all reorder simultaneously. Upstream sees a demand spike that no end customer created. The shortage is then real, and it was manufactured by the correlation rather than by the shock.
The amplification of demand signals upstream through a supply chain is a long-standing observation in operations research, usually discussed as the bullwhip effect. [I] Our claim is narrower and specifically about automation: shared software converts what used to be heterogeneous human misjudgement, which partially cancels, into homogeneous machine judgement, which does not. Diverse error averages out. Identical error compounds.
This is why the common proposal to fix fragility with better forecasting is, in this specific respect, pointed the wrong way. [I] A more accurate shared model makes every operator wrong at the same moment with greater confidence. Accuracy and independence are different axes, and only one of them is being optimised.
Regionality as topology, not sentiment
"Buy local" is usually argued on grounds of emissions, employment or national interest. Those arguments are real and are not ours here. The resilience argument for regionality is structural and separate.
Ecological modularity is a food web organised into weakly connected compartments, so that disturbance in one compartment spreads less to the others; compartmentalisation increases food-web persistence (Stouffer & Bascompte 2011).
[I] A regional supply chain is a compartment. The resilience contribution of regionality is not that local producers are better, or more trustworthy, or lower-carbon. It is that a region is a boundary across which a disturbance propagates less, provided the regions are not all running the same decision rule — which is the condition that makes regionality do any work at all. Twelve regional operators on one shared forecasting platform are one operator with twelve warehouses. Regional geography without regional decision-making buys nothing.
Two further ecological results sharpen what to aim for:
Response asynchrony is the condition in which components rise and fall out of step because they respond differently to shared conditions, so that aggregate output varies less than any single component does (Loreau & de Mazancourt 2008, not re-read). The portfolio effect is the resulting reduction in variability of the aggregate — the statistical mechanism shared by a diversified financial portfolio and a species-rich community (Thibaut & Connolly 2013, not re-read).
So the target is not "more suppliers" and not even "more regions". It is asynchrony: operators whose reorder timing is genuinely out of step.
A criterion you can measure before the shock
The practical problem with resilience is that it is usually only measurable in retrospect, which makes it unfalsifiable in advance and therefore easy to claim. The response-diversity framing gives a quantity that can be measured on ordinary operating data, with no disruption required.
[I] We propose the following, stated plainly enough to be argued with:
[I] Do not count suppliers. Measure the correlation between their responses.
[I] For each pair of nodes in the network, take the time series of their ordering decisions — quantity, timing, or both. Compute the pairwise correlation of those series over a period containing at least one shared external signal. Report the mean pairwise correlation, and report the worst-case cluster: the largest set of nodes whose mutual correlation exceeds some threshold.
[I] A network of forty suppliers whose ordering decisions are almost perfectly correlated has, for resilience purposes, roughly the diversity of a network of one.
This has three properties we consider necessary for the measure to be worth anything. It is computable from data an operator already holds. It is falsifiable before a disruption rather than after. And it will frequently return an uncomfortable answer for networks that score well on supplier-count metrics — which is the point, since a measure that always agrees with the existing one is not adding information.
[I] We would also expect the measure to be resisted, because the correlation it reveals is usually the direct result of a standardisation programme that someone was promoted for.
The governance form that matches the topology
A modular network needs a governance structure that is also modular, or the decisions re-centralise and the topology becomes decorative.
The relevant body of work is Ostrom's, on the conditions associated with durable self-governed commons: clear boundaries, locally fitted rules, participation by those affected, accountable monitoring, graduated sanctions, cheap conflict resolution, recognised rights to organise, and nested layers of governance (Ostrom 1990).
[I] Two of those eight do specific work here. Locally fitted rules is the direct antidote to the shared decision rule — it is the institutional statement of response diversity. Nested layers is what allows regions to coordinate on standards (physical interfaces, data formats, safety requirements) without coordinating on decisions (when to reorder, what to forecast). Those two kinds of standardisation are routinely conflated and have opposite effects on correlated failure: interface standardisation raises substitutability, which helps, while decision standardisation raises correlation, which hurts.
The older institutional principle is the same shape. Separation of powers divides the authority to propose, decide and carry out action among independent bodies so that no single one can act unchecked (Montesquieu 1748; Madison, Federalist 51). [I] An automated supply chain in which one platform forecasts, decides and executes has collapsed all three functions into one body, and the argument against that arrangement is two and a half centuries old.
The part that is usually left out: the people in it
Supply-chain automation is normally justified by removing human decisions. There is a measured cost to doing that which is not usually counted in the business case.
Job control is the decision latitude a worker has over how and when tasks are done (Karasek 1979). Job strain is the combination of high psychological demands with low job control, and is associated with mental strain and cardiovascular risk — the Whitehall II cohort found job control contributing to social variation in coronary heart disease incidence (Karasek 1979; Marmot et al. 1997). Meaningful work is work experienced as significant because it offers skill variety, task identity, task significance, autonomy and feedback (Hackman & Oldham 1976).
[I] This matters to the resilience argument and not only to the ethics of it. The local dispatcher who overrides the model because she knows the bridge is out is response diversity. She is a decision rule that differs from the platform's. Automation designed to eliminate her removes, in one move, both the latitude that protects her health and the heterogeneity that protects the network. A regional automation design that keeps a human override at each node is buying two things with one decision, and the second one does not appear in the business case at all.
We state the direction of this claim carefully. The epidemiology on job control is strong. The link from retained human override to measured network resilience is our inference and we have not tested it.
A caution from our own system
[measured] We have run the component-diversification experiment on ourselves and it did not work. In our own question-answering system we varied both the language model and the knowledge graph — two of the three obvious components — and failed to decorrelate failures, because the correlation lived in a routing layer that neither variation touched. The result is reported in The Correlation Problem with the experiment named.
[I] We think this generalises as a warning rather than a result. Diversifying the components you can see does not establish independence if a shared layer sits underneath them. For a supply chain, candidate shared layers include: a common data provider, a common weather or demand feed, a common cloud region, a common certification regime, and common personnel trained on the same methodology. A network can be genuinely diverse at the supplier layer and perfectly correlated at the forecast-input layer, and the supplier-count metric will show health throughout.
We offer this as a specific thing to check rather than as a proof. We have one instance of it, in a different substrate, in a system much smaller than a supply chain.
Timing and the stopping rule
Two older ideas bear directly on automated optimisation and we think they earn their place here.
Wu wei is effective action without forcing: acting early and lightly rather than late and hard, and declining the exertion that creates the resistance it then has to overcome. The Tao Te Ching, chapter 63, has it as acting "without (thinking of) acting" and anticipating "things that are difficult while they are easy". It is not passivity; it is a claim about timing and proportion. [I] The automated analogue is a control system with a small, early, frequent correction rather than a large, late, confident one — which is also, not coincidentally, the intervention profile least likely to synchronise with everybody else's.
Sufficiency is knowing what is enough, treated as a form of wealth rather than a constraint on it; in the Tao Te Ching, the person who knows sufficiency is counted rich, and the pursuit of more is the mechanism by which enough is lost. It is the counter-principle to optimisation without a stopping rule.
[I] We take that last phrase literally as an engineering requirement. An automated supply chain optimising for cost or inventory turns with no stopping rule will drive buffers toward zero, because every individual buffer reduction is locally justified. The stopping rule has to be stated in advance, as a floor, in the same terms the optimiser understands — not as an exception a human invokes under pressure, since the whole difficulty is that the pressure arrives exactly when the floor is most needed.
What would falsify this
We would consider the central argument wrong if:
- A network with high measured pairwise response correlation absorbed a shared shock as well as a matched low-correlation network. That is the direct test and we have not run it.
- Shared forecasting software were shown to reduce ordering synchrony in practice — plausible if the software deliberately staggers or randomises its recommendations, which would be a design response to this very argument and would count as the argument succeeding rather than failing.
- Response correlation turned out not to be measurable from ordinary operating data at useful resolution. This is the most likely practical failure of the criterion proposed above.
Limitations
This is a design argument, not an empirical result. We have not deployed, measured, or observed a regional automated supply chain. We hold no proprietary logistics data and have ingested none. The ecological and institutional findings are standard results in their fields, cited, and several are flagged not re-read in our own knowledge packs. Everything marked [I] is our inference, including the central claim that shared decision rules rather than lost capacity are the dominant failure mode — which is an argument from structural analogy and is exactly the kind of transfer that should be checked in the destination domain before being believed.
The one quantity we actually measured is the negative one reported above: that varying two visible components of our own system did not decorrelate its failures.
How the cross-domain links in this paper were found. The connections between ecological response diversity, algorithmic monoculture, commons governance and the classical material are not literature review. They are relations declared in SomaSoft's curated knowledge packs and traversed in our concept graph — for instance an edge asserting that wu wei prevents correlated failure, which sits in the graph alongside one asserting that algorithmic monoculture causes it. A 2,500-year-old text and a 2021 computer science result meet at the same node. We mention the method because it is the actual provenance of the argument's shape, and because the honest counterpart is that such a path is a prompt for thought rather than evidence of anything: the edge asserting it was written by us.