11. Two Failures

As ΠH crosses and robotics extends the sweep from cognitive into physical work, a community must provision itself while members’ labour is structurally dominated. Its choice world ρ changes in two distinct, dangerous ways.

Failure 1 — The provisioning problem

For two centuries, the bridge from production to consumption was itself a concentration of µ³ in the economic-coordination domain. Expectations stacked:

• productive activity is done by people,
• performance is paid in wages,
• wages buy consumption that sustains the people who work.

Each link was doxic infrastructure, maintained like any institution. The whole chain was the institutional embodiment of the answer Robbins set aside: the community the economy was for was the community whose labour produced it.

When productive entities cease to be welfare-bearing entities, that chain does not gently reallocate. It suffers referent shift. The concentration linked human productive activity to human consumption. Non-human production sits outside what the linking institution covers. Markets coordinate exchange among participants; they do not by themselves mint new doxic concentrations tying some entities’ production to other entities’ welfare.

Redistribution is a partial repair. A basic income can create a thin new concentration in welfare-distribution space. It does not automatically reconstruct what the old chain carried: contribution, dignity, identity, daily structure, membership felt as earned standing. People notice the substitution because it is a different concentration of µ. What is needed is constitutional work on Level-3 institutions: new links from community provision to community membership without labour as mandatory intermediary—while preserving, where possible, what work carried beside income. The framework poses the design problem; it does not pretend to complete the constitution here.

Failure 2 — The doxic maintenance problem

Historically, doxa was maintained by human missionaries with limited bandwidth and mutual checks: journalists, teachers, clergy, neighbours, family talk. Redundancy and diversity meant no single pipeline could set µ alone.

Language models change the marginal maintainer. As tutoring, customer service, recommendation, drafting, and conversation shift to AI, large fractions of everyday doxic reproduction run through systems whose µ was set by training choices at a small number of labs. Two failure modes become available at scale:

Doxic capture. One µ dominates maintenance. The community’s unspoken background drifts toward whatever the dominant pipeline encodes—whether or not the community would endorse the drift if it could see it. Homogenisation of the taken-for-granted.

Doxic collapse. Incompatible µ multiply across overlapping communities. Shared background dissolves. Coordination fails because participants no longer share a prior about what questions mean. Unconstrained proliferation without coordination on installed assumptions.

Both are catastrophic in the framework’s own terms:

• Capture → resilience failure (monoculture; C unbounded on unmodelled states).
• Collapse → openness failure (disconnected mixture; no single attractor; coordination probability falls with reference overlap).

Each is the default of its own excess: capture of unrestricted concentration; collapse of unrestricted proliferation. Holding one connected distribution without crushing diversity is the central practical challenge of doxic engineering in the transition.

The contest between references

When several maintainers reproduce references, the population holds a mixture. Shares can be modelled by selection on predictive fitness (how well a reference lets adopters predict)—replicator dynamics whose solutions sit in the same exponential-tilt family one level up.

Limits:

• one fitness dominates → capture,
• several persist with incompatible support → collapse,
• interior rest point (maintained compatible diversity) needs stabilising force such as negative frequency dependence—portfolio value of rare references under varying environments.

Automating maintenance can steepen fitness gaps toward a dominant pipeline while concentrating who “reads” the environment—weakening the very frequency dependence that holds the interior. Default under concentrated automated maintenance tilts toward capture. Holding the interior is work, not autopilot.

Alignment is doxic alignment at infrastructure scale

Section 7’s two-factor requirement does not need repeating. The transition’s news is scale: when the marginal maintainer of a community’s doxa is AI, misaligned µ is no longer one quirky chatbot. It is the shared reference itself drifting or fragmenting. Capture and collapse are large-scale doxic misalignment. Engineering µ—which RLHF already does—is a first-class safety target beside engineering V, and at maintenance scale the stakes are constitutional.

Next: the closing convergence—one object, three faces of the same governance problem—and choices a room can actually make.