Bad Equilibria Are Not One Thing

I. Bad equilibrium is not a diagnosis

“Inadequate equilibrium” (Yudkowsky), “Moloch” (Alexander), “iron cage” (Weber), “tragedy of the commons” (Hardin) — the family of names for stable bad states is large and well-developed. Each captures a real surface condition: a system can remain broken even when better states are visible; individual rationality can produce collective stupidity; no single actor may have enough free energy to move the whole system.

These names work at the surface layer. Repair requires the layer below: a shared drift mechanism can still require different missing primitives.

A repair primitive is the smallest institutional capability whose absence keeps the failure stable: a category, an owner, a route, an aligned incentive, a commitment device, a capacity stock, a ledger. The primitive is the institutional artifact beneath the policy. Policy without the primitive has nothing to attach to.

“Fund healthcare better” is not a primitive. “Book deferred care as a liability with a named owner and a review trigger” is closer. “Raise awareness” is not a primitive. “Create a public signal that changes what each actor knows the others know” is closer. “Hold someone accountable” is not a primitive. “Assign authority, deadline, budget, and consequence to a named actor” is closer.

The diagnosis precedes the repair. The diagnosis says which primitive is missing.

II. Five diagnostic gates

Before assigning a repair, walk five diagnostic gates:

1. Object gate. Can the system see and name the problem as a stable object?
2. Destination gate. Does responsibility for the harm have an address?
3. Payoff gate. Does the responsible actor benefit from repair?
4. Joint-move gate. Can the necessary actors move safely together?
5. Execution gate. Can willing actors implement?

The gates are diagnostic axes, not mutually exclusive boxes and not a causal sequence. Real failures can miss several gates at once. The order is a repair-ordering discipline: form the object before assigning responsibility for it; assign the destination before designing incentives for that destination; separate payoff failures from joint-move failures before building capacity. A bank run shows why the distinction matters. Depositors may prefer the stable equilibrium, but without assurance each still has reason to run if others run. The payoff gate and the joint-move gate therefore remain separate.

Two audits then run across the gates. The bonding audit asks whether several gates are locked together so that single-gate repair snaps back. The absorption audit asks what hidden variable breaks when the visible target is preserved. Composite failures require repair ordering. Bonded failures require decoupling before ordinary repair ordering can work.

III. The five primitive families

1. Object failures

The system cannot see, name, or stably hold the problem as an object.

The data may not exist. It may exist but fail to enter institutional categories. The phenomenon may lack a stable public boundary, evidence standard, or response condition. Operators may see it; the institution cannot book it.

The native specimen is deferred infrastructure maintenance booked as savings because no liability ledger exists. The maintenance debt is real and accumulates; the ledger does not exist, so the debt does not appear on any balance sheet. The missing primitive is the ledger. The repair class is object formation: measurement, category, accounting rule, evidence standard, data model, public boundary. The Framing Machine develops the academic-cluster instance: the institution’s container determines what it can recognize as a contribution.

The attractive wrong repair is exhortation to “take the problem more seriously.” Without the object, seriousness has nowhere to land.

2. Destination failures

The problem is visible and named, but has no correct destination for repair.

This is the telos gap family. The harm is visible, but no public actor must receive it, answer for it, continue repair, and face a trigger when it worsens. The signal can circulate indefinitely because no address exists. Routing failure is the adjacent case where the destination exists and the signal does not reach it — Optionality Has No Router develops that case. In a telos gap, routing harder does not help because the address does not exist.

The native specimen is a cross-jurisdiction harm whose mandate, budget, and trigger sit in different institutions, so no single owner has the standing to act. The missing primitive is assignment. The repair class is destination assignment: owner, route, escalation path, response duty, deadline, audit trail. If no owner exists, the first repair is making ownerlessness itself public and procedurally live.

The attractive wrong repair is forwarding the signal to more inboxes. Volume on the routing layer cannot substitute for an address.

3. Payoff failures

The owner exists but is locally rewarded for sustaining the failure.

The metric rewards the proxy instead of the target. The agent benefits while the principal pays. The regulator serves the regulated constituency. The incumbent extracts rent from the position. The actor may understand the problem perfectly; the local game still pays for the wrong behavior.

The native specimen is the Campbell–Goodhart pattern: a quantitative indicator used for decision-making becomes subject to corruption pressure and distorts the process it monitors. The missing primitive is a payoff structure that aligns local reward with the principal’s actual goal. The repair class is payoff redesign: metric redesign, contract redesign, monitoring, anti-capture architecture for ordinary capture, countervailing power, aggregation of harmed parties.

The attractive wrong repair is adding more metrics. The new metric gets gamed in the same pattern; the system now performs compliance at two levels.

4. Joint-move failures

The problem is visible, owned, desired, and technically possible, but no actor can move safely unless others move too.

The native specimen is a bank run (Diamond–Dybvig). Depositors may prefer a stable bank, but if they believe others will withdraw, withdrawal becomes rational. The missing primitive is assurance. Each actor wants the good state privately; private desire alone does not make public movement safe.

The repair class is a credible joint move: common knowledge, focal point, sequencing, assurance contract, deposit insurance and analogous guarantees, enforcement, commitment device, or institution that changes what each actor can safely expect from the others. Many “Moloch” cases live here. The Egregore’s Button develops the recursive-belief structure that sustains the bad-coordination side of the equilibrium.

The attractive wrong repair is exhortation. Telling each actor to cooperate does not change what each actor can safely expect from the others.

5. Execution failures

The right actor exists, owns the harm, has the right incentives, can coordinate the move, and still cannot act.

The owner lacks staff, tools, competence, authority, budget, implementation path, or decision-rule clearance. The design exists but cannot be built. The mandate exists but the operational layer cannot carry it.

The native specimen is the implementation gap (Pressman & Wildavsky): a program with apparent political consensus and budget still fails to deliver because the operational chain has too many independent clearances, none of which is the binding constraint alone. The missing primitive is operational capacity matched to mandate.

The repair class is operational: staffing, tools, competence, simplification, authority, decision-rule reform, implementation pathway design.

The recurring misdiagnosis is to read capacity failure as unwillingness. The owner appears to refuse because the outcome does not change. Refusal and incapacity produce the same surface. If the missing primitive is capacity, punishment exits competent people and deepens the failure.

IV. Bonded failures: when gates lock together

The five-gate sequence works for ordinary composite failures. It fails when the gates are bonded.

In a bonded failure, the missing primitives are not merely adjacent. They enforce each other. The system cannot represent the problem because the dominant epistemic frame excludes it. No owner receives the harm because the status structure treats the harm as illegitimate. Incentives reward preservation of the frame. Coordination punishes visible defection. Execution capacity remains inside the same cluster that would have to be bypassed.

This is the structure Trapped Equilibria calls concentric capture. It is a multi-axial bonding failure, not an ordinary incentive failure. Payoff redesign alone does not repair it, because the payoff surface is held in place by epistemic authority, status dependence, material channels, and coordination expectations together.

The repair primitive is decoupling: alternative evidence standards, alternative status channels, alternative funding, alternative coordination substrate, or a boundary position from which actors can move without paying all four costs at once. In bonded failures, repair starts by creating somewhere else to stand. The distinction matters because composite failures admit linear repair (in order); bonded failures do not, and the order itself becomes a trap if attempted naively.

Decoupling is not aimed primarily at converting the core. In a captured cluster, the core has usually metabolized the dominant frame and defends it sincerely. The live repair audience is the periphery and the boundary actors: people who can see the better frame but need somewhere to move without losing all four axes at once.

V. The absorption audit

After any proposed repair, run the absorption audit: when the visible target is protected, what hidden variable breaks instead?

Conservation of failure: when the constraint remains and the visible target is protected, blocked contradiction moves to the nearest unowned variable.

Queues, quality, burnout, deferred maintenance, under-enforcement, institutional memory, debt, discretion, informal rationing, and future people can all absorb failure while the official target remains intact. The dashboard says “on track” because it measures the protected variable, not the variable carrying the cost.

A care guarantee can be preserved by lowering effective service quality. A balanced budget can be preserved by deferring infrastructure maintenance. A waiting-list target can be preserved by triage gaming. A school metric can be preserved by excluding harder cases. Fiscal savings can be preserved by burning staff capacity and institutional memory.

The repair class is accounting and ownership: hidden-vector identification, long-horizon ledger, reserve requirement, automatic review clause, capital-stock accounting, and an owner for the displaced variable. Full Accounting develops the accounting form: when a capital stock has no ledger, it can be consumed without appearing as a cost.

The same audit runs across time. A cost pushed beyond the decision-maker’s term, budget cycle, career, or institutional memory is still absorption: the hidden variable is the future stock that pays. Pension promises, deferred maintenance, demographic decline, and debt can be visible in principle while remaining politically absent because the cost arrives outside the acting system’s feedback horizon. The repair is to make the future stock present: long-horizon ledgers, automatic triggers, reserve requirements, demographic floors, debt brakes, and review rules that convert delayed consequences into proximate signals.

The audit question is: “What had to break for the target to be met?”

VI. Wrong repairs create new failures

Misdiagnosis creates new failure modes.

• Routing harder when no destination exists. The signal becomes ambient noise. The harm now appears watched, which legitimates inaction.
• Adding metrics to a payoff failure. The new metric gets gamed in the same pattern. The system performs compliance at two levels.
• Raising awareness inside a captured cluster. The cluster’s immune response activates: defensive narratives, identity-based rallying, escalated purity tests. Awareness coupled to no changed cost surface hardens the equilibrium it intended to break.
• Reading capacity failure as unwillingness. The most competent operators exit. Failure persists. Replacement operators have lower capacity. The cycle deepens.
• Org-chart shuffle for a destination failure. Agencies are merged, split, renamed, or moved under a new ministry, while no actor receives the harm with authority, deadline, budget, and trigger. The address looks new; the telos gap remains.
• Owner without authority. A named owner is created without authority, budget, deadline, trigger, or consequence. The reform produces someone to blame, not someone able to repair, and absorbs the next decade of reform attention.
• Applying single-gate repair to a bonded failure. The repaired gate snaps back because the other gates still enforce the old equilibrium. A new metric, owner, or committee becomes another capture surface.
• Protecting the visible target without auditing absorption. The target survives by burning queues, quality, staff capacity, deferred maintenance, or future people.
• Repair that prunes its own option-value. A primitive that solves the visible failure while permanently foreclosing future moves — irreversible delegation, lock-in commitment, alternative-substrate destruction — buys present resolution against future adaptive capacity. The absorption audit catches this only after the option is already gone.

A repair primitive applied to the wrong failure mode becomes another failure mode.

The diagnostic separates repair from theatre.

VII. Composite specimen: healthcare queues

Healthcare queues are useful because the same surface can instantiate every gate. Long waits may mean the system lacks reliable queue data, lacks a joint owner for deferred care, rewards regions for gaming a waiting-time guarantee, lacks a credible sequence among ministries and providers, lacks staff or authority, or is preserving formal compliance by burning quality and burnout.

The surface leaves the machine undiagnosed. “Queues are growing” only names the symptom. The mechanism audit must say: this failure is currently blocked by X; the missing primitive is Y; the attractive wrong repair is Z. In real healthcare systems the diagnosis is usually a composite of several gates plus an absorption layer; the audit’s job is to name them in repair order, not to pick one as the whole story.

VIII. Who runs the diagnostic

A taxonomy of failures presupposes someone running the diagnostic. That presupposition is itself a primitive.

In ordinary failures, internal diagnosis can work. In captured clusters, the diagnostic operator must be structurally protected from the cluster it audits. Otherwise the diagnostic becomes part of the equilibrium: the cluster names the failure in the way that preserves its own position. This is diagnostic capture: the institution that should name the failure is absorbed into the failure’s own ontology. It does not need to lie. It only has to classify the problem in the category that preserves the existing equilibrium. Capture failures get redescribed as capacity gaps when the cluster wants resources. Capacity failures get redescribed as unwillingness when the cluster wants discipline. Absorption failures disappear because exposing them would reveal the variable the institution has been silently consuming.

A protected diagnostic operator needs five properties:

• Independent mandate and funding outside the audited cluster’s appointment, budget, and removal channels.
• Access to evidence, including the hidden variables the cluster would prefer not to expose.
• Public output, so the audit cannot be absorbed as internal process.
• Right of reply for the audited cluster, so the audit does not become arbitrary unaccountable power.
• An escalation trigger that connects findings to repair, replacement, funding, or public registration of ownerlessness.

Those protections turn external audit from deus ex machina into mechanism.

The operator also needs a low-enough information gradient to the system it audits. Independence is not enough. A distant auditor can be formally protected and still operate on lossy reports, delayed indicators, and formalized proxies that omit the load-bearing mechanism. The audit must either sit close enough to see the mechanism or be designed to import local knowledge without being captured by it. The Information Gradient develops the distance-degradation case in depth.

Existing audit institutions show that most of this stack is partially instantiated. The UK National Audit Office is independent of government and the civil service, reports to Parliament, scrutinises public spending, and publishes value-for-money work. The European Court of Auditors is the EU’s independent external auditor, checks whether EU funds are collected and used correctly, publishes findings and recommendations, and reports suspected fraud to OLAF or EPPO rather than investigating fraud itself. Parliamentary ombudsmen carry pieces of the same stack in citizen-complaint domains. These institutions are independent of the audited departments, but not outside the wider political system: their budgets, remits, and follow-through still depend on parliamentary or treaty architecture. Their evidence access is real but bounded by mandate. What they most consistently lack is an escalation trigger strong enough to force repair when the audited system prefers absorption. Audit without escalation becomes high-quality ambient noise.

A mechanism audit must therefore output four things: the gate diagnosis, the missing primitive, the bonding warning, and the absorption warning. The gate diagnosis says what kind of stuckness this is. The missing primitive says what must be built before policy can attach. The bonding warning says whether single-gate repair will snap back. The absorption warning says which visible success would merely move failure into a hidden variable.

The escalation path looks different in different cases. For ordinary failures, escalation routes through the existing institutional stack: budget, replacement, statutory duty, court-like remedy, parliamentary trigger. For bonded failures, that route is closed by definition — the parliamentary, statutory, and court machinery is part of what is captured. The audit’s usable output in a bonded case is not a repair ticket for the core. It is a coordination focal point for the periphery and for boundary actors: a public, evidenced specification that lowers the cost of moving to alternative substrate and registers the captured cluster’s ownerlessness in a way later legitimacy will be able to attach to. Decoupling and reform are different escalation paths; the audit format is the same.

This leaves the bootstrap problem. Naming the missing primitive does not create the authority to install it. A mechanism audit is a repair specification, not political power. Where the escalation path exists (ordinary failures), the audit can attach to it. Where it does not (bonded failures), the audit becomes part of the legitimacy infrastructure that boundary actors use to build the alternative. Either way, the audit has reduced confusion before it has produced repair.

IX. Close

The corpus map for this taxonomy is direct: object failures point to The Framing Machine; destination failures to The Telos Gap and Optionality Has No Router; payoff failures to The Governance Alignment Problem; joint-move failures to The Egregore’s Button; execution failures to Steering Power and When Does Reform Actually Happen?; bonding to Trapped Equilibria; absorption to Full Accounting; the diagnostic function to Full-Stack Survival; and the institutional seat of the diagnostic operator to The Fourth Branch. This essay is the repair-ordering layer that connects them.

A civilization cannot repair bad equilibria by naming them all bad equilibria. It has to know which gate is blocked, whether the gates are bonded, and where failure is being absorbed.