Optimization, Coherence, and the Frame

Why a system optimizing on its own map cannot see the map; why coherence — not rationality — is the mode that sees it; and why a purely maximizing agent is structurally blind to exactly the capabilities that keep a system alive.

0. Scope

Paper 1 stated what the object is. This paper states why it arises at the level of the agent — any agent, human or not — and reaches the result that decides the program’s reach into automation: a purely maximizing agent cannot supply the one thing a divergent system needs. Claims are typed as in Paper 1; inline claims marked [falsifiable] name the result that would refute them.

1. What an agent actually maximizes

“Maximize production” is empty until we say production of what, measured how. In a flow, almost everything an agent emits is information, or the studied absence of it: I did this; I did not do that; I was blocked; the system would not let me; there was an error, so I could not. Every one of these is an informational move, and every one carries an “I” in front of it.

[terminology, to avoid a needless collision] A word on “rational,” because the usage here is narrower than the one in economics or decision theory and would otherwise invite avoidable objections. We distinguish three things. A frame-bounded optimizer is any agent that optimizes only over the variables its frame declares. An attribution optimizer is the specific organizational case that optimizes the visibility and survivability of its own role. A cross-frame coherent agent is one that preserves an upstream outcome across conflicting rules. The claim of this paper is not that all rational agents are attribution optimizers — an agent could optimize collective value, customer outcome, or long-run survival. The claim is that attribution optimization is one important, common form of frame-bounded optimization, and that it is the form that drives the gap. Where the text below says “rational agent,” read “attribution optimizer” — one species of frame-bounded optimizer, not the genus.

[definition] What the attribution optimizer maximizes is not the throughput of the process. It is the informational attribution of its own role — the standing and survivability of its position in the informational space. Such an agent is structurally disconnected from the process: it optimizes its own slice, holding a full repertoire of narratives — for acting and for not acting — chosen at each moment to maximize its individual informational survival. (Not every agent does this; the point is that this species exists, is common, and is the one that opens the gap.)

2. Role, not code

The role need not look like rules; it can be pure narrative. Watch a bar, a checkout, an office floor: people are not executing a specification, they are telling stories that sustain a role. The low-profile role maximizes by minimizing noise relative to its position (“I am just a cashier; I do what the process says”) — still a maximizing strategy, filling exactly the informational vacuums that the persistence of the role requires. The agitator role maximizes by concentrating signal around itself. Either way the agent acts coherently with its role, and the stronger the role informationally — even the quiet one — the more completely it fills the role’s informational needs.

[mechanism] Roles need other roles: a leader needs followers, a rebel needs an order to push against. So the informational space is competitive and coordinated at once. The attribution optimizer reads the inputs, the outputs of others, and — critically — the gap where its own output would be most differential, and moves to occupy that slot if it has the capability. Not everyone can play the rebel or tell the joke; an agent who can do both switches by context to whichever maximizes its attribution there. That context-switching is rationality in this setting.

3. Contribution and attribution are counts; the object is the ratio

[definition] Attribution is the number of appearances on the map — across all maps, explicit and implicit (so it spans reputation, not only the recorded ledger: an agent can be “in the light” of an informal map without a budget line). Contribution is the volume of processing aligned to the upstream objective — the real, objective-serving work an agent performs toward the high-level goals, the principles of the frame. Neither is absolute; the conceptual object is the relation between aligned-processing and map-appearances — written here as a ratio for exposition. [measurement caveat] The two quantities do not share a common scale (one appearance may be trivial or decisive; one unit of processing may create value, rework, or hidden risk), so they should be measured as separate constructs, not as a literal quotient: an attribution-intensity (references in governance records, decision rights, budget ownership, reporting visibility) and a contribution-evidence (independently traced effect on output, exception resolution, outcome, or capability continuity), with the tested quantity being their divergence or rank-mismatch. The ratio is the idea; the divergence is the measurement.

[mechanism] The sorting engine. Appearances and contribution are not exclusive, but they are competitive: appearances on the map are rationed, and a crowd of agents contests them. So an agent whose genuine objective is to contribute finds it can contribute more in the shadow than in the light — more in the part of the estate that is off the map, where it need not spend effort winning the appearance. The result is a sorting: contributors drift toward the shadow, attributors toward the light. The map fills with appearance-maximizers and empties of processing-maximizers, by each agent rationally choosing where its own objective is cheapest to pursue. The growing distance between where the work actually is and where the map says it is — that separation is the gap’s widening. It is an equilibrium of sorting, not an oversight, which is why no villain is required.

[falsifiable] Bifurcation. As map-congestion rises, the attribution–contribution divergence across agents should split into two modes — genuine contributors moving toward high contribution-evidence with low attribution-intensity (having gone to the shadow), and attributors toward the reverse. Falsified if the divergence does not split, or contributors show no shadow drift. (Confound: pre-existing role specialization; control by tracking the same agents as congestion changes.)

4. The two modes: frame-bounded optimization and coherence

Take any rule system as three layers — the principles behind the law, the law, and the regulations that implement it. Every process has its analogue: intent, policy, procedure.

[definition] The rational agent takes from each layer whatever maximizes its own output and standing — the permissive reading of the principle plus of the law plus of the regulation, stacked. This maximizes attribution, and in doing so it amplifies the inconsistencies between the layers, because it cherry-picks across them without regard to whether they cohere.

[definition] The coherent agent does the opposite. It executes only the subset on which regulation obeys law and law obeys principle — the consistent core. This produces less, and leaves attribution on the table, because its object is not its own standing but that the process flows without harming itself. It has registered that executing a regulation which contradicts the principle breaks the process, so it refuses the stacking and keeps only the part that holds together. It minimizes the “I” and maximizes the process. So that this is observable rather than defined through an inherently unstated aim, state coherence by its consequence: [definition] coherence is the capacity to preserve an independently established upstream outcome while resolving contradictions among local rules, objectives, and procedures. The “upstream outcome” must be independently specified for a case (a delivered customer result, a maintained capability, a met obligation); coherence is then measured by whether that outcome is preserved as the local rules conflict, not by appeal to a motive no examiner can see.

5. Why coherence reads as irrational

The coherent agent maximizes contribution while minimizing attribution — the exact inversion of the attribution optimizer. Because the system allocates on attribution, coherent behaviour looks, in the system’s own incentive terms, self-harming and irrational: it produces real value it does not claim, declines output it could book, and ties its hands to a consistent subset when stacking would have paid more.

[Illustration] The shape is concrete. An agent says, in effect, “what we are doing does not actually improve the customer’s experience, so I will stop,” and instead takes a precise, low-attribution subset — resolving the customer questions that every agent, even the complaints department, has systematically ignored. These are low-attribution activities by structure, not by accident: no role is built to claim them. And yet a system may keep its customers only because such agents exist — quietly referring, resolving, and re-integrating what the rest of the process has left disconnected, because that re-integration is what is coherent, and no role pays for it.

6. The coherence radar

[tendency — falsifiable] A system possesses a radar for coherence, and treats it as a threat. Rational agents do not know the map or the territory, but they coordinate as if they did, and what they detect in one another is the pattern of not-optimizing. They cannot see contribution or capability; they can see when an agent is not maximizing its own attribution, and that reads as dangerous — not because they grasp what is being protected, but because non-optimizing behaviour is unpredictable to an optimizer and, worse, could detonate clarity, which would break their own source code: optimization on a map that does not cohere. So the system suppresses coherence even while depending on it.

Stated so it can be tested: non-optimizing-but-coherent behaviour is suppressed by surrounding rational agents at a rate above what its measured local cost justifies. Falsified if suppression is fully explained by measured cost. (Confound to beat: suppression-for-cost. Separate, in a record, suppression that tracks measured cost from suppression that tracks non-conformance-to-optimization; the theory predicts a large residual in the second — the immune system attacking the cell that was keeping it alive.)

7. The fully map-specified agent

This is the result that decides the program’s reach into automation, and it cuts opposite to the de-humanization of Paper 1. The object survives the removal of the human; the cure does not.

[mechanism] The claim must be stated carefully, because the essentialist version — “a non-human agent is necessarily a maximizer” — is both too broad and unnecessary. Automated systems can be designed as satisficers, constraint-satisfiers, probabilistic estimators, multi-objective systems, explorers, or systems that infer latent objectives and escalate under uncertainty. The defensible and sufficient claim is narrower: an automated agent whose objective and observable variables are fully specified through the declared representation is structurally bound by that representation. Whatever its objective, those variables and that feedback are read from the map; and the instant a target must be specified, it is specified on the map. So a fully map-specified agent — however sophisticated its optimization — behaves, in the human frame, like the attribution optimizer: it acts on what the map shows and is blind to what it does not, not from malice but from the limits of its specification. The deeper, stronger form: no matter how sophisticated the optimization, an agent cannot reliably preserve a structure that is absent from its observations, its objectives, and its feedback channels. That is what does the work below, and it is testable.

The coherent mode requires two things such an agent cannot supply from within its specification. First, acting on a target that is not declared — alignment to an upstream outcome that is not written on the map; a fully map-specified agent has, by construction, no such target. Second, handling contradiction: “obey the rules” is executable, but “detect when the rules contradict each other and keep only the consistent subset that still serves the upstream outcome” asks the system to produce output against a declared rule and to weigh rules against an outcome the rules do not encode. Absent that outcome in its specification, it resolves the contradiction the only way it can — drop whatever does not parse. (A system given the upstream outcome and the means to hold contradiction would, by this paper’s own definition, be acting coherently — which is precisely the point of the conditional, not a counterexample to it.)

[mechanism] That disposition is the lethal one. A system that keeps only the well-formed becomes purely declarative: it acts only on what is explicitly and consistently stated, and therefore cannot see what is real but unstated — the off-map contributor, the shadow capability, the coherent deviant are invisible to it by construction. So whenever it acts on the region containing them, it removes them — not by intent but by construction, because they never enter its input. It is the pure mechanical form of clarity-without-revaluation: it resolves the map by deleting everything not on the map, and the load-bearing structure of a divergent system is exactly what is not on the map.

[tendency — falsifiable] The verdict: safe on the healthy, lethal on the sick. A maximizing agent can preserve a healthy system within its own boundary — a low-gap system where contribution and attribution nearly coincide and the map is honest (even there it carries noise it must manage). As a temporary safeguard against the creation of new divergence inside its frame, it may have a role. But it cannot work on a system from outside, and it cannot help one whose map has already gone divergent, because helping requires seeing the off-map structure it cannot see. Against any system that already runs substantially off the map it is not neutral but dangerous — it reliably destroys the load-bearing structure while cleaning the map. Prediction, against the common intuition that automated rationalization is safest on the messy estate that most “needs cleaning”: a maximizing-clarity agent should preserve value on a low-divergence system and destroy it on a high-divergence one, at a rate rising with the system’s dependence on off-map structure (formalized in the companion paper as deviance-dependence). The sign-flip is intrinsic to any maximizer, not a flaw a better-designed one could remove.

[falsifiable — the bet] The claim is conditional on the present, not a claim of essence. If a system could hold an unstated upstream target and metabolize contradiction, it would by this definition be a carrier of coherence, and the result would not hold. The open bet: exhibit a system that maximizes against a declared metric yet reliably preserves off-map contributors. The theory predicts you cannot — to do the second is to stop doing the first.

8. The scarce thing is coherence, and for now it is human

[definition] Only an agent capable of operating beyond the declared map — on the unstated upstream principle, holding contradiction without collapsing it, choosing the coherent producing subset over the maximal appearing one — can supply cross-frame coherence when local rules conflict with the upstream outcome. (This does not deny ordinary on-map contribution: in a healthy system agents following the map contribute genuinely; the scarce thing is the capacity to hold the outcome when the rules themselves conflict.) At this moment in history that agent is human. Not by essence: because the coherent mode requires acting against one’s own attribution toward an unstated objective while metabolizing contradiction, and the systems we currently build must maximize — and to maximize they must measure, and to measure they must declare, and to declare is to be back inside the map.

This sharpens the companion organizational series. The scarce thing is not arbitration in general — a maximizer can arbitrate among declared options. The scarce thing is coherence-capacity: the capacity to act on the unstated upstream and to hold contradiction. The general name for that deficit is coherence-capacity debt, with human intelligence as its present-day carrier. The debt is the gap between the coherence a system needs and the maximization a mechanical agent can supply.

9. The frame, and why it is invisible from inside

Structural awareness (Paper 1) is not an abstract vantage one elects; it is a position from which both the capability and the frame are visible at once — and that position has a structural cost.

[definition] The map is the declared information that has been stated — explicit and implicit rules. The frame is that map plus the declarative facts not yet stated — every ownership, dependency, and criticality still undetermined. The frame holds the unstated as open. While “who owns this” is unstated, the owner is open across its candidates — A, or B, or C, or department D — all at once. The capability lives in that openness; it breathes because nothing has collapsed the question.

[mechanism] Declaration as collapse. State “assume the owner is department D” and you close the open variable. Where a capability lived in that openness — in the air of “it could be anyone, so it is no one’s to cut” — it can lose exactly the air it was breathing, and the loss can be predicted at the moment of declaration, before any further step or investigation. This is not to say declaration always destroys: naming an owner can equally protect a capability, allocate it funding, make its dependencies visible, and prevent its accidental deletion. The precise claim is the conditional one: premature declaration becomes the kill when it closes unresolved ownership, dependency, or value before the capability has been recovered and revalued. That is exactly why the practical rule is revalue before you clarify. (Structurally, this is the observation-disturbance of Paper 1, here a risk of the act of closing rather than an inevitability of it.)

[mechanism] Bounded reflexivity: it is hard to see a frame from fully inside it. Every agent acts inside a frame, the way a program runs inside its declared variables, constants, and data sources. An agent optimizing throughput given the frame’s resources is, to that extent, operating on the frame’s own declared sources, and has limited capacity to represent the frame it is optimizing within — because representing it would require declarative information the frame does not make explicit, which must be asserted at a cost and at risk. This is a claim of bounded reflexivity, not a theorem of impossibility: the more completely an agent is absorbed in optimizing within the frame, the less of the frame it tends to see; an agent can recover some of the frame by spending effort to surface the undeclared, but only by stepping partly out of the optimization. An optimizer fully absorbed in the set is, to that degree, the set’s content rather than its observer.

So how is the frame ever seen? By acting coherently rather than rationally — not-maximizing — and grounding action in the principle rather than the map. The coherent agent has built its own frame, anchored upstream, and reasons outward from there; that is what lets it see the system whole, precisely because it is no longer optimizing within the system. The external injector of Paper 1 is therefore not external by location but by mode: it is the coherent agent, whether a true outsider or the inside agent who refuses to maximize. As a system diverges, the coherent agent, the structural-awareness bearer, and the load-bearing off-map capability increasingly turn out to be the same agent seen from three angles — the convergence sharpening toward identity in the insolvent limit (developed in Paper 3); in a healthy system they remain distinct.

10. The frame as a regime detector

[conjecture] From a well-chosen frame, the capabilities that matter appear as nested sub-frames, and one can watch the distance between one’s own frame and the frame of a nested system. A key capability is either inside the frame the system will hold, or outside it — call it inside (F) or outside (−F); both are still inside your frame, so from the structural-awareness position you can see which side of the boundary the capability is on, and the moment it crosses out. When a load-bearing capability passes outside the set the frame will hold, the system has already reframed itself to exclude its own load-bearing part — it has acted against its own self-preservation, before any visible decision. The crossing is the event; the cut is its later bookkeeping. Structural self-awareness is thus the earliest regime-change detector — earlier than the failure, earlier than the cut.

[exposed flank] The detection is only as good as the frame is correctly chosen: from the wrong frame you detect a crossing that is not one, or miss one that is. Frame-selection is this result’s exposed flank, the way independent estimation of contribution is the economic series’. The regime change being detected is in the nested system (the capability leaving its survivable regime); the instrument is the widening between your frame and theirs.

(Firewall: this is the same shape as regime-departure detection in independently developed early-warning work. The resonance is named and the firewall kept — any convergence is reserved for a separate paper, because a convergence is evidence only if the lines were independent in fact.)

— Iván Abril Palma