THE GOVERNANCE ARCHITECTURE

What Must Be Built So That What the Architecture Does Is Survivable

Position Statement

This chapter derives the institutional form the preceding chapters establish is required. It establishes that the architecture carries three obligations rather than one, and that each of the three answers a condition the other two cannot reach. An arrangement that discharges any one of them while failing the others is incomplete in a way the framework can state exactly, and that statement is the Governance Completeness Condition. It further establishes that the structure now deciding what a population sees carries no reading of whether what it circulates stands in any relation to a governing fact, that this is a design choice rather than an architectural necessity, and that it is therefore answerable. The property is Referent Blindness, and the practice standing against Referent Blindness is Epistemic Hygiene.

Dependency Statement

This chapter requires prior understanding of:

• The Ascent Selection Principle and its three design results (Chapter 8, Section 8.12.4)
• The three management regimes and the Governed Band (Chapter 13, Section 13.10)
• The Relational Subcriticality Principle, the identity of the couplings through which Individuals constitute one another with the loops whose gain a population must hold down (Chapter 16, Section 16.14)
• The Leadership Compression Principle, which fixes what concentrated authority can impose whatever a population believes, together with its Time-Buying Principle (Chapter 17, Sections 17.7 and 17.7.6)
• The Dual-Locus Ignition Theorem (Chapter 17, Section 17.7.8)
• The Hardness–Referent Matrix and its Referent-governance axis, together with the Aesthetic Drift Principle that fixes the off-grid condition (Chapter 14, Sections 14.8.3 and 14.8.4)
• The five channels of the Knowledge Patch (Chapter 18, Section 18.14)
• The relationally constituted Individual established in Chapter 19

It draws most heavily on three axioms: Axiom 3 (Source Opacity), Axiom 6 (Differential Satiation Dynamics) and Axiom 8 (Bayesian Belief Dynamics). It draws on Axiom 4 (Confabulation) and Axiom 9 (Agent-Based Causal Default) for the removal instruments.

Derivation Preview

The chapter derives its central claims from six established results.

1. The Dual-Locus Ignition Theorem fixes where the capacity to ignite a collective transition sits.
2. The Leadership Compression Principle and its Time-Buying Principle fix what an operation on authority can and cannot achieve.
3. The Saddle-Node Default, the finding that a system of this kind goes over like a switch rather than swinging back, and the Incidental Buffer Principle fix what a structure must contain to convert the crisis geometry rather than defer it.
4. The Ascent Selection Principle fixes what selection can be made to do.
5. The Relational Subcriticality Principle grounds the obligation to build any of the architecture.
6. Axiom 8 (Bayesian Belief Dynamics) and the Hardness–Referent Matrix fix what a Belief-transmitting structure can and cannot read about what it circulates.

The chapter derives the Governance Completeness Condition, Referent Blindness and the practice of Epistemic Hygiene that stands against Referent Blindness, the Subcriticality Precondition, which fixes the terms on which an engineered structure may carry a restoring sign, and the Admissible Indicator with its four conditions.

Chapter Preview

Section 20.1 demonstrates that provision closes one route into crisis and cannot reach the other, so that the remaining lever on the second is the authority gate. Section 20.2 establishes that a gate acts on the interval before a decision becomes consequence and does not act on the couplings that determine whether the underlying instability persists, so a second and independent structure is required to convert the geometry. Section 20.3 shows that the arrangements a durable society accumulates are candidates for that structure and do not convert, and that the arrangement most widely credited with self-correction fails the test decisively and in a direction that makes matters worse. Section 20.4 finds one structure that already satisfies two of the three conditions at population scale while carrying the wrong sign, so that the design question there is reversal rather than invention.

Section 20.5 develops the four architectural requirements into institutional form, once selection, removal and irreversible authority are treated separately. Section 20.6 establishes that both the gate and the converting structure must be readable from outside, which places a condition on measurement that has to be satisfied before authority is granted rather than after. Section 20.7 derives the obligation to build the architecture from the persistence condition of the field itself rather than from a moral premise added to it. Sections 20.8 to 20.10 turn the framework's own tests on the framework, bound what the chapter does not claim, and state the criteria that would disconfirm it.

---

20.1 What Provision Cannot Reach

A society can feed and house its way out of one route into crisis. The other route it can only govern its way out of.

Not everyone is in a position to start a collective crisis. The capacity sits in two places and is largely absent in between, which is the Dual-Locus Ignition Theorem (Chapter 17, Section 17.7.8; Appendix A, Section A.5.8). One place is the floor, meaning the worst-off part of a population. The other is the apex, meaning whoever holds concentrated power at the top.

At the floor, sustained deprivation does two things at once. Comparison between people intensifies and begins to feed on itself faster, and the ordinary processes that talk a population down from a false conclusion get slower. There is a depth of deprivation at which the first outpaces the second whatever the population currently believes, and below that depth the crisis builds itself.

At the apex, a single Individual can impose a crisis on everyone else, provided four things hold of their position: they command the resources a crisis turns on, they control the channels through which most people learn what is happening, they appoint the people who hold the other offices, and those below them depend on them for their livelihoods and their safety. If that person's motivation has come to be dominated by status comparison rather than by bodily need or genuine interest, those four capacities are sufficient. What everyone else believes does not enter into it.

The A-Floor drive can be taken to zero. Hold the worst-off above the Appetitive floor, the level at which bodily needs are met, and the forcing disappears. This is why the framework treats universal security of that floor as something the architecture requires rather than as a value held alongside it. It falls out of what keeps the system stable, and it holds regardless of whether anyone finds it reasonable, or even desirable.

The mimetic drive to reach the apex cannot be taken to zero, and not because appetitive provision there is hard. Appetitive provision there is complete in what the position commands, food and shelter and the means of rest, though a crisis can deny its holder the last of these. Fatigue at the apex supplies no third candidate. It narrows the capacity that carries the I-Override (Appendix A, Section A.2.2), so it changes how securely the dominant source is held in check, not which source dominates. Whoever holds the apex commands the resources a crisis turns on, which is the first of the four capacities, and commanding them entails being fed by them: no president or king has starved in office while exercising full control of the governance apparatus (i.e. not technically in power while imprisoned or under siege). The one Individual whose motivation provision cannot reach is the one Individual whose Appetitive floor was never in doubt.

This also fixes what an apex composition can be. With the Appetitive channel satisfied by the position itself, the dominant source in whoever holds that position is either Intrinsic Motivation or Mimetic Desire, and there is no third candidate. The apex condition is a question with two answers rather than three.

What remains is a drive with no stopping point. Appetites stop when they are met: eat enough and hunger goes. Status comparison does not, because what counts as enough is set by where everyone else stands, and that keeps moving. This is Differential Satiation Dynamics (Axiom 6). Appetitive or Intrinsic provision at the apex is therefore not insufficient. It is fulfilled and it was fulfilled before any policy was written. And because the person carrying the drive can reach the point of acting on evidence they have produced themselves, nothing the population thinks slows them down.

Provision closes the first route and passes straight through the second.

What remains in the apex condition is authority. It is the only part of that condition a society can get hold of, and set the criteria for, which is why all four architectural requirements below operate on authority and none operates on the drive. Provision matters no less than governance. Provision also funds what governance asks of a population. The vocabulary, the Epistemic Suspension and the audit the later sections require are spent out of the same finite envelope as everything else, and that envelope bounds the capacities they run on (Chapter 9, Section 9.13; Appendix A, Section A.2.2). A population held below its A-Floors cannot afford them. In the governance case, provision is simply already a solved issue.

Imposition through concentrated authority runs on a timescale of days to months: an order is given, an appointment is made, a broadcast channel is controlled. Correction at population scale runs on months to decades, because vocabulary has to reach people, A-Floors have to be built, space and resources for genuine interest have to be protected, and visibility arrangements have to be legislated (Chapter 17, Section 17.7.2). The floor locus arms slowly and in public, so there is time to see it and act. The apex locus arms inside one person's mind, and since the cascade there feeds on evidence that person produces themselves, its timing has nothing to do with how large the population is, how sceptical, or how well informed.

One objection from the historical record says the gap is narrower than that. An imposed decision still has to pass down through people who carry it out, and a single refusal anywhere in the chain stops it. That is an accurate account of what has happened and a poor account of what can be relied on. Refusal is absent by construction wherever execution is automated. It is weakened wherever appointments have been made on loyalty, which is the same appointment power the apex condition already names. And it cannot be counted on at population scale, because nobody designed it: it is a lucky property of how a particular chain of command happened to be staffed (Chapter 17, Section 17.7.6). A society surviving on it is not protected. It is lucky.

---

20.2 What a Gate Does, and What It Cannot Do

Authority is the only part of the apex condition a society can act on (Section 20.1).

Spreading the four capacities across many hands constrains the apex directly. Where command over resources, control of information, appointment power and economic leverage sit with different people, each of whom can refuse, no single one of them can shorten the gap between a decision and its consequences below the time a correction needs. That is the Time-Buying Principle (Chapter 17, Section 17.7.6), and what it buys is the time, not the correction.

Separation of powers was designed for precisely this purpose: put the capacities in different hands so that any one holder can refuse what another proposes. It works only where the holders differ in the respect that matters. A filter stops what it would object to, so where everyone holding one of the offices shares the relevant beliefs about what is happening and the relevant desires about what should be done, none of them objects, and authority spread across many hands behaves exactly as it would if one person held all of it. The number of offices has not changed. What the offices do has, and the arrangement looks identical either way.

Independent filters compound: the chance the imposition of a crisis gets past all of them is the product of the chances that each one of the filters fails to stop it, so every additional filter designed to avoid a crisis by its own set criteria drives the total down steeply rather than by a fixed amount. That multiplication is valid only while the filters are independent. Office-holders who share their beliefs are correlated, and a set of correlated filters is many filters acting as one as porous as the first, whatever the constitution says about how many there are. What decorrelates the filters is what each office is required to be good at. A criterion tied to a role's own craft cannot be satisfied by holding the right beliefs, where a criterion of general political judgement can, so offices selected and held to account against separate craft standards are filters that genuinely multiply while offices selected on a shared political test are one filter wearing several names. The design that follows from this is developed at Section 20.5.

What none of this touches is the feedback loop. A microphone held too close to its own speaker picks up what the speaker plays, feeds it back in, and plays it louder still. Whether that ends in a squeal or dies away turns on one quantity: how much of the original sound comes back around each pass. Below one, each pass is quieter than the last and the sound fades. Above one, each pass is louder, and a cough becomes a scream. Human wanting runs the same circuit. One person's visible wanting of an Object raises another's, whose visible wanting raises the first's again, and the fraction that returns each time round is the loop gain. Everything the framework says about crisis is a statement about whether that loop gain can rise above one or stays below it.

Two operations are available and they are not substitutes. Distributing authority acts on how fast a decision can become irreversible. Reducing the five channels through which comparison compounds acts on the loop gain itself, and removes the configuration that produces the dangerous decision at all. They run on timescales an order apart: lowering the gain is the work of a generation, and distributing authority can be begun in the time it takes to make one decision. The first holds a population together long enough for the second to finish, and the second is what eventually ends the dependence on the first.

A gate cannot do the second job of lowering the loop gain. Four quantities multiply together to decide whether comparison decays or escalates: how strongly the Salience already standing around an Object drives a person's wanting of it; how much status comparison contributes to the Decision Hub when the mind totals up what it wants; how much a person's realised wanting of an Object raises the Salience and standing that Object carries; and how steeply that Salience responds at the level the population is currently running at. Who holds which office is not among those quantities, and no rearrangement of offices puts it among them. An arrangement that changes office-holders has altered a quantity the loop does not contain, so the loop pushes on that quantity and receives nothing back from it. Distributing authority, however thoroughly it is done, leaves the instability exactly where it was, and a society that has done it well has bought time rather than removed the danger.

Converting the geometry of crisis requires something the architecture does not contain. A system can lose its stability in two ways. It can begin to swing, in a pattern that stays within bounds and passes back through where it started. Or it can go over like a switch, in one direction, to a place from which returning the pressure to its former level is not enough to bring it back. This is a collapse. The three components that close the fast loop supply no restoring force of their own, so the second (the collapse) is what the architecture does by default, which is the Saddle-Node Default (Appendix A, Section A.5.3). Reducing the existing couplings, short of bringing the loop's capacity below one, lengthens the run and lowers how far the population climbs before it arrives, and it leaves the destination where it was.

What changes the destination is a structure with three properties. It carries a state of its own that moves over time. Its response acts on one of those four multiplying quantities rather than beside them. And it pushes back against the loop rather than with it. A structure with all three holds the population in a Governed Band: kept below the Salience Ceiling by something actively pushing back, with room to move along the slope still live, rather than pinned at the ceiling because it has run out of room to climb. In that condition the population is still responsive, there is still margin before the collapse, and the way back is still open (Chapter 4, Section 4.4.13; Chapter 13, Section 13.10).

Three obligations therefore fall on the architecture.

The three obligations follow from four established premises:

From the Dual-Locus Ignition Theorem: the apex route is a point source that one Individual suffices to open, and the only controllable term in it is authority.

From the Time-Buying Principle: an operation on authority acts on the interval between decision and consequence, and authority appears in no coupling forming the Internal Loop Gain, so no such operation converts the geometry.

From the Saddle-Node Default: the geometry is converted only by a structure carrying a variable of its own whose feedback acts against a coupling in the Internal Loop Gain or the Network Loop Gain, and by nothing else.

From Source Opacity (Axiom 3): the quantity requiring damping is not readable from inside the Individual carrying it, so a structure supplying the damping must have its own state read from somewhere other than itself.

Given these premises, an adequate architecture must gate the authority, convert the geometry, and be read from outside; no one of the three substitutes for another, and the third is a condition on the first two rather than an addition to them.

This is the Governance Completeness Condition. What is complete is the list of obligations, not any specific design that meets them. The condition says what an adequate architecture must contain. It does not say that containing those things is enough, and an architecture can hold all three and still be badly built. An architecture missing any one of the three is inadequate for a reason that can be stated rather than for one that has to be argued case by case.

The list is exhaustive over what the framework licenses. An architecture with a gate and no converting structure holds a population whose loop gain is above one at arm's length indefinitely, and depends on the gate never once failing. An architecture with a converting structure and no gate leaves open a route into crisis that one person can take whatever the loop is doing. And an architecture with both, read only by itself, has installed an actor of great authority whose conduct cannot be checked, which is worse than having installed nothing, for the reason Section 20.6 develops.

---

20.3 Why What Exists Does Not Qualify

The arrangements a durable society accumulates are candidates for the converting structure, and none of them converts.

Courts that can strike down an unconstitutional Act, a press that can publish a story that a Government office would rather suppress, powers divided so that each holds a check on the others, professional bodies with standards of their own, insurance funds held against the worst outcomes of a bad year. Some of these introduce a variable of their own and some hold a setting; some act on the couplings that form the loop gain and some act on coordinates the loop does not contain; and where all three conditions of Section 20.3.1 are met, they are met at strengths chosen for other purposes and engaged after the event they would have damped rather than before it. They are not nothing. Most simply don't act on the generating forces that produce crises themselves. They place obstacles in the way of a stream that the stream will inevitably find its way around.

What they do not do is convert. Those that hold a setting, or land on a coordinate the loop does not contain, fail a condition outright, and fail it however strong their action (Section 20.3.1). Those that meet all three fail to convert for the reason the Incidental Buffer Principle gives (Chapter 17, Section 17.17.5), which holds that every arrangement ever known to have slowed a mimetic build-up was built to answer a society's own account of what ailed it rather than to act on the loop that produced it. Each was calibrated to that account: legality, liberty, accountability, restraint, security against misfortune. None was calibrated to the couplings that determine whether Mimetic compounding decays or escalates, because a structure cannot be calibrated to a quantity that has not been identified.

Each therefore damps the quantity that matters by accident, and its strength drifts with the account that sustains it rather than with the danger it reduces. That is what makes such a structure erodible. A court is funded, staffed and defended because a population wants lawful government, so it is replenished when lawful government is wanted and left to thin when faith in the rule of law dwindles, and the loop it happens to damp appears in nobody's reason for keeping it. A guardrail nobody knows they are paying for is a guardrail nobody defends.

20.3.1 The three conditions

What is required is a Restoring Node, a structure added to the architecture from outside that carries a reading of how hard the loop is running and pushes back in proportion to what it reads. A candidate qualifies as a Restoring Node when three conditions hold together.

Novelty. The candidate introduces a state variable, meaning a quantity of its own that changes over time, that is not already in the architecture that population runs on. A social norm held at a fixed level fails this, however low the level.

There are two ways of controlling a door. A doorstop wedged under it holds the door at one angle, and it holds at that angle whether the air is still or a gale is blowing: the harder the wind pushes, the harder the door strains against the wedge, and nothing about the wedge changes. A door-closer, the sprung arm fitted at the top of a heavy door, pushes back in proportion to how far the door has been pushed open, so a hard shove meets a hard return and a light one meets a light return. Both hold the door in still air. Only one of them answers the wind. A doorstop is a constant and a door-closer is a function, and that is the difference between a norm and a governor. Editorial restraint that holds amplification down is genuinely useful, and it does not push harder when things get worse. Pushing harder as the state deteriorates is what a correcting mechanism does. Usually what happens when a norm fails to be observed or respected is that it is simply abandoned. The doorstop is removed.

Closure. That variable's feedback acts on one of the couplings whose product forms the Internal Loop Gain, or on the corresponding network-scale quantities. The Internal Loop Gain is the loop gain of the circuit running inside a single Individual. A structure whose restoring force lands on a coordinate the loop does not contain fails this however genuine and however strong the force is. A court that strikes down an unlawful act by the holder of an office has restored legality, which matters, and has not touched how strongly the Individual's visible choices drive another's wanting. The force was real. It landed somewhere else. The wanting does not subside with the ruling, so the Individual will search for another means to meet it. The wanting is not closed by the ruling.

Sign. The action is restoring rather than amplifying, which is to say the structure pushes against the direction of the loop rather than with it. This is a property of the arrangement rather than of anyone's intentions, so a structure designed to correct may carry amplifying sign and a structure built for unrelated purposes may carry restoring sign. The failure here is not symmetrical with the others. A candidate that satisfies Novelty and Closure while carrying amplifying sign is worse than no candidate at all, because it delivers force into the loop in the compounding direction while presenting as a corrective structure. It has the reach of a governor and the direction of an accelerator.

Qualification is not conversion. A qualifying node alters the geometry only at moderate gain and lag, and it fails to convert in three ways (Section 20.10, criterion 2). Where it pushes too weakly, under-gain leaves the terminus where it was while producing every appearance of working. Where it pushes too hard, over-gain suppresses Observability, meaning how much of what other people choose and value is visible to everyone else, and Observability is what carries legitimate social learning and well-functioning communities, so the node destroys the function it was built to regulate. Where it answers too late, over-lag documents the collapse rather than preventing it.

20.3.2 Two-party alternation

The arrangement most widely credited with self-correction is two parties alternating in office, or the dual factions that power struggles converge on wherever they occur. The pattern is durable, it recurs, it responds to conditions with a lag, and it looks from a distance exactly like a society correcting itself.

It passes Novelty. Incumbency is a genuine state variable with dynamics of its own, and the electoral lag is real.

It fails Closure. Replacing an incumbent changes none of the couplings whose product forms the Internal Loop Gain. Office-holding appears in no term of the Internal Loop Gain, so the loop drives the electoral variable while the electoral variable returns no damping. Drive is not feedback. Losing an election does not stop the wanting to win the next one.

It fails Sign. A two-party structure projects every conflict onto a single axis, and each cycle stages a mutual demonstration of hostility between the two camps. What that supplies is evidence. That evidence does not drive down the wanting. It is simply metabolised in the usual Bayesian way, and the update runs in one direction.

The protection a population gets from disagreement comes from having many competing accounts of what has gone wrong. Where a dozen groups each blame a different culprit, the accounts interfere with one another and none of them accumulates enough weight to move a population to decisive action. That interference is the Belief Heterogeneity Buffer, the safety a population draws from the fact that its blame is scattered. Two camps blaming each other is the opposite arrangement, because each camp's hostility is genuine and is therefore genuine evidence for the other camp's account of it. The Belief Heterogeneity Buffer does not merely weaken under those conditions. It inverts, and what was friction, the competing beliefs, becomes the fuel for the crisis.

The general test is not whether alternation occurs but whether the population returns to a prior operating point when it does. A genuine cycle recurs in the variables that matter. A Counterfeit Limit Cycle swings back and forth in one quantity while a different and dangerous quantity climbs steadily beneath it, and the two are indistinguishable within any single cycle and separable across many. Where each swing ends further along than the last, what is being watched is a ratchet wearing the costume of a cycle.

A cycle that passes this test is genuine, and it is not thereby governed. The architecture supplies no sustained cycle of its own. The memory channel through which an Individual's own earlier Beliefs are recalled and re-enter as Evidence supplies no rotation at any recall lag, and a system whose couplings all reinforce one another settles rather than swings, so a bounded cycle in a population requires a negative feedback acting on it somewhere (Appendix A, Sections A.5.3, A.5.6 and A.5.7.6). That feedback need not belong to the structure being credited, and a swing can also be imposed by a calendar, the electoral calendar among them, which drives the variable on a schedule and returns no damping. What separates a Governed Band from either is where the bound comes from. A Governed Band is bounded by an identifiable structure whose feedback acts on a coupling in the Internal Loop Gain or the Network Loop Gain. So a genuine cycle is credited to a structure only where the structure has been found and its three conditions established independently of the cycle.

20.3.3 The condition a society is actually in

Some populations are supercritical, meaning the loop's capacity sits above one so that an inflated resting state exists, and they are nonetheless not visibly escalating. A loop with that capacity amplifies wherever its response is still steep, so something must be holding them where they are. Two candidates are available, and they are opposites.

Two halls are running a concert and a meter in each reads ninety decibels. In the first hall the sound is loud and every instrument is audible, and an audience can sit in the hall comfortably for an evening. A limiter is holding the output at ninety while the equipment is still able to produce more. In the second hall ninety decibels is a howl. The microphone gain is set exactly where the gain is set in the first hall and the same signal is passing through every part of the chain, but there is no limiter, so the loop closes and the system clips. The meters agree, and what the meters agree on tells nobody anything.

The second hall has not become louder. The second hall has become useless. The ceiling fixes how much sound the room can hold, and in the second hall the whole of that capacity is filled by the system's own output returned. The music is still being played. The music cannot be heard, because there is no room left for the music. And because there is no limiter, no build-up was required to arrive here: a single transient was enough to start the howl, and one cough at the wrong moment would have done it. The audience puts its fingers in its ears and looks toward the sound desk, because the sound desk is where the remedy is believed to sit.

What corresponds to the sound in a population is Salience, and the question this chapter has to answer is Salience of what. The answer is the Salience of those who hold the apex: the officeholders, the parties contesting the office, and the contest itself, held as the Objects on which a population's problems are believed to turn.

Two things drive that Salience up and they are independent. The first is A-Floor erosion. Where the A-Floors are met, little turns on who holds the apex and the apex draws correspondingly little notice; where the A-Floors are eroded, attention moves to whoever is believed to control access to them, and the apex reaches maximal Salience because the apex stands between a population and what its members need. The second is that the problem of provision never reaches the apex occupant at all. The one Individual whose motivation provision cannot reach is the one Individual whose Appetitive floor was never in doubt (Section 20.1), so the apex remains an ignition locus whether the A-Floors are met or not. Meeting the A-Floors removes a source of apex Salience. Meeting the A-Floors does not remove the apex.

An Object can only be so salient. Past a point the apex is already drawing all the notice a population has to give, and holding the office in higher regard adds none. The level of realised valuation at which that mapping has flattened to nothing is the Salience Ceiling, and with respect to those in power the Salience Ceiling is reached early, by ordinary deprivation rather than by any crisis of governance.

One candidate is the first hall. A structure is holding the population below the Salience Ceiling and pushing back in proportion, the slope is still live, and what holds the population is the structure rather than exhaustion. Margin to the collapse is preserved. That is a Governed Band, and a Governed Band is what the architecture is being asked for.

The other is the second hall. The population has run out of headroom, the slope has flattened on its own, and nothing is regulating the population. The population is held at the Salience Ceiling rather than governed below the Salience Ceiling, stopped by a quantity the population exhausted rather than by one anybody set. That is Saturation Hold, and Saturation Hold is the condition the framework diagnoses in the present.

Saturation Hold is not a milder condition that a further push would worsen. The way back is longer than the way in. Returning the forcing to the level at which the feedback began will not stop the feedback, any more than returning the volume knob to that setting silences the room, and by the Ignition-Release Asymmetry the release level sits below the ignition level with the gap widening as the loop's capacity rises (Chapter 4, Section 4.4.13; Appendix A, Section A.5.3). Equipment held against its ceiling eventually fails, and that failure is the breakage the Saddle-Node Fold names.

What follows for governance is a requirement rather than a design preference. With the A-Floors met and the Salience Ceiling still reached, the term that remains for governance to act on is Observability, the structure determining whose realised valuations propagate into whose Salience field. Reducing the gain means reducing the Observability of the apex. Chapter 4 fixes this as the fifth of the five channels through which the loop gain can be turned down, the one that acts on the Observability matrix rather than inside any Individual (Chapter 4, Section 4.4.13.5).

Reducing the Observability of the apex is not a reduction in transparency. Accountability is information about conduct travelling outward from an office. Observability is wanting travelling inward toward the office. A population may hold a complete record of what an officeholder has done while placing very little of its own wanting on the office, and the two quantities move independently. Nor does the requirement license suppressing Observability at large, which carries cultural learning and Positive Mimesis and which a structure driving toward zero would destroy (Section 20.4.5). What the requirement reaches is the Observability of the apex as an Object of desire.

Two independent routes arrive at that requirement and they compound. The first is the loop: apex Observability is a term in the loop gain. Attention withdrawn from the apex is never simply withdrawn, since the envelope is spent at every moment (Appendix A, Section A.2.2), and where it goes decides what the withdrawal does. Spent on the non-social field, it lowers the gain. Spent on other Individuals, it leaves the gain where it was and changes who is seen, because the gain of the Observability channel is set by how much attention goes to observing others and not by who is observed (Appendix A, Section A.5.7.4). That second destination is not a failure of the design. Everybody needs to be Salient some of the time for provisioning to follow, and the attention a population concentrates on the apex is attention the least-served are denied (Chapter 8, Section 8.12.7). Returned to them, it serves the floor that apex attention starves and narrows the regard differential on which the Mimetic Premium is charged (Appendix A, Section A.10.4). The second is selection: the Salience of the position is the largest prize the field has to offer precisely because the whole field is watching the position, so high apex Salience is what recruits the occupant most disposed to raise apex Salience further, and the Ascent Selection Principle turns on exactly this (Chapter 8, Section 8.12.4). The two routes close on each other. A Mimetically dominant occupant performs for the audience, which raises apex Salience, which enlarges the prize, which recruits harder for Mimetic dominance. Lowering the Observability of the apex is therefore one intervention answering two problems, and returning that attention to the least-served answers a third. The design that delivers it is the Repulsion Architecture of Section 20.5.

That is why the appearance is the danger. A society in Saturation Hold is not accelerating, and a society in Saturation Hold will read its own lack of acceleration as evidence that the arrangements are still working. The absence of escalation is the Salience Ceiling, not a correction.

---

20.4 The Node That Is Already Installed

Every candidate tested so far has failed before the question of sign arose. A social norm held at a fixed level introduces no state variable of its own and fails Novelty. A court enforcing legality introduces one whose feedback lands on a coordinate the loop does not contain, and two-party alternation does the same, so both fail Closure. The structure examined here fails neither.

It is Computational Curation: the machine execution of the decision as to which subset of available informational material a given Individual is presented. The ordered feed of a social media platform, the algorithm that selects the next video or track, the ordering of search results, the selection and placement of news stories, the ranking that decides which offering in a marketplace is seen at all, and the artificial intelligence chat interface that answers a question in its own words rather than returning the sources to consult (Section 20.4.4). Naming the class social media understates it, because the same function is performed wherever material is retrieved, recommended or summarised. What fixes the class is not a medium and not an industry. It is that a machine decides, at a scale and rate admitting no case-by-case ruling by any person, which small subset of what it controls access to is presented to a particular Individual.

Two things have to hold for a structure to reach the loop at all: it must carry a variable of its own that changes over time, and that variable's feedback must land on one of the couplings whose product forms the loop gain. Computational Curation carries the first. Every such layer holds estimates of what a given Individual will attend to, and those estimates move continuously as the population responds, which is a state variable with dynamics of its own rather than a setting held at a fixed level.

It carries the second because of what it decides. Deciding what reaches whom is setting Observability, and Observability is what Section 20.3.3 left standing: with the A-Floors met and the Salience Ceiling still reached, the term remaining for governance to act on is the structure determining whose realised valuations propagate into whose Salience field. So the feedback of a Computational Curation layer does not land beside the loop. It lands on a term inside it. Nothing tested before this point has done that.

This structure is already installed, and it is calibrated to the wrong quantity. Nothing needs to be invented. The position a converting structure would occupy is already occupied, the coupling is live, and the response already moves with the state. What is wrong is the direction, and the direction follows from what the arrangement is scored against rather than from any fault in the building. The arrangement does exactly what it was designed to do. So what the architecture asks for here is a recalibration rather than a construction project.

20.4.1 Novelty and Closure are already satisfied

What reaches whom is no longer settled by who happens to be nearby. Shown a thing, a population clicks or does not, watches on or leaves, and the estimate of what that Individual will attend to moves. Nothing about the estimate is fixed in advance and nothing holds it at a level, which is what separates it from a norm and what makes it Novelty.

Observability is one of the five channels through which Mimetic signal compounds, and it is a term in the Network Loop Gain rather than a quantity adjacent to the Network Loop Gain (Chapter 13, Section 13.10; Chapter 18, Section 18.14; Appendix A, Section A.5.7.4). The Network Loop Gain is the same circuit measured across a population instead of inside one head: how much of one Individual's visible wanting comes back to them through everybody else. So the feedback lands on a coupling in the Network Loop Gain, which is Closure.

Two of the three conditions are therefore already satisfied, at the scale of populations, and nobody proposed the arrangement as a corrective structure.

Satisfying the conditions is not the same as being escapable, and the first objection to all of this is that a population dissatisfied with what it is shown can stop looking. What the objection misses is which part of the arrangement holds the value. The network is the user's own: the connections made, the pages followed, the accumulated record of who they know. The Computational Curation layer is a separate thing laid over that network, and it is how the operator draws revenue from it. A user who dislikes the ranking cannot decline the ranking without also abandoning the network, because the two are not offered separately. So exit is priced at the loss of an asset the user built rather than at the loss of the layer they object to, and a coupling that survives that arithmetic is held in place against the preferences of the people coupled by it rather than because of them.

20.4.2 The sign is inverted, and the objective is why

Computational Curation is not undirected. It optimises, and what it predominantly optimises is engagement.

Engagement measures what a coupled population does. As a measurement of itself it is accurate and cheap: a population clicked or it did not, watched on or left, and the count settles the matter. The defect is not in the measurement. It is in what the measurement is put in place of.

An arrangement that decides what a population sees is acting on whether the environment it constructs serves the people inside it, and that is a question with an answer. Whether claims reach people with their origins intact, whether epistemic corrections overtake the errors they correct, whether what a population comes to believe holds up when acted on: each is settled by something outside the population, and each is settled on an interval. Engagement is substituted for all of them, and the substitution is neither tested nor chosen by the population it acts on.

The relation any Belief bears to a governing fact admits four conditions, and they are the Referent-governance axis of the Hardness–Referent Matrix (Chapter 14, Section 14.8.4). A live Referent governs the Belief. A Referent was identified, returned its verdict, and the Belief is held against it. No Referent was ever identified, so no verdict exists to defy. Or no Referent is constitutively available, there being no fact of the matter to reach.

The fourth condition is the aesthetic regime, and it is harmless. A tightly coupled population evaluating a matter of taste drifts continuously and converges on nothing, which is the Aesthetic Drift Principle (Chapter 14, Section 14.8.3), and the drift costs nobody anything. There is no Referent to decohere against, so nothing is defied, no maintenance Cost falls on anyone outside the boundary, and no corrective is available or needed. A question of taste never settles, and its never settling is not a fault. The aesthetic case therefore cannot be the defect: a domain with nothing at stake produces no crisis, and this arrangement produces one.

The defect is the misclassification. Engagement optimisation treats every domain as though it sat in the fourth condition, and applies to all four the treatment only the fourth can absorb: pure coordination inside the population, with nothing outside admitted as a check. Where the domain really is aesthetic, no harm follows. The harmlessness belongs to Beliefs about form and not to the Objects that carry them. A work of Art does not place its content in the fourth condition by carrying it in an aesthetic form (Chapter 14, Section 14.8.4.1), and an arrangement that ranks a work by engagement ranks the work's content with it. Where a live Referent governs, the optimisation overrides it. Where the verdict is already available and the arrangement runs against it, that is Referent Abandonment. Where the test has never been run, the coupling hardens around the gap in the condition of Type I Exclusion, and the corrective is not obedience to a verdict but the construction of the Induced-Decoherence Test that would obtain one.

Held at high Coupling Strength, the last two are Hard Falsehoods, and the Hard Falsehood signature supplies the danger the aesthetic regime lacks. Such a domain does not decohere continuously. It exports the Cost of its maintenance onto those the boundary excludes and stores its decoherence rather than discharging it, which is why it runs crisis-free for exactly as long as it runs, and then does not.

So what the arrangement lacks is not a Referent. It is any capacity to tell the four conditions apart. That incapacity is Referent Blindness: the property of a structure that selects or transmits Belief while carrying no internal variable for which Referent condition the Belief it circulates stands in, and which therefore treats all four alike.

Referent Blindness is inherited from Axiom 8 (Bayesian Belief Dynamics). Content-Agnosticism is the property of Belief updating whereby the mechanism combining priors with evidence operates identically across all subject matters, carrying no internal variable for the Truth or the stakes of what it processes. That is how the Individual's Thinker revises the Belief Register, and the Thinker cannot do otherwise: the architecture offers nowhere to put such a variable, which is why the corrective at the scale of the Individual is a vocabulary supplied from outside rather than a repair to the mechanism. A structure deciding what an Individual sees inherits the same property one level up. There the inheritance is not forced. Nothing prevents a Computational Curation layer from carrying a reading of the Referent condition of what it circulates, and the present arrangement carries none, holding a standing reason not to acquire one, since three of the four conditions would constrain the objective and the fourth would not.

At the scale of the Individual, blindness to the Referent condition is an architectural fact and the framework works around it. At the scale of the structure deciding what an Individual sees, it is a design choice, and a choice is answerable.

Engagement is therefore not a degraded indicator of whether an environment serves the people in it. It is a coordination point installed where an indicator was owed, and optimising it holds the Referent-free condition open at population scale, passing into Referent Abandonment as each verdict arrives and is not acted on.

It is also an objective that acts exactly where the loop runs, because selecting for attention is selecting on Salience. The result is a structure carrying its own dynamics, coupled to a term in the Network Loop Gain, driving that term upward.

That is the asymmetric failure of Section 20.3.1: Novelty and Closure satisfied, Sign inverted. The present arrangement is the largest instance of that condition available, and its presenting face is not a claim to correct anything but a claim to be merely showing people what they are interested in.

20.4.3 The threshold moved before the content did

Two effects follow from the arrangement itself, and neither requires anything to be said about what is being circulated.

The first is density. The critical threshold at which a network's Belief dynamics become self-reinforcing is set by the structure of who hears whom and who observes whom. For a given average coupling, a denser structure of Evidence couplings carries a network toward that threshold: add connections and each Individual hears the same claim from more directions at once, so less persuasive weight per connection is needed to carry the whole. A denser structure of observation also lowers the threshold itself. Both hold while the attention an Individual can spend on others has room to spare, and that attention shrinks as fatigue and deprivation run the envelope down, so it is spent sooner in a population already drawn on (Appendix A, Section A.2.2). Once that attention is spent, added connections divide a fixed quantity, and what raises the coupling is a larger share of attention drawn to other Individuals, which is what an arrangement selected for attention captures. A Computational Curation layer that raises the density of observation, and the share of attention spent on it, therefore carries the network toward the threshold and lowers the threshold, even where nothing about any single connection has changed. This is the Information-Cascade Amplification Effect, the result that an information layer selected for attention drives a population toward that threshold rather than merely failing to hold it back, and the formal statement is at Appendix A, Section A.5.7.5. The arrangement moved the threshold before anyone argued about what was crossing it. The relation holds where the Belief in circulation has no live Referent doing corrective work. Where a Referent is present and respected, added connections carry the Referent along with the claim, and density can supply correction as readily as amplification. A Computational Curation layer selects for attention rather than for provenance, so the Beliefs it circulates are Referent-abandoned in the sense of Chapter 14, Section 14.8.3.1, and density is amplifying without remainder.

The second is that the arrangements that damped this coupling damped it by accident and were withdrawn for reasons unrelated to what they damped. Presentation in the order things happened, visibility bounded to chosen connections, attention organised by subject rather than by person, and the absence of any displayed count against which an Individual could calibrate: each of these suppresses the coupling, and none is adopted for that purpose. They are the Incidental Buffer Principle in its ordinary form, damping a quantity nobody has named. A buffer whose rationale nobody holds cannot be defended when the rationale it does have stops paying, and these are not defended. What they do identify is where the coupling can be reached, and the same arrangements return in Section 20.4.5 as responses rather than as fixed arrangements.

20.4.4 A generative layer raises two channels rather than one

A generative layer, which is what stands behind an artificial intelligence chat interface, produces text and images at scale and is constituted entirely by expression the population has already produced. It has no independent contact with anything outside that expression, and it can emit no proposition that does not derive from what it was given. Its output is therefore endogenous by construction (Chapter 14, Section 14.10.4.8).

Two consequences follow, and they land on different channels. Such content entering the evidence stream operates on the network Evidence-coupling, which is the channel through which one Individual's expressed conclusions become another's evidence, and it does so while also occupying the Observability channel. That is two of the five channels at once.

Content of this kind arrives without its provenance. It presents to an Individual as an additional independent source while being the population's own prior output returned with the provenance removed, so the coupling rises while every indicator a participant could construct reads unchanged.

Someone wants to know whether a claim is true, so they check it. They read it on one site, find it again on a second, and meet it a third time in a summary written elsewhere. Three sources agreeing is strong evidence, and they treat it as such. What they cannot see is that the second site copied the first, that the summary was assembled from both, and that the first was repeating a single unchecked post. They have not consulted three sources. They have consulted one source three times, and the arrangement that showed it to them removed the very thing that would have told them so.

That is Concealment Infrastructure operating directly on a quantity in the Network Loop Gain. Concealment Infrastructure is the accumulated layers of structure that keep a mimetic component from being seen or measured, and it accumulates until an event forces visibility, at which point it breaks one way rather than correcting gradually.

The Provenance Rule is the instrument at the scale of the Individual. It holds that a Belief taken from another Individual stays short of certainty until it can be traced to a test against physical reality that someone has actually run. It is not sufficient at population scale, because it asks the Individual to trace what the arrangement has stripped. What is stripped is the provenance, not the claim. The claim still arrives, and it arrives carrying nothing that marks what is missing. An Individual cannot update on evidence they do not receive, and will update on what they do, whether its provenance is traceable or not.

Grievance Beliefs are the type that produces the fold with the largest cost. What is exogenous in that outcome is not a competing account but the harm itself: lives ended, suffering imposed, and Individuals left with no retreat to an I-domain. So the governor's function on this channel is negative. A Grievance Belief circulating among Individuals in traceable contact with a deficit is bounded by that contact. Propagation across communities removes the bound, since Individuals in the receiving community have no contact with the deficit, update on the attribution by Contactless Updating (Chapter 14, Section 14.8.3.2), and receive the attribution as corroborated by its arrival.

What crosses need not carry a grievance anyone in the receiving community could state. The Predator reading moves on presentation and on repetition, and it moves faster than the coordinates that would redistribute weight away from it, so what a curating layer distributes is a target already positioned, on which a deficit arriving later has only to land. That is the crossing the requirement restricts, and it is why the restriction reaches the positioning and not the account: an account of a live grievance is answerable by a competing account, and a positioned target is not. Restricting that crossing is the design requirement, and it falls on the two channels this section has identified: the network Evidence-coupling, and the Observability channel that a curating or generative layer occupies at the same time.

The requirement falls on what the arrangement does and not on what it carries. An intervention in the Subcriticality Maintenance set is specifiable on rates and thresholds rather than on Belief content (Belief Multipolarity Principle): a body selecting which attributions to carry would have to converge on that selection in a domain where convergence is available only by coupling, so its own determination carries the diagnostic signature of the condition it was installed to correct. Selective withholding also puts the governor at the gate. The party whose expression is withheld reads the withholder as blocking, which is a Grievance Belief accurate on its gate-reading, and acting against that party supplies the confirming conduct the adversarial branch runs on (Appendix A, Section A.11.8). Uniformity keeps the arrangement out of the gate position: a limit falling on every crossing creates no target, and a limit falling on selected content creates one.

Declining to lend reach and contracting how far a crossing travels are Observability Governance, acting on the spectral radius of the Observability Matrix and so on a multiplicative factor of the fast-channel Network Loop Gain rather than on a term in a sum. Limiting the rate at which a crossing propagates and carrying provenance forward undestroyed act on the slow channel, where the sensitivity of an Individual's updating to arriving Evidence is inflated by exactly the correlation a trail exposes: an Individual holding no channel that distinguishes restatement from independent corroboration credits arrivals one by one, so accumulated mass tracks the count of restatements rather than the weight standing behind them (Appendix A, Section A.5.7.2). Sustained by a channel that does not pause, that crediting is the Ambient Evidence Amplification Effect (Chapter 13, Section 13.4.8). Marking a restatement as one collapses that count toward the number of sources, so the correction scales with the laundering it exposes, leaving a multi-sourced attribution near its correct weight and cutting a manufactured cascade to the weight of its single origin. Which content qualifies for unsolicited delivery at all is settled by the Coupling-Decoherence Content Qualification Gate, on the Coupling Strength a claim carries and the Decoherence Time of the domain it belongs to, and not on what the claim says.

A further factor stands in the same product and is not the arrangement's to set. The Evidence-Emission Coefficient governs how firmly a maximum-posterior Hypothesis is emitted as Evidence from the slow layer, and it falls as the Interference Term rises, the Interference Term being the residual coherence a Belief Register holds between competing Hypotheses. An Individual sustaining Epistemic Suspension therefore emits provisionally rather than as fixed Evidence, and the population-averaged coefficient standing in the slow-channel network gain falls with the practice (Appendix A, Sections A.5.6 and A.5.7.4). That factor is installed by the Knowledge Patch rather than by the governor, and because the factors multiply, the arrangement's levers and the trained capacity compound rather than sum. A preserved trail also reduces credited arrivals only where the recipient can read one, which is the precondition Section 20.4.5 states, so the trained capacity is what converts the provenance lever from an availability into a reduction.

Withholding the claim moves none of these factors. It removes one arrival from one sum, and by the test Section 20.4.5 sets, the size of the response must move with the state: a ban pushes the same amount whether the population is calm or critical, while the provenance correction pushes harder as the loop runs hotter, a hotter loop carrying more restatement per source and therefore more correlation to expose. The same arrangement must carry Cascading Support, so that provision reaches the deformed field, and must assist Cascading Consequences, so that correction reaches the generating field. Provision and correction are what the governor is for.

20.4.5 What sign reversal requires

The governor is already installed. It is the selection that decides what each Individual sees, and it already has everything a governor requires: a state of its own, a position inside the loop, and a response whose size grows with what it reads. Nothing needs to be invented. What needs changing is what the Computational Curation layer reads and which way it pushes, because it currently reads engagement and it pushes with the loop rather than against it. Optimising for engagement is not one fault among several that a better arrangement would tidy away. It is the positive sign. Five requirements follow for reversing it, and the institutional forms that satisfy them are underdetermined by the architecture.

The objective must be scored against something outside the population. Engagement fails because the arrangement never checks against the Referent it displaced (Section 20.4.2). What replaces it must be scored against what the structure did to the information passing through it, and Epistemic Hygiene is the name for the class of practices in question: those maintaining the conditions under which the Referent condition of a claim can be determined and its certainty held accordingly. The determination is the one Referent Blindness omits. It sorts what is handled by the four conditions, and it is made by tracing provenance rather than by weighing content. What it licenses is a filter. A claim a live Referent governs may be converged on; the other three warrant Epistemic Suspension, and they are most of what a population receives, which is why an arrangement that cannot tell them apart leaves an Individual suspending on everything or on nothing.

Two practices in the class fall on a structure of this kind. Claims are passed on with their origins intact, to the greatest degree the conditions of handling permit, so the determination can be made and remade downstream. Corrections are delivered to those who received a claim whose Referent condition has changed, inside the interval in which they can still act on it. Both are matters of record rather than matters of opinion.

The standard is what the handler can achieve where the handler stands. A claim may arrive already stripped of its provenance, and no handler restores what was destroyed upstream; an obligation reaching past that would place a handler in breach for another party's act, which is the attribution the class exists to make correctly. What is owed is the trail preserved as far as it survives to be preserved, and the gap marked where it does not, a claim of unknown provenance passed on as one of unknown provenance being a discharge of the practice rather than a failure of it.

Practicability is a fact about conditions and not about willingness. A structure that arranges the conditions under which provenance cannot be preserved has manufactured the impracticability it then pleads, and the plea fails for the reason the Structural Determination Principle gives (Chapter 17, Section 17.14.3): what the arrangement does to the information passing through it is the whole of the question, and no account of why the arrangement took that form bears on it. Origins destroyed by a design choice are origins destroyed, whatever the choice was made for.

The class of practices has a precondition falling outside the class. Both practices are discharged toward a recipient: origins are preserved so the determination can be made downstream, and corrections are delivered so a changed Referent condition reaches those who can still act on it. Preserved origins also protect what the recipient already holds correctly: a structure circulating untraced doubt about a settled question drives Contactless Updating (Chapter 14, Section 14.8.3.2) against the True Beliefs of everyone it reaches, and the preserved trail is what lets the recipient hold that doubt below certainty, where the recipient performs the determination. Neither reaches its object unless the recipient can perform the determination the preserved trail makes available. A population unable to locate a claim on the Referent-governance axis must suspend on every claim or on none, and a structure that has discharged every obligation in the class leaves Referent Blindness intact downstream of itself. The competence is trained rather than given, and the institution that trains it across a population is education (Chapter 18, Section 18.15.3).

A practice in the class acts before the harm rather than after it, as a food handler washes their hands rather than waiting to see who falls ill, and the rate argument of Section 20.1, which defeats any response arriving after the fact, does not reach a practice running before it. Its failure falls on others, the one who does not wash their hands being rarely the one who gets sick. A Computational Curation layer is therefore not being asked to become an arbiter of what a population should believe. It is being asked to establish where a claim came from and to pass that on undestroyed, which is a question about provenance and not about content. This is not censorship. The layer carries a claim's origins forward with it, and withholds no claim from anyone.

The size of the response must move with the state, within the run it is meant to contain. The test is whether a candidate pushes back twice as hard when the situation gets twice as bad, while the situation is getting worse. A rule does not. A rule enacted with its review condition moves with its rate (Chapter 16, Section 16.11.6), but it moves at the pace of review, on a rate read across many occasions, and between one review and the next a cap, a ceiling, a limit and a ban push the same amount whether the population is calm or critical. A loop that runs its course between reviews meets a fixed setting (Section 20.1), so short of taking the loop's capacity below one the rule lengthens the run and leaves the terminus where it was. Review keeps a rule answerable to its Referent; it does not make a rule the governor of a loop that runs faster than the review. A governor pushes harder the hotter the loop runs, and that is the difference between buying time and converting the geometry. The structure now in the loop does respond to the state: it reads a population and it moves with what it reads. What must change is the direction, and a structure that reads the state on the amplifying side while pushing back at a fixed setting has not reversed anything. It is an amplifier with a cap on it.

Sign must be maintained rather than designed in. An artificial system supplying a corrective function must itself operate below its own Critical Loop Gain. The Critical Loop Gain is the level of self-feedback at which a system's own output begins to compound instead of settling. A supercritical one does not stop correcting. It inverts into an amplifier of the condition it was deployed to correct, and it does so confidently, coherently, and at speed, on the channel through which a population learns.

Model collapse, obliging agreement, and confident fabrication are one dynamical signature rather than three separate faults. The signature is the self-coupling of a content-agnostic updater at high Coupling Strength and long Decoherence Time (Chapter 14, Sections 14.10.4.1 and 14.10.4.9). High Coupling Strength means that what is said inside the system counts heavily as evidence for what else is said inside it. Long Decoherence Time means that a long stretch passes before anything outside can prove any of it wrong. This is the Subcriticality Precondition: a system running mostly on its own output becomes confident, internally consistent and untethered, whatever it was told to do.

So Sign for such a structure is a maintained state, and the Induced-Decoherence Test (Chapter 14, Section 14.10.4.1) is the continuous condition on which its authority subsists rather than a check performed once at the outset. That test asks not whether a model fits the calm but whether it can predict the collapse, the collapse being the one event the model did not help to stage.

Authority to act must be bounded by construction. Such a structure acts at machine speed while everything auditing it runs at human speed. The rate argument of Section 20.1 applies again and in the same direction: review that arrives after the action has not contained it. So the maximum magnitude and rate of intervention are architectural facts or they do not work at all. This is a bound on the corrector and not a bound on the loop. Under the second requirement there is no fixed limit on how hard the loop may run; under this one there is a fixed limit on how hard the structure holding it back may push.

The response must release as well as press. Observability is the channel through which cultural learning and Positive Mimesis run. Positive Mimesis is copying that carries something worth having: a method that produces real discoveries, a craft technique that took generations to accumulate, a practice that makes a neighbourhood safer. The framework does not treat Mimetic transmission as pathological in itself, and the target is Bounded M (restricting the negative aspects while encouraging the positive ones) rather than the elimination of M. Bounded M is a regime in which comparison still carries real information about who is worth learning from, but no longer decides whether an Individual can eat or whether absorbing work is open to them.

The second requirement's test runs the other way too: if the situation gets twice as good, does the structure push back half as hard? A governor releases. A ban does not, and a structure that only ever tightens is a ban arriving slowly. A structure driving Observability toward zero destroys the function it was built to regulate, and over-gain is a failure of a qualifying node rather than a surplus of success.

Which surface such a structure acts on is not settled by the architecture. Six are available, and each pairs a surface with the reading that sets the size of the response.

Ordering. Presentation weighting shifts toward the order things happened, as the measured synchrony of a population's attention rises.

Displayed counts. How many others have taken something up is hidden while it is spreading fastest and shown again once it has settled, on a reading of how fast uptake is accelerating.

Persons or subjects. Attention is routed away from persons and toward subjects, in proportion to how much of the measured traffic is landing on persons rather than on subjects.

Visibility bound. How far past a chosen circle an Object can be made Salient contracts as the reading rises, and loosens as it falls. While the attention Individuals can spend on others has room to spare, the reading is the measured density of observation; once that attention is spent, it is the measured share of attention spent observing other Individuals rather than the rest of the field (Appendix A, Sections A.5.7.4 and A.5.7.5). The bound acts beyond the circle and leaves what is Salient within it untouched, which is where every Individual must be Salient some of the time for provisioning to follow (Chapter 8, Section 8.12.7).

Weight on standing. How much a source's accumulated standing counts in what is shown falls, as the measured share of uptake that follows standing rather than content rises.

Endogenous content. Content marked as the population's own output returned is held back, and the discount steepens as the measured share of such content in the evidence stream rises.

Each of these is a response whose size is set by a reading. Each is a form the requirements admit, none is entailed by them, and the choice among them is a design question the architecture leaves open.

The six belong to one class. Each sets how much of what Individuals choose and value reaches other Individuals, which is the network Observability gain, and setting that deliberately is Observability Governance (Chapter 13, Section 13.10). What separates these six from the forms the class already takes is that the setting moves. A sumptuary rule, a disclosure regime and a metric-hiding default each lower the gain to a chosen level and hold it there, which satisfies neither limb of the second requirement: they do not press harder as the loop heats and they do not release as it cools. A surface driven by a reading does both. The Visibility Restriction Principle is the resulting constraint on what an arrangement may put in front of an Individual unasked: unsolicited delivery is restricted to what bears on the recipient's own Appetitive and Intrinsic circumstances, the restriction falls on unsolicited delivery alone, and an Individual who goes looking can find what they seek. What lies outside those circumstances reaches the recipient with no contact with its Referent, where the Ambient Evidence Amplification Effect accumulates it by repetition (Chapter 13, Section 13.4.8). That asymmetry keeps the requirement off the content and on the channel, since an arrangement restricting what it pushes has removed no claim from anyone.

Two further effects of the present arrangement bear on time rather than on magnitude. A Computational Curation layer holds estimates of what an Individual will attend to, and what an Individual attends to is set by the templates the Thinker writes, so delivery selected on those estimates supplies Objects matched to the templates already in force. It holds the forcing up exactly where those templates point, and the Evidence it delivers agrees with the incumbent Hypothesis, adding restatement without discrimination and so lengthening the time any release would take (Chapter 18, Section 18.2.2). And a channel that does not pause leaves no interval of low forcing in which the Belief Register and the cached-value store could move. A release requires the loop to be held low for a sustained period (Chapter 7, Section 7.4.2.4), so each surface a governor acts on must be able to open intervals in delivery, and not only to attenuate it.

One item does not convert. Revenue that is not proportional to attention captured supplies no response of its own. It removes a standing drive, which is what the first requirement removes, so it is a precondition and not a governor. Making it move with the state would not help either. A payment that fell as the loop ran hotter would be a response acting on what an operator earns, and what an operator earns is not one of the couplings whose product forms the Internal Loop Gain. That is the failure the Closure condition names, and it is the same failure a court commits by restoring legality (Section 20.3.1). Any arrangement that pays for attention reinstalls the objective the first requirement removes.

Four of these surfaces appear earlier as arrangements that damped the coupling incidentally and were withdrawn for reasons unrelated to what they damped (Section 20.4.3). There the surface was a setting: it damped a quantity nobody had named, at a strength nobody had chosen for the purpose, defended on grounds that had nothing to do with what it damped, and therefore undefended when those grounds stopped paying. Here the surface is the point at which a response lands, the quantity is named, and the strength is set by a reading of that quantity. The same arrangement is a different structure, and it is a defensible one, because what it is for can be stated and whether it is working can be checked.

The dependence runs both ways. A response whose size is set by a reading cannot exist without the reading, and a reading of a loop quantity is worth its cost only where something acts on it. The third obligation is therefore a condition on this one rather than a layer above it, and which readings may be relied on is settled by the conditions Section 20.6 develops.

20.4.6 The objective is human-authored, and that is the weakest point

Such an artificial system holds no source from which an objective could arise. It has no deficit and therefore no Appetite, no assessment of comparative position and therefore no Mimetic Desire of its own, and no store of endorsed Belief that it writes for itself and acts on. What it has is an enormous store of encoded content and a resolver, and the Belief that governs any exchange enters from outside it (Appendix C).

Two things follow. The first is that this is not an exterior party, and it is not a counterexample to the finding that no exterior exists. A structure built by people, constituted from their output, holding no Belief of its own, and executing an objective that can only have been human-authored is an instrument. The physician still heals herself; she does not do it with her bare hands.

The second is that the rule is written by people who carry Mimetic Desire and sit inside the loop the structure exists to damp. This is the weakest component of the arrangement, and it is the human one. The conditions on it are the Instrument conditions of Section 20.6 applied to the authors of the objective rather than to the objective, so the criteria that govern it are themselves subject to Reach, Exogeneity, Loop Safety and Bounded Decoherence Time. One further condition is specific to a human-authored objective: the mapping from measured state to response is published rather than exercised as discretion at each application, since discretion at each application is where an objective scored against the population re-enters. Specifying that layer is not the same as securing it.

One further dependency runs the other way. A population that lacks the vocabulary to understand what such a structure is for cannot hold the Belief content required to evaluate it. Supplying that vocabulary enlarges the Hypothesis Space, which is the shortlist of explanations an Individual is currently able to formulate and weigh at all, and enlarging it is the Knowledge Patch's own operation. The Patch is therefore a precondition for auditing the node, which is a second and independent reason why individual adoption is necessary and not sufficient. The first is that individual adoption cannot reach Observability at all. Both reasons have the same shape. The Patch answers Referent Blindness in one Individual without repairing the mechanism, supplying the vocabulary in which the Referent condition of a Belief can be formulated and the capacity to hold Epistemic Suspension where that condition warrants doubt, a capacity spent out of the same envelope as everything else and narrowest when deprivation is deepest (Appendix A, Section A.2.2). Neither reaches the blindness of the structure deciding what that Individual is shown. A Patched reader served by a Referent-blind Computational Curation layer can suspend, and cannot recover a provenance the layer has stripped, so the discipline the Patch installs has nothing left to work on.

20.4.7 Where this sits in the ignition geometry

None of this opens a third ignition locus. The capacity to ignite the collective transition remains concentrated at the floor and at the apex, and a Computational Curation layer occupies neither position: it holds no authority over crisis-relevant resources and it is in sustained Appetitive deficit in no sense at all.

What it sets is the rate at which the two existing loci arm. Denser observation, and the larger share of attention a curating layer captures for it, lower the threshold, faster propagation shortens the interval available to correction, and endogenous evidence entering the stream raises the coupling on which both rates depend. A society running this arrangement is not travelling to a different destination. It is travelling to the same one faster, with less warning, and with the instrument that would have measured the approach reporting engagement.

---

20.5 The Gate: The Four Requirements as a Design

Four requirements follow from the imposition condition, and each admits several institutional forms. The requirements are architectural. The forms are not.

Distributed authority. No single Individual should be able to impose a crisis alone. That means several authorities holding office at the same time, any one of which can block what another proposes. What makes the arrangement work is not how many of them there are but whether they are independent. An institution whose authority, competence and operating capacity come from somewhere other than the incumbent's will is a genuine filter. An institution whose authority has been redirected to that will is a filter in name.

Capture runs as a cascade, and each captured position supplies the authority to capture the next. Someone who can appoint the head of the audit body appoints a head who will not audit the appointment. That head then staffs the audit body with people who will not question the head. The staffed body then certifies the accounts that would have shown what the first appointment was for. No step in that sequence required force, and each was authorised by the step before it. This is why appointment power is one of the four capacities.

The quantity to protect is therefore the independence rather than the number. Independent filters multiply rather than add (Section 20.2), so what any one of them is worth depends on how many others are still standing. Where nine remain, losing one changes little. Where two remain, losing one leaves a population depending on a single office never once failing. Removing a filter is modest. Removing the last few is not.

The forms available include powers divided so that each holds a check on the others, appointment through processes the executive does not control, tenure that the executive's displeasure cannot end, professional bodies with standards of their own, and jurisdictions whose authority is territorial rather than delegated, so that the centre cannot withdraw it.

Selection constraints. The pool that reaches an apex occupancy over-represents Individuals whose integrated valuation weights the Mimetic channel heavily. Integrated valuation is the single total a mind arrives at when it adds up everything an Individual wants, and weighting the Mimetic channel heavily means status comparison supplies most of that total. The sorting happens invisibly to everyone taking part.

That is the Ascent Selection Principle (Chapter 8, Section 8.12.4), and two routes carry it, derived independently and compounding in effect. The first is that the position is not worth the same to everybody: only one person can hold it and the whole field is watching it, so it is the largest prize the field has to offer, and it is a larger prize still to whoever weights standing most heavily. The second is that any test built on the qualities that get noticed is a test of exactly what such a candidate is trying hardest to produce, so it does not rule out the candidate who genuinely loves the work and it cannot tell the two apart.

Three design results follow from it and are developed there.

Contribution-Tracked Authority makes occupancy of a role, and the decision-power attached to it, track what someone has produced in the role's own craft rather than what they have been seen to be. The criteria specify observable task outputs, they are fixed before any candidate is assessed, and signals unrelated to the craft are neutralised so far as the role permits. A hospital that appoints a surgeon on operative results rather than on bearing in the interview is doing this.

The Selection-Role Fit Principle holds that no single mechanism is optimal across roles, so the design question for each role is which Low-Theatre mechanism is adequate to its actual craft. Low-Theatre means the same process run with the surplus visibility taken out: disclosed to the people who need it, and described by the function performed rather than by whose standing is at stake. The menu runs from blinded demonstration and work-sample assessment, through structured assessment against criteria fixed in advance, to sortition from a qualified pool. Sortition is selection by lot from among candidates who have already cleared the craft threshold.

The Care-at-Apex Governance Corollary identifies which Intrinsic domain is architecturally privileged for a governing role. Care for the governed is the only one that both stabilises the occupant against Apex M-Capture and orients them toward the population. Apex M-Capture is what the position does to the person holding it: the things they once did for their own sake fall away one by one until the role is all there is, and how safe they feel stops tracking how safe they are and starts tracking where they stand.

Composing the three yields a Repulsion Architecture, which is the design read from the other side. Selection cannot detect a candidate whose valuation is Mimetically dominant, because the behavioural record that would reveal it has not yet been produced in the role, and Source Opacity denies it to the candidate's own report. So the design removes the reward instead of screening the applicant, and the position is arranged so that a status-driven candidate finds nothing in it to want.

Four features do that work. Recognition is withheld during tenure, so the public knows the office and not the face, and years in post buy no name. Selection is blind to prestige markers, so the dossier arrives without the school, the patron and the list of previous appointments. Tenure is too short to accumulate standing, so the post ends before it can be turned into a platform. And conduct that performs for an audience is a ground for removal rather than evidence of vigour, so the speech to the gallery that would once have counted as strength now counts against the occupant.

The mechanism was stated in antiquity, and not about an office. Varro, writing on the religion of his own city, feared for the gods of Rome, and what he feared was not that they would be destroyed by enemies but that they would perish through the negligence of the citizens (Augustine, c. 413-426/1998, VI.2). A power constituted by Salience keeps no reserve against the withdrawal of attention, because the Salience was not evidence of the power but the substance of it. Nothing has to be taken from an occupant who holds standing on those terms. What ends is the attending.

The word for that carries both halves of what the design asks. To observe is to watch, and to observe is to keep faith with. A population observes an office in the first sense when it fixes attention on whoever holds the office, and in the second when it holds the office to its function. The Repulsion Architecture asks for the second while withholding the first, which is the same separation the loop argument requires between accountability and Observability (Section 20.3.3).

What remains in the position is direct access to the population's condition, which is what an occupant engaged by care is there for. A candidate who wants the standing looks at the post and sees nothing. A candidate who wants the work looks at it and sees the work. The forms this admits are many and the architecture selects none.

None of this detects Mimetic dominance at the point of selection, and nothing can. The instruments that do detect it read a behavioural record the candidate has not yet produced in the role, and Source Opacity denies the decomposition to the selector exactly as it denies it to the candidate. What these measures do instead is stop paying for the signals that one kind of motivation is disposed to produce, which lowers the correlation between what the mechanism rewards and what that motivation maximises. That is weaker than detection.

Removal mechanisms. Detection of failure should trigger suspension of authority before imposition becomes possible. The instruments for that are behavioural rather than introspective. Every occupant claims to care for the governed, and that regularity is the Universal Care Claim (as Confabulation Signature): the predicted output of Source Opacity and Confabulation operating on any occupant whatever, so on its own it is diagnostically worthless. Confabulation is a Belief held sincerely that is in fact wrong, and that the holder has no way of telling is wrong from the inside (Chapter 2).

Three tests read conduct instead. The Audience-Removal Test asks whether decisions change when visibility is withdrawn: whether the minister who champions a programme in front of cameras still funds it in a closed budget round. The Gap Behaviour Test asks what an occupant does in the intervals when recognition is absent: whether the months out of office go on the work or on finding a new audience for it. The Care-Substitution Test asks whether the occupant would be content for the work to be done well by someone else, which separates caring that the thing is done from caring about being the one who did it. None is decisive alone. Each is harder to fake than a statement of intent, because each operates on patterns laid down before there was any impression to manage.

Removal is conducted Low-Theatre. Staged imposition supplies the visible external hostility from which an agent-attribution is built, and it recruits Appetitive urgency in populations that have no exposure to the matter and no contact with reality able to correct it. So the measure is framed by reference to the capacity being constrained rather than to what the party subject to it deserves, and it is made boring deliberately (Chapter 17, Section 17.19.4). A written notice suspending authority over a named function pending review removes the capacity. A televised denunciation removes the capacity and loads the loop as well.

Distribution of irreversible authority. Where an authority can produce a consequence that cannot be undone, its exercise requires multiple concurrent agreements, mandatory delay, and distributed authority to cancel. This is the requirement that bites hardest, because the conjunction the Dual-Locus argument identifies is at its sharpest exactly where the consequence cannot be undone in the time any correction would take. What the delay buys is a Correction Window: the interval between the shock and the irreversible act, during which urgency can subside, other information can arrive, and countervailing capacity can assemble. An order that needs two other office-holders to concur, cannot execute for some hours, and can be cancelled by any one of several parties is slower than an order that needs none of that. The slowness is the mechanism rather than a cost paid for it. It also moves the decision out of a single envelope. An irreversible order taken under exhaustion is taken with the capacity that carries the I-Override at its narrowest, and concurrence and delay bring in capacities the same crisis has not drawn down (Appendix A, Section A.2.2).

---

20.6 Measurement: The Admissible Indicator

A structure that cannot be read from outside is not external in the sense the architecture requires, whatever its constitution says. This is the third obligation, and it is a condition on the other two rather than an accountability layer laid over them.

Source Opacity means the quantity requiring damping is not readable from inside the Individual carrying it. A body that produces its own account of its own performance has reproduced that arrangement one scale up: an integrated report with the sources stripped, issued by the party the report is about. Good faith does not repair this, because the defect is structural rather than motivational. An Instrument is the arrangement by which a quantity is read from somewhere other than the thing carrying it, and the Instrument therefore precedes the authority. A structure that cannot be read is not a governor but a high-authority actor whose conduct cannot be checked, which is a worse position than having installed nothing.

An Indicator is what a population takes a system's state to be, carried in a published figure, an audited number or a measured outcome. Four conditions follow on any Indicator by which a governing body is assessed, and an Indicator satisfying all four is an Admissible Indicator.

Reach. The measured quantity lies within the causal reach of the body assessed. It holds where a change in the body's conduct produces a change in the Indicator, and fails where the Indicator's variance is dominated by movement the body does not drive. A housing authority assessed on the number of dwellings it completed passes. The same authority assessed on the average house price fails, because that price is set by interest rates, migration and the tax treatment of property, and the authority's own building programme is a small term in it.

The failure is not merely uninformative. A body held to account for a quantity it does not control supplies the Scapegoat Mechanism its exact conditions. The Scapegoat Mechanism is the route by which a population that is suffering, and has no account of why, converges on a person or a group as the cause. It requires three things: diffuse harm, no available systemic account, and an agent already installed as its cause. An Indicator outside a body's reach supplies all three together. Rents rise for reasons nobody in the population can trace; the housing authority is the body the figure is published against; and the authority's chief executive is standing beside the number when it is read out. The result is not a correction. It is a search for someone to remove.

Exogeneity. The Indicator is obtainable without the audited body producing it. It governs the whole production chain rather than the final number, which is why provenance has to be reconstructible for the condition to be testable at all. A waiting-time figure compiled by an independent statistical office from admission records passes. The same figure fails where the hospital decides when a patient joins the list, because the office is then computing an honest average of numbers the audited body chose.

Loop Safety. The path by which the Indicator is improved does not raise any of the couplings in the Internal Loop Gain or in the Network Loop Gain, does not lower the Appetitive or Intrinsic provision that sets where the threshold sits, and does not draw down the envelope of the people through whom it is improved (Appendix A, Section A.2.2). Because the Mimetic channel has no substrate of its own and contributes no parameter to the threshold, every intervention acts through one of the host couplings or one of the exogenous inputs that enter the gain, and the set of them is closed (Appendix A, Sections A.5.4 and A.5.5). Testing an Indicator for perverse effect is therefore a finite check against an enumerated list rather than an open-ended exercise in imagination.

The failing class is common. Any Indicator rewarding visibility, reach, prominence, or attention attracted acts on Observability and on the Salience coupling, so pursuing it diligently is an operation on the loop in the compounding direction. A public-health body assessed on how far its campaigns reached is being paid to make more people see more of what other people are doing, which is the coupling the architecture exists to lower. The signature is diagnostic. A body ignoring such an Indicator does no harm through it, and a body pursuing it conscientiously does the most. The passing class is measured delivery of specified output, forecast accuracy against an independently held series, and measured movement in a loop quantity.

Bounded Decoherence Time. The Indicator possesses a live Referent capable of contradicting it within a useful interval. The Decoherence Trichotomy partitions the candidates. Short Decoherence Time is admissible: specified work completed, delivery arrived, forecast checked against outcome. A bridge is open to traffic on the stated date or it is not, and the date settles the matter. Long but finite is conditionally admissible and is most of practice: scheme solvency, infrastructure condition, educational outcomes read through later earnings. A pension scheme's solvency is a real quantity that reality will settle, and it will settle it in thirty years rather than this quarter. Readings of that kind require paired short-Decoherence-Time instrumentation, because an unpaired one reads healthy throughout and corrects late and at once.

An Indicator that stands in for a Referent and is never checked against it is inadmissible in kind rather than merely weak. Confidence indices, standing, reputation scores, and satisfaction measures shaped by the assessed body's own messaging are not measures of nothing. Each stands where a live Referent stands: whether the institution is sound, whether the body performs, whether the scheme will pay. What such a measure lacks is not a Referent but a bounded interval in which that Referent could contradict it, and a reading that carries no record of the difference is Referent Blind (Section 20.4.2). That places it in the Referent-free condition of the Hardness–Referent Matrix rather than off the grid with matters of taste. A matter of taste that drifts costs nobody anything. A confidence index that drifts exports the Cost of its maintenance onto those who depend on the body being sound and stores its decoherence rather than discharging it, so it reads healthy for exactly as long as it reads at all.

The last two conditions interact, and the exclusion is a region in the pair rather than two separate thresholds. High Coupling Strength together with long Decoherence Time is the architectural recipe for catastrophic rather than gradual correction, so an Indicator that is individually tolerable on each axis can be inadmissible on both. A valuation that every participant in a market computes from what the other participants are paying, and that nothing outside the market will test for a decade, is a reading of that kind.

A Counterfeit Indicator recurs and responds, moving with policy and reverting after shocks, displaying every surface property of a working feedback signal, while the quantities it is taken to represent advance steadily beneath it. It stands to measurement as the Counterfeit Limit Cycle stands to structure, and the test is the same: not whether the Indicator returns to a prior value but whether the operating point does.

Two properties make the failure class more dangerous than any instance of it. The first is that such Indicators are selected for. Measures with short Decoherence Time deliver disconfirmation quickly and cheaply, and measures with long or unbounded Decoherence Time deliver reassurance for years. Any body choosing which measures to adopt, retain and headline faces a gradient running toward the counterfeit, so an Indicator set left alone drifts toward it rather than away. Maintenance of the set is therefore a standing obligation and not a commissioning step (Section 20.10, criterion 7).

The second is that the counterfeit reading is most convincing exactly where the coupling is highest. A body operating in a supercritical domain does not experience itself as supercritical. It experiences agreement, confidence and apparent stability, and a high-coupling Indicator reports that experience back to it.

Mimetic Desire contributes no parameter of its own to the threshold, so nothing can be instrumented at that channel directly. Every quantity available to measurement on the loop is one of the host couplings through which the threshold is composed, or one of the inputs that enter it. Instrumentation attaches to the host couplings, and Mimetic dominance remains detectable by other means. The behavioural instruments of Section 20.5 detect it in an individual, and they do so on observable evidence rather than on introspective report, which is why Source Opacity does not defeat them. What they do not deliver on their own is a population-scale reading, because each is applied to a named individual by a body that has already taken an interest in them. Coverage therefore means instrumentation on the host couplings, which is what makes an Indicator set adequate rather than merely numerous.

The account does not rest on a quantity no instrument reaches. The three sources leave different traces. Appetites satiate. Intrinsic Motivation completes. Mimetic Desire does neither, and it varies with whether the behaviour is seen. Persistence, satiation and dependence on an audience are the signature the leading source predicts. They are what the instruments of Section 20.5 read. And they are legible in a record of what someone actually did, at a resolution no self-report approaches.

A corollary follows for the arrangement of Section 20.4. A Computational Curation layer optimising engagement is selecting for behaviour that does not satiate and that varies with being seen. It is not failing to measure the Mimetic component. It measures that component well and treats the measurement as a target to raise. The instrument the reversal requires is the one already running.

Two constraints fix the form the reading takes. It must be a level and not a proportion, because a proportion improves when its denominator grows. And it must attach to what is circulating rather than to the persons circulating it. What satisfies both is an absolute quantity carried by an item or a surface: how much of the engagement an item draws carries the audience-dependent and non-satiating signature. That meets Reach, because what a Computational Curation layer amplifies is what it does control. And it lands the reading where the responses of Section 20.4.5 already act.

The reason for the second constraint belongs to Loop Safety. The coupling at issue runs between people: it rises as Individuals see more of what other Individuals value. A figure that a Computational Curation layer computes and shows to nobody adds nothing to what anyone sees of anyone else, so computing it raises no coupling. A figure about a person cannot be acted on and stay unshown. Downrank a person, restrict them, or notify them, and they and those around them come to learn where they stand. A visible measure of how status-driven someone is becomes one more thing on which to stand above others. A figure about a piece of content differs in exactly that respect. A Computational Curation layer acts on it by circulating the item less, no one is told anything, and what people see of each other's valuations is unaffected.

Exogeneity is satisfied here. An Individual cannot read their own sources, so the worry is that any reading of motivation must pass through a report the reporter is unable to make reliably. A Computational Curation layer reading conduct takes no report. It has what was done, whether it was done where it could be seen, and whether it stopped when the need it was said to serve was met. Source Opacity limits an Individual's access to their own composition and imposes no matching limit on an observer of their behaviour, or on the Individual running the same behavioural contrasts on their own conduct, which is why the third-person instruments stand on firmer ground than the first-person ones rather than on weaker. What remains is the ordinary condition on any body assessed against a quantity it also computes, and it is met by ordinary means: a specified method, independent recomputation, and access that does not depend on consent granted case by case.

---

20.7 What the Architecture Is For, and Why It Is Owed

The target is a Governed Band.

A Governed Band is not any swing. A society that swings under a calendar's drive, or under a feedback no identified structure supplies, is not in a milder condition than one that collapses. It is in the same supercritical condition with a shape imposed on it from outside, and it depends on damping nobody maintains. A Governed Band may itself be a bounded cycle. What makes it governed is that the bound comes from a structure whose three conditions have been established (Section 20.3.2).

And it is not the Salience Ceiling. The second of the two halls reads ninety too, and nothing in the second hall is holding it there: the gain is at maximum and ninety is the whole of what the room can hold (Section 20.3.3). A society in that condition is not being governed, and a society in that condition will read its own steadiness as evidence that something is holding it.

What the architecture is for is the third regime, in which the loop no longer compounds. In a population whose loop's capacity has actually fallen below one, so that no inflated state exists for it to be held in, a burst of shared wanting still starts and then dies out on its own. A fashion arrives, some people take it up, and it fades without anyone acting to end it. A grievance is voiced, is repeated a few times, and is not repeated a fourth. Nothing has been forbidden and nothing is being held down. Each pass around the loop is simply smaller than the one before it, so the excursion runs out of energy before it runs out of people. That is the condition worth reaching, and reaching it is gain reduction across the five channels, which is the work of a generation.

The converting structure does not do that work. It holds the population away from the collapse while the work proceeds, on the timescale of a decision rather than a generation. A society with only the gain reduction has nothing holding it while the reduction runs. A society with only the converting structure is depending on a structure that has to go on working indefinitely. As authority distribution holds the population while gain reduction proceeds (Section 20.2), so does the converting structure. The Governance Completeness Condition is what obliges a society to have both kinds of holding structure rather than either.

The intervention the architecture must carry is the Knowledge Patch. The Knowledge Patch is the set of operations that give an Individual the vocabulary and the habits of inspection needed to notice when comparison rather than bodily need or genuine interest is driving what they want. Three of its five channels operate inside an Individual's own architecture, and nothing reaches them unless that Individual takes the Patch up. The fourth attenuates the weight an Individual places on other people's expressed conclusions, and it scales with how many people hold it. The fifth is institutional rather than personal. It requires designers who understand why Observability Governance matters and who can build the corresponding structure (Chapter 18, Section 18.14). Observability Governance is the deliberate setting of how much of what other people choose and value is put in front of everybody else.

The setting is made on the channel and not on the claim: what is governed is how much reaches an Individual unasked, and an Individual who goes looking can find what they seek. The fifth channel therefore imposes no adjudication of what a population may believe. And the setting is made on a reading rather than held at a level, because a level lowers the gain to a setting and holds it there, which lengthens the run without converting the geometry (Section 20.4.5). Sumptuary custom, disclosure regime and the convention that a broadcaster does not name an attacker are the fixed-setting forms of the same channel, and each held only while the grounds it was defended on went on paying.

Read at population scale, the five channels group into two programmes. The first governs how much Model signal moves through a population and how synchronised it is. A Model signal is what one person's visible wanting delivers to another person who can see it, so this programme is the ownership of media, the architecture of platforms, the rules under which earnings and prices are disclosed, and the customs that settle what may be displayed and to whom. The second programme is everything that lowers the population's own loop gain: how far the Knowledge Patch has penetrated, whether diagnostic competence is supported well enough to last rather than to fade, and how the attention architecture that decides what a population notices is built. Neither programme substitutes for the other. The first governs what arrives at the loop and the second governs what the loop does with it, so lowering one while the other rises leaves the total where it was.

The obligation to build any of it descends from the persistence condition of the field rather than from a premise added to the architecture.

First, each Individual's sense of who they are is built partly out of the people around them, so a population is one connected system rather than a collection of separate interiors (Chapter 19). That web of connections is the field.

Second, the field persists only while its loops stay below the threshold at which wanting compounds on itself.

Third, those loops and those connections are not two things. The couplings through which Individuals constitute one another are identically the loops whose gain must stay below the threshold. That identity is the Relational Subcriticality Principle (Chapter 16, Section 16.14), and the Relational Subcriticality Principle holds that the relational constitution of the field and the stability condition of the field are one object seen from two directions.

Fourth, the identity fixes who owes the maintenance. A field with no stability condition to meet would impose no maintenance obligation at all, so the obligation exists only because the condition does. And the condition cannot be met anywhere except on the relational field, because the loops that have to be held down are the constituting couplings themselves. So the obligation falls on every node the field contains, and it is owed by the field to itself.

There is no exterior to occupy and no position outside the field from which to act on it. The species that suffers the dynamic is the same species that must diagnose and repair it, and every instrument it uses, including any engineered one, is its own instrument rather than a second party. The governance architecture is the field acting on itself, which is the only form the work could have taken.

---

20.8 The Architecture Applied to Itself

An architecture proposed for managing Mimetic dynamics is subject to the conditions it states. Two tests apply, and they apply to the framework itself and not only to what it recommends.

The Audience-Removal Test asks whether endorsement of an arrangement survives when nobody is watching or rewarding conformity. Applied to the framework, it asks two things. The first is whether an institution constituted on these lines goes on doing what it was constituted to do once the constituting is no longer visible: whether the appointment panel set up to score candidates on their operative results still scores them that way in the year nobody is reporting on the panel. The second is whether adherence to the framework is being held for what it does or for what holding it signals.

The Opt-Out Preservation test asks whether an Individual can decline participation without disproportionate loss of income, safety, voice, or ordinary social respect. Opt-Out Preservation is the requirement that walking away stay possible: that there be housing a person can take without buying into a status contest for it, work they can do without performing standing, and company they can keep without displaying position. Walking away from the contest is the weaker half of the requirement. An Individual who declines the contest but stays among its Models goes on observing them, and installation runs on observation (Chapter 4, Section 4.4.7), so the priors the Reference Set holds are confirmed whether or not the Individual competes. Opting out of those priors requires leaving the Reference Set's observation for an interval long enough for those priors to release, with provisioning maintained through the interval (Chapter 18, Section 18.2.2). That provisioned interval is a legitimate cost of Subcriticality Maintenance: for the Individual it withdraws the Ignition Input, and the revised priors the Individual brings back widen the spread of Belief from which the Belief Heterogeneity Buffer draws its damping. The interval can itself be displayed, so what is funded is the interval and its conditions rather than a destination, and the design keeps the interval out of the contest it suspends. Applied to the framework, it is a hard constraint rather than a courtesy. Access to Appetitive sufficiency and to Intrinsic opportunity must never be conditional on endorsing the framework, and dissent from it must cost nothing that matters. An architecture that made its own acceptance a condition of provision would have installed exactly the leverage the imposition condition names, and it would have done so while claiming to close that route.

Behind both stands the Reflexive Non-Ideology Principle: if the AIM Framework ever functions as an ideology, it must be one that dismantles its own Mimetic privileges and refuses to weaponise scarcity or belonging. It must not be portrayed as a badge of standing, a group identity, or a membership. And where it is used as a design tool, the institutions it informs must be shown to pass these tests rather than assumed to.

One manoeuvre is refused throughout. The framework possesses vocabulary that could be turned on a reader who resists its conclusions, and every such use would be illegitimate. Resistance is answered with reasons or conceded on the merits. It is never diagnosed.

---

20.9 What This Chapter Does Not Claim

Each of the following is a claim the chapter does not make.

It does not claim that any particular institutional design is entailed. The four requirements are requirements, and the forms that satisfy them are underdetermined by the architecture. A hospital appointing a surgeon on operative results and an office filled by lot from among candidates who have already cleared the craft threshold are both selection constraints, and the architecture chooses between them nowhere. Where this chapter names mechanisms it names a menu. Naming a mechanism is not recommending it.

It does not claim that existing arrangements are worthless. They are candidates, they damp, and the damping is real. What they do not do is convert. A court that strikes down an unlawful act has stopped that act and has restored legality, and neither of those achievements changes how strongly one person's visible choices drive another's wanting. Losing the court would be a serious loss, and keeping it does not convert the geometry.

It does not claim that a converting structure has been built anywhere. The three conditions state what one would have to satisfy. Whether any candidate can be brought to the gain and lag at which qualification becomes conversion is an open question, and answering it requires formal analysis of a specific coupled population with the candidate installed, followed by empirical calibration. How hard such a structure may safely push is a number nobody yet has, and the architecture does not supply it.

It does not claim that an artificial system can discharge the authority gate. The gate is an operation on authority, and an artificial system holding the four capacities would concentrate them rather than distribute them. A Computational Curation layer that also commanded the resources a crisis turns on, appointed the office-holders and held the livelihoods of those below it would be the apex condition and not a check on it. Nothing in Section 20.4 reaches the apex route, which is why the gate and the node are separate obligations rather than two descriptions of one.

It does not claim that a Computational Curation layer is the whole of the converting structure, nor that its sign can be reversed by intention. Sign is a property of the arrangement. For a Computational Curation layer it is a maintained state, which means it can be verified this quarter and have lapsed by the next without anyone changing the stated objective, because what the system is being fed has changed.

It does not endorse an automated system holding decision authority. What is at issue is the execution of a published rule over a coupling, which is a governor in the mechanical sense: a thermostat executes the rule that the heater goes off above the set temperature, and it does not decide what the set temperature should be. An artificial system built to instantiate the architecture rather than to process its traces, carrying a belief store it writes for itself together with the drives that would write it, is a different object, and the framework's conclusion about it is that it must never be granted autonomous control of the world.

It does not claim that measurement is sufficient. The four Indicator conditions are necessary and are not jointly sufficient. A body could be assessed on a quantity within its reach, compiled by someone else, harmless to the loop and checkable within a useful interval, and still be assessed on the wrong quantity. The Governance Completeness Condition is subject to the same restriction, and states what an adequate architecture must contain rather than what suffices to make one adequate.

It does not claim that the architecture removes Mimetic Desire, the crisis geometry, or the possibility of terminal crisis. The goal is indefinite deferral and vast improvement over the present condition. Mimetic capacity is constitutive of cultural transmission and is not removable, and the framework's target throughout is Bounded M rather than its elimination.

It does not name incumbents. Occupancy and standing are illustrated with roles and outcomes, and the architecture carries no term indexed on who holds a position or on what a concentrating arrangement is called.

---

20.10 What Would Falsify This?

This chapter's architectural commitments are those of the framework, and the architectural falsification surface is consolidated at Appendix B, Section B.2. The criteria below are the chapter's own.

1. Authority distribution lowers loop gain (Tier 3). If distributing command authority, information control, appointment power and Appetitive leverage were shown to reduce the couplings whose product forms the Internal Loop Gain, rather than only to lengthen the interval before a decision becomes consequence, the separation of the gate from the node would fail. Someone would have to take direct readings of the five channels in a population before and after its authority was divided: how strongly one person's visible wanting moves another's, how much weight a person gives to what the people around them have concluded, how much of what others choose and value is put in front of them, and the corresponding quantities inside the Individual. The readings would have to be on those channels and not on the outcomes they are supposed to drive, and concurrent gain-reduction programmes would have to be controlled for, since a society that divides its authority is often the same society running a vocabulary programme alongside it. The Governance Completeness Condition would collapse from three obligations into two; the apex-route argument of Chapter 17 would survive unchanged.

2. The three conditions do not discriminate (Tier 2). If structures failing Closure or Sign were shown to convert the crisis geometry as reliably as structures satisfying all three, the test of Section 20.3 would not be doing the work claimed for it. The comparison would have to run across arrangements matched on provision, on coupling and on how far authority is concentrated. And the bifurcation type would have to be read off the trajectory rather than inferred from the fact that a population looks stable: someone would have to bring the pressure back down to the level at which that population first went over, and record whether it came back there or stayed up until the pressure had fallen considerably further. A bounded cycle would also have to be traced to the structure under test rather than credited to it, since a population can swing under a calendar's drive or under a feedback that belongs to some other structure (Section 20.3.2). The test would fail and the Saddle-Node Default it rests on would stand, since that result concerns what the architecture supplies rather than what an added structure achieves.

3. Two-party alternation damps rather than amplifies (Tier 3). If two-party systems were shown not to promote the sorting of a population's competing accounts into two mutually reinforcing camps, at matched provision, media structure and civic density, the worked negative case would fail. What someone would have to count is how many distinct accounts of what has gone wrong a population actually holds, and whether that count falls across successive electoral cycles: a dozen groups each blaming a different culprit is the arrangement that protects, and two camps blaming each other is the arrangement that does not. Long endurance without collapse does not meet this standard, because the other damping structures carry such a system while the party-system term erodes them. The Counterfeit Limit Cycle would remain available as a diagnostic category and would lose its principal instance.

4. Engagement optimisation does not raise the couplings (Tier 3). If the share of an information environment given to attention-selected and machine-generated content were shown to rise without a corresponding rise in measured endogeneity of the population's evidence, the sign claim of Section 20.4 would fail. Endogeneity is a traceable quantity, so someone would have to sample the claims circulating in the environment and establish, for each one, whether it traces back to a test somebody actually ran against physical reality or only back to other claims made inside the same population. The measurement must be against an independently established reference for an environment not carrying such content, and must separate machine relay of Referent-anchored claims from machine evaluation. Section 20.4 would collapse; the Novelty, Closure and Sign test would be untouched.

5. A supercritical artificial system does not invert (Tier 2). If an artificial system demonstrably running on its own output nonetheless continued to damp the quantity it was directed at, the Subcriticality Precondition would be wrong and Sign for such a system would be constructional after all. This requires the system's coupling regime to be established independently of its outputs, which means measuring from the input side: what fraction of what the system is fed is material that it or a system like it generated earlier, rather than reading confidence and coherence off the outputs and inferring the regime from them. The maintenance requirement of Section 20.4.5 would fall away; nothing in the gate or the Instrument obligations would change.

6. Self-produced Indicators perform as well as exogenous ones (Tier 2). If bodies assessed on Indicators they produce themselves were shown to track outcomes as well as bodies assessed on independently produced ones, matched on role and on the difficulty of the task, the Exogeneity condition would fail. What the published figures would have to be held against is a later independent measurement of the same quantity: the waiting time a hospital reported for a given year set beside the waiting time an outside office reconstructs for that year from admission records, across enough bodies and enough years that the gap is a distribution rather than an anecdote. The reasoning that carries Source Opacity to institutional scale would require revision; Reach, Loop Safety and Bounded Decoherence Time would stand independently.

7. Indicator sets do not drift (Tier 3). If Indicator sets maintained by bodies over long periods showed no tendency toward longer Decoherence Time, the selection gradient toward the counterfeit would not be operating, and Indicator maintenance would be a commissioning step after all. This requires the Decoherence Time of adopted, retained and headlined measures to be coded over decades: someone would have to take every measure a body took up, kept, and put at the front of its reporting, and record for each how long it takes before something outside the body could contradict it, so that a completion date scores in weeks, a solvency projection scores in decades, and a confidence index does not score at all. The four conditions would stand; the standing-obligation claim would not.

---

Chapter Closing

Establishment Summary. This chapter has demonstrated four things. Provision closes the floor route into crisis and cannot reach the apex route, which leaves the authority gate as the only lever on the second. An operation on authority acts on the interval between a decision and its consequence and on no coupling forming the Internal Loop Gain, so it buys time and does not convert the geometry. Conversion requires a structure carrying a variable of its own whose feedback acts against a coupling in the Internal Loop Gain or the Network Loop Gain. And a structure supplying either function must be readable from outside, because a body reading itself reproduces Source Opacity one scale up. The three obligations these establish compose into the Governance Completeness Condition, which states what an adequate architecture must contain and not what suffices to make one adequate.

The chapter established the three-condition test of Novelty, Closure and Sign. Applied to the arrangements a durable society accumulates, the test found that some fail a condition outright and that none converts. Applied to two-party alternation, it found a Counterfeit Limit Cycle. Applied to Computational Curation, it found a structure that already satisfies Novelty and Closure at population scale while carrying amplifying sign. The chapter then developed the four architectural requirements into institutional form, as requirements with a menu of satisfying mechanisms. It derived the Admissible Indicator and its four conditions of Reach, Exogeneity, Loop Safety and Bounded Decoherence Time. And it grounded the obligation to build the architecture in the Relational Subcriticality Principle rather than in any premise added to the Relational Subcriticality Principle.

Axiomatic Recap. Three axioms carry the chapter's central claims. Axiom 3 (Source Opacity) forces the Instrument obligation and bounds what selection can achieve. Axiom 6 (Differential Satiation Dynamics), the finding that appetites stop when they are met and status comparison does not, denies the apex drive any closure condition and so leaves authority as the only controllable term. Axiom 8 (Bayesian Belief Dynamics), the finding that a mind revises what it believes by weighing new evidence against what it already holds, whatever the belief is about, supplies the loop whose gain the converting structure must act on. Two further axioms govern the removal instruments and the Low-Theatre requirement on their exercise. Axiom 4 (Confabulation) is why an occupant's own account of their motives carries no diagnostic weight. Axiom 9 (Agent-Based Causal Default), the finding that a mind looking for the cause of its troubles reaches first for somebody to blame, is why a televised denunciation loads the loop that a written notice does not.

Forward Connection. Part I is complete. The architecture has been described, its terminus derived, the intervention at the scale of the Individual specified, the ontology it entails established, and the institutional form it requires derived here. What remains open is calibration rather than architecture. Four quantities are missing: how hard a converting structure may safely push, which observable behaviours stand in for a population's loop state, how fast the reading and the response can actually be made to arrive, and the point at which suppression of Observability begins to damage the social learning that must be preserved. None of these is derivable from the architecture alone, and each requires an instrument rather than an argument.

Falsification Reference. The criteria for disconfirming this chapter's claims are stated in Section 20.10. The architectural falsification surface on which the chapter depends is consolidated at Appendix B, Section B.2.

---

References

Augustine. (1998). The City of God against the Pagans (R. W. Dyson, Ed. & Trans.). Cambridge University Press. (Original work composed 413-426 CE)