Agency in the Emergent Multiverse
Where the laws meet the branches
Force the coin. By the accounting of the kybit, changing a fair 50/50 distribution into certainty registers one kybit relative to that baseline. The fingers, solenoid, or other controller uses physical resources, but only a restricted equilibrium-reference realization earns the lower bound \(k_B T \ln 2\) from the information divergence itself. Part I’s robust result was narrower than a universal exchange rate: agency is physically implemented, resource-bounded, and limited.
Now run the same intervention in the Quantum Branching Universe (QBU) — the branching ontology mapped in its own chapter. An imperfect controller will usually leave some nonzero Measure on tails; a model that assigns tails zero does not secretly restore it through branch abundance. So what did the joules buy? If the causal comparison is sound, they helped implement a policy associated with a different conditional distribution of later records than the stated baseline.
The volume’s two halves — physical control and the branching model — meet here. David Wallace’s The Emergent Multiverse supplies the most developed version of the relevant picture: branches emerge through decoherence, worlds are patterns rather than fundamental furniture, and rational decision is formulated for agents inside the quantum state. Set that machinery beside the proposed agency laws and a translation becomes possible. Whether the translation is a derivation is the question this chapter must keep open.
Control in a Branching Model
Start with the first law: physical control requires a resource-consuming implementation, with a kybit-proportional lower bound only under stated thermodynamic conditions.
In Wallace’s picture an agent and its decision process are physical components of unitary evolution. The agent does not stand outside the wavefunction and choose amplitudes. What can be compared are conditional distributions: the Measure of outcomes correlated with policy A versus the Measure correlated with policy B, from a declared Vantage. If a causal model supplies an appropriate inactive or alternative-policy baseline, the divergence between those distributions can be measured in kybits. That makes it a possible unit of agent-relative control in the QBU. It does not yet show that an agent literally redistributes global Measure or that the work cost is determined by the divergence.
Wallace’s quantum decision theory argues that agents satisfying specific rationality axioms should maximize Measure-weighted expected utility. The Deutsch–Wallace program thereby offers a decision-theoretic route to Born weighting, whose epistemic counterpart is developed in Probability Without Collapse. This is a normative connection between rational preference and Measure, not a thermodynamic proof that accurate Credence minimizes joules per kybit. Better models often reduce wasted action in real systems; the theorem itself does not establish an energy optimum.
Decoherence Underwrites Decay
The second law: a finite isolated agent cannot replenish its total remaining control budget indefinitely.
Decoherence and agency decay share an environment but should not be collapsed into one mechanism. Decoherence arises when phase information disperses into environmental degrees of freedom, making interference between coarse-grained records inaccessible. Maintaining a controller against noise also requires correction, repair, and usable gradients. Both processes involve an open system coupled to an environment, but one cannot infer that every “branching event” consumes one corresponding unit of an agent’s free-energy budget.
The emergent multiverse is therefore compatible with the second law rather than an independent derivation of it. Cut an agent off from usable gradients and its finite physical organization eventually loses the ability to model and correct. In QBU language, drift is the distribution produced by the comparison dynamics without the modeled intervention. No one needs to “pay” for Measure to follow unitary evolution; payment enters when a physical controller maintains the correlations that distinguish its policy from that baseline.
No Frictionless Branches
The third law: perfect, frictionless control is physically impossible.
Quantum mechanics blocks the fantasy of an agent selecting one realized branch while deleting the others. It also imposes uncertainty, finite distinguishability, and limits on control under incomplete access. But unitary operations can be logically reversible, so the argument should not claim that every quantum manipulation necessarily dissipates a fixed amount. The physical controller as a whole still faces finite precision, preparation, correction, and actuation costs.
This kills a fantasy that branching universes invite: the agent as branch-selector, surfing worlds like channels. Nothing in the physics supports it. The more defensible claim is that every embedded controller remains bounded and that Everettian branching offers no escape hatch from those bounds.
Meaning Is Found in Measure
Now return the opening question to the table. If multiple modeled outcomes occur, does your agency matter?
Yes, conditionally on the decision-theoretic role assigned to Measure. You do not select one exclusive future. Your physical policy is correlated with a different Measure-weighted distribution of outcomes than alternative policies, and a rational agent who accepts the Born-weighting assumptions has reason to care about that difference.
Measure is obtained from squared amplitudes over the relevant decoherent components and defined more carefully in Measure, Vantage, Branchcone. A higher-Measure event sector is not more real than a lower-Measure one, but the decision-theoretic program says it should count for more in expectation and evaluation. Agency, on this reading, concerns which Measure-weighted outcomes are conditionally associated with the agent’s policy. Older language about “concentrating” or “shifting” Measure is retired in favor of that explicit causal comparison.
In a deterministic multiverse where every modeled outcome exists, the framework locates practical significance in Measure rather than branch count. You do not control what exists; you enact one physical policy rather than another, and the policies have different Measure-weighted consequences. The further claim that agents are ethically answerable for those weights belongs to measure responsibility. Physics supplies the weights and correlations; it does not supply the value judgment by itself.
Sharing a World
Everything so far treats the agent alone inside the branching model. But steering is rarely solitary. We persuade, coordinate, promise, compete — and all of that presupposes something the multiverse makes suddenly non-trivial: that the agent you are dealing with is in your world. The multiverse contains observers beyond counting. What makes a particular one share yours?
Observers in the QBU are not abstract Cartesian minds; they are physical patterns embedded in decohered records, who join branches rather than create them. For questions of cross-branch continuity, model an observer with a Strong Pattern Identifier (PI) that combines functional constraints with provenance from a declared ancestor:
\[ \text{PI}_{\text{observer}} = \{\, P : P \text{ matches the defining pattern of a given observer across branches} \,\} \]
These PIs pick out continuations treated as the same observer for the analysis. The identity criterion is declared by the model rather than discovered from resemblance alone.
Observers do not enjoy arbitrary access to all possible decompositions. They operate through stable environmental records, spatial location, object permanence, and learned interpretive models. I propose that two observers share an effective world at a Vantage \(V\) to the extent that they remain mutually observable through a common environment and their models support communication. That motivates an Observer Class at \(V\):
\[ \mathcal{O}(V) = \{\, \text{PI}_i : \text{PI}_i \text{ is coherent and communicative from } V \text{ forward} \,\} \]
— the modeled set of observers sharing enough physical and semantic structure for interaction. Sharing comes in degrees. Define Observer Class Alignment (OCA) as a placeholder for that degree:
\[ \text{OCA}(\text{PI}_A, \text{PI}_B, V) = \text{degree of effective basis overlap at vantage } V \]
Candidate observables include mutual predictability of measurement outcomes, shared semantic encoding, common environmental records, and functional compatibility of memory and causal models. No single metric is derived here. OCA is a research proposal for describing degrees of practical world-sharing, not a settled physical quantity.
The modest implication is enough: coordination and shared empirical inquiry require overlapping records and interpretable channels. Ethical standing does not. An agent may matter despite distance, disability, future location, or inability to communicate, and Volume 5’s agent binding must not be reduced to OCA. The observer-class proposal concerns practical interaction inside an emergent world; it does not define the full boundary of obligation or ontology.
The Honest Ledger
The objections chapter promised that this chapter would keep the ledger on the framework’s known gap. The gap is this:
The claim that intentional distributional control has a general quantitative relationship to policy-conditioned Measure differences is not derived here. The thermodynamics establishes resource constraints and a lower bound in a restricted realization. The QBU supplies a modeled arena. The bridge between them — a general relationship among causal control, KL divergence, work, and conditional event weight — remains a commitment of the framework. In the spirit of the ledgers Volume 2 keeps, here is what a genuine derivation would need to supply:
A defined baseline. The claimed difference is a comparison, and the comparison class is a counterfactual: the Measure distribution over the Branchcone forward of the agent’s Vantage had the agent’s predictive machinery been inert — the baseline distribution \(P_{\text{initial}}\), restated as a fact about event weights. That counterfactual must be constructed with the branch-based semantics of Causality and Counterfactuals and shown to be well-defined in a branching ontology, not smuggled in from single-world intuitions.
A dynamical bound. A demonstration identifying the physical conditions under which a predictive control process changes conditional branch weights relative to that baseline, and by how much. A fluctuation-theorem-style result might bound divergence by dissipated work, but it would have to show that the relevant reference distribution and protocol satisfy the theorem’s conditions rather than assuming a general Landauer exchange rate.
A no-conjuring proof. Unitarity preserves total normalized Measure. Any policy difference must concern correlations between action processes and later event sectors — never the creation, destruction, or global reallocation of amplitude. The derivation must show that this correlational difference is strong enough to underwrite the counterfactuals the framework asserts, and no stronger than unitarity permits.
A triviality check. In a deterministic wavefunction, every Measure distribution is what unitary evolution delivers, agents included. A derivation must locate the causal difference made by the policy variable rather than redescribe Measure as flowing through an agent-shaped channel. Until that intervention-relative difference is rigorous, the objection that the comparison is bookkeeping rather than achievement remains live, and I list it here as live.
These are honest debts. What stands independently is narrower: agents are physical control systems; Everettian agents are components of unitary evolution; alternative policies can be compared by their conditional Measure-weighted consequences; and practical world-sharing depends on common records and channels. The claim that “when you spend the joules, the weights move” compresses several missing steps and should be retired. The proposed object is an intervention-relative causal comparison, not a theorem about modifying global amplitude.
The three proposed laws remain compatible with the emergent multiverse, and the translation sharpens their open questions along with their claims. Part III turns inward toward what it means to choose when the bookkeeping contains a branching structure of realized outcomes — beginning with Everett’s Demon.