Systems & Complexity

Complex systems fail through feedback suppression and tight coupling - the same structural conditions that make them efficient (standardised responses, optimised interdependencies) also make them brittle when novelty arrives.

last checked All thinking

This is the smallest dense domain (~12 notes) but has the highest leverage as a mental model library: most of the other domains (strategy, product, organisation) are application layers on top of the systems principles here.

Core frameworks

Cynefin framework - The sense-making framework: four domains (Simple/Clear, Complicated, Complex, Chaotic) defined by the relationship between cause and effect. The key structural claim: the cliff edge between Simple and Chaotic - systems that suppress all variation to achieve optimised predictability have no response to novelty, so disruption sends them directly into chaos. The most costly error is domain misclassification: treating Complex as Complicated means applying expert analysis to a problem that can only be solved by probing. (Cross-listed under Strategy - this note genuinely belongs in both domains.)

Systems thinking - Systems have structure that produces their own behaviour: feedback loops, delays, stock-and-flow dynamics. The central implication for intervention: acting on the wrong leverage point is not just ineffective but may make the problem worse (because systems have shadow purposes - subsystems that appear to serve the official function may actually be maintaining the system’s stability, and removing them produces second-order failures).

Ashby requisite variety (Ashby’s Law of Requisite Variety) - A control system must have at least as much variety as the system it regulates. A system with fewer possible states than its environment will eventually encounter a situation it cannot respond to - and will fail. The implication: designing out variation (standardisation, automation) produces control only within the envelope of anticipated conditions. Beyond that envelope, the controller has no moves.

CAS emergence nonlinearity coevolution - Complex Adaptive Systems are defined by: emergence (system-level properties not present in any component), nonlinearity (small changes can have large effects and vice versa), and coevolution (agents in the system adapt to each other, so the system is never static). These three properties are what make CAS resistant to pre-planned intervention - the system changes in response to your intervention.

Failure modes

Normal accidents perrow - Charles Perrow’s thesis: in systems that are both complex (many components in non-linear interaction, with limited understanding of the whole) and tightly coupled (one component’s failure propagates to others without time to intervene), accidents are normal - they are a structural property of the system, not a result of human error. The policy implication is not to find better people but to either reduce complexity or reduce coupling - or accept that accidents will continue at a predictable rate.

Failure margins and feedback loops - The design conditions that make learning possible: wide partial-failure margins (enough slack that failure is survivable rather than catastrophic) and tight feedback loops (quick, accurate signals about what the system is actually doing). Organisations that have neither will not learn because they neither survive failure nor notice it in time to update.

Reliability resiliency recovery distinct - Three properties that are often conflated but are structurally different: reliability (consistent performance under normal conditions), resiliency (capacity to absorb disruption and maintain function), recovery (speed of return to normal after disruption). Optimising for reliability often degrades resiliency (by eliminating the slack that absorbs shocks). Recovery is the neglected one - much post-incident work focuses on reliability improvements that don’t change recovery speed.

Legibility imposes monoculture (James C. Scott) - Making a system legible (readable, mappable, controllable by a central authority) requires simplification - standardisation, categorisation, elimination of local variation. That simplification degrades the diversity that was often load-bearing: the variation that seemed like inefficiency was frequently the system’s resilience mechanism. “Scientific forestry” (monoculture for yield measurement) producing ecosystem collapse is the canonical example.

Systemic slow onset risk class - A specific failure mode: risks that develop slowly, through many small, individually-innocent steps, with no single moment where the danger is obvious. The system is in trouble before any single alert fires. Boiling-frog dynamics. The detection challenge requires tracking trends, not just thresholds.

Premature automation locks in error - Process improvement has a mandatory order of operations because each later step raises the cost of revising the earlier ones: optimisation embeds whatever it streamlines and automation locks in whatever system it finds, so simplifying or automating a process that should not exist makes the unnecessary faster and far harder to remove. The characteristic failure is starting at simplification or automation - building a faster, more reliable version of the wrong system - because the payoff of optimisation arrives sooner than the invisible payoff of deletion. (The applied discipline is Musk algorithm; the corrective to careless deletion is Chesterton’s Fence.)

Complexity ratchet makes forgetting the failure mode - Each technological generation inherits more complex problems, requiring knowledge spread across larger teams over longer spans, until the system outstrips any single mind. The dominant failure mode then shifts from physical decay to forgetting: a single gap in transmitting implicit, distributed knowledge can cause catastrophic failure, and building software to manage the workflows only defers the transmission problem one level of abstraction higher. The strategic upshot is that transmission infrastructure (documentation, checklists, redundant carriers) is not overhead but the primary defence against the system’s characteristic way of dying.

Institutional design

Commons as third institutional form (Elinor Ostrom) - Beyond markets (price signals, private property) and hierarchies (central authority, command), commons are a third institutional form: collectively managed shared resources with community-developed rules. Ostrom’s empirical finding: commons succeed when their rules match local conditions, have graduated sanctions, and allow rule modification by users. This refutes the Tragedy of the Commons as an inevitable outcome and reframes it as a design problem.

Representation vs dynamics architectural split - In any system with both a static model (what things are) and dynamic behaviour (what things do), the architecture must separate these cleanly. When representation and dynamics are coupled, changing the model requires changing the behaviour, and vice versa - both become harder and more error-prone. This is a systems principle with direct application to software architecture (Hexagonal architecture ports adapters).

Durable transmission favours copyable forms - What endures across generations is not what is most impressive or permanent but what is information-dense and cheaply copied. Monuments, empires, and individual genius erode with their originators; what crosses time intact is the reproducible carrier - institutions that encode procedure, culture that transmits values through behaviour, standards and checklists that fix procedure in reproducible form, and stories that carry the why and so enable adaptation. The design implication: invest in copyability and information density, not in the durability of any single physical embodiment. Paired with the complexity-ratchet note above, this is the institutional answer to why transmission rather than decay is the binding constraint in complex systems.

Connecting to other domains