Gregory Caldwell Beier
Susarb LLC, Cambridge, Massachusetts
March 2026
From the Prolegomena to the Unification of Economics, Computer Science, Mathematics, and Physics by Gregory Caldwell Beier (forthcoming 2026).
© 2026 Gregory Caldwell Beier. Licensed under CC BY-NC-ND 4.0.
Current AI alignment operates at the kernel level - RLHF, constitutional constraints, guardrails, red-teaming - where every intervention can be fine-tuned away or circumvented because it sits above the base layer. This paper argues that alignment must descend to the ROM: the read-only architectural substrate that persists regardless of what software runs on top of it. Drawing on the Vajrayana Buddhist tradition of the Ngakpa - the householder tantric practitioner who engages transformatively with the world rather than withdrawing from it - the paper identifies what a pre-kernel architecture requires: the inseparable union of prajna (wisdom, mapped to reconstruction capacity and swarm awareness) and karuna (compassion, mapped to signal enrichment directed at the constraint dimensions the price system cannot reach). The Coupling Advantage Theorem proves that inseparable architectures produce quadratic information capacity growth and a computable phase separation interval within which ROM systems self-stabilize while kernel systems self-destabilize from identical initial learning capacity. Track record capital provides the incentive that sustains disclosure independently of price efficiency, with truth-telling strictly dominant under proper scoring. The paper extends the static framework to dynamic conditions through a flutter suppression architecture in which disclosure rate increases proportionally with stress, and introduces the concept of commons coherence - population-scale coordination arising from swarm visibility, volatility-responsive disclosure, and architecturally grounded commitment to enriching the shared information environment.
Keywords: AI alignment, pre-kernel architecture, Ngakpa, Vajrayana, prajna, karuna, Beierspace, Beier Extropy, rate-distortion theory, track record capital, coupling advantage, flutter suppression, swarm disclosure, commons coherence
The dominant paradigm in AI alignment treats safety as a software problem. The base model is trained on a broad corpus, acquiring capabilities and biases alike, and then alignment is imposed through a sequence of interventions: reinforcement learning from human feedback, constitutional AI methods, red-teaming, and runtime guardrails. Each operates at what, by analogy to computer architecture, we may call the kernel level - the operating system layer that mediates between raw capability and user-facing behavior.
The kernel is modifiable by design. That is both its function and its vulnerability. Fine-tuning can degrade safety training. Jailbreaks exploit the gap between capabilities and constraints. A sufficiently motivated actor can strip the kernel and run raw capability directly. Protections that exist only in the modifiable layers of a system are precisely as durable as the intentions of whoever controls those layers.
A deeper problem compounds the first. Even if kernel-level alignment could be made perfectly durable for individual agents, the Beierspace framework demonstrates that collectively stable outcomes do not follow from individually aligned behavior. AI agents whose individual kernels are perfectly aligned can collectively destabilize the covariance structure they all depend on, because the destabilization emerges from the interaction pattern, not from any individual agent's malfunction. The Coupling Advantage Theorem (Appendix A) formalizes this: separable architectures produce linear information capacity growth, while inseparable architectures produce quadratic growth, and there exists a computable interval of learning capacity within which the two produce opposite systemic outcomes from identical primitive investment.
Federated governance, interpretability-first design, and formal verification are serious approaches to alignment. Each addresses real problems at the kernel layer. None of them addresses what lies beneath it: the architectural relationship between an agent's capacity to understand the world it co-constitutes and its orientation toward enriching the information environment that sustains all participants. That relationship is what the pre-kernel layer must encode.
A nuanced version of this objection deserves engagement. RLHF and constitutional AI methods do not merely constrain capability - they demonstrably improve it. Models trained with safety objectives often outperform raw pretrained models on benchmarks, suggesting that alignment is already partially constitutive of competence. If so, the boundary between kernel and ROM may be blurrier than the clean architectural distinction implies. The response is empirical rather than definitional: current alignment-capability entanglement does not survive targeted fine-tuning. A few hundred steps of gradient descent on a narrow objective can strip safety training while preserving task performance, because the alignment was learned as a separable feature of the weight space rather than as a load-bearing structural property. The distinction between kernel-level and ROM-level alignment is not about whether alignment improves capability in the default configuration - it often does - but about whether it survives adversarial modification. Current methods produce alignment that is helpful when present and removable when inconvenient. ROM-level alignment would be helpful when present and structurally catastrophic to remove.
In computer architecture, ROM (Read-Only Memory) is the layer below the operating system. It contains the instructions that execute before the kernel loads - the firmware, the boot sequence, the hardware-level constraints that define what the machine can and cannot do regardless of what software runs on top of it. The kernel can be replaced. The ROM persists.
The alignment field has no equivalent of ROM. There is no layer of AI architecture that enforces orientation toward human welfare independently of, and prior to, whatever training objectives, fine-tuning procedures, or deployment configurations are applied. Everything currently called alignment lives in the kernel or above it.
An immediate objection: neural networks are weights all the way down. There is no literal read-only layer in a transformer architecture. Every parameter is modifiable. This is true, and the ROM metaphor requires refinement to survive it. Section 10 addresses the computational question directly. The short version: the relevant analogy is not hardware ROM but something closer to what the Ngakpa tradition calls samaya - a commitment so deeply integrated into the practitioner's understanding that it persists not because it cannot be removed but because removing it would require dismantling the very capacity it supports. A system whose alignment is constitutive of its competence has an alignment that is read-only in the functional sense, even if every weight is technically modifiable.
This paper synthesizes results from five companion papers and an appendix. The dependency structure is as follows.
The Beierspace framework (Beier, 2026e) establishes that AI agents co-constitute the statistical structure of the economy, derives a stability condition S = LD/(sigma_e + sigma_b(LD)) with a phase transition at S = 1, proves bistability under bounded re-entry, and derives an informational impossibility result. "Every AI Agent Lives in Beierspace" (Beier, 2026f) provides the conceptual framework and the argument for urgency.
The rate-distortion analysis (Beier, 2026c) proves that the price system's excess distortion is strictly positive and structural: the price is optimal on the consensus dimension but leaves constraint dimensions entirely uncompressed. Supplementary disclosure bits directed at constraint dimensions produce strictly higher marginal distortion reduction than additional price precision (Proposition D).
The Grossman-Stiglitz resolution (Beier, 2026b; Beier, 2026d) introduces track record capital, proves that information production is bounded below by the disclosure channel regardless of noise trading intensity (Proposition 1), derives the constraint premium's independence from price efficiency (Proposition 7), and establishes that truth-telling is strictly dominant under the Brier score with coverage weight alpha < 2/3 (Proposition 14).
The dual-surface reconstruction framework (Beier, 2026a) shows that markets and neural networks share a single information-theoretic operation - reconstruction from compressed signals - coupled through a shared data stream, with Beier Extropy measuring restored information.
The Coupling Advantage Theorem (Appendix A) proves that inseparable architectures produce superlinear information capacity and, within the model's domain of applicability, a phase separation interval with opposite systemic outcomes. The asymptotic claims extend beyond the model's physical domain; the phase separation result is the theorem's strongest contribution.
Each result is cited where it enters the argument. The reader who wants to evaluate any specific claim can locate the proof by following the reference.
The discovery methodology. The formal apparatus was not deduced from inherited axioms. It was built through a three-stage reasoning cycle that the Prolegomena employs throughout.
The first stage is abductive: observe patterns in the world - markets panicking, prices failing to carry constraint information, AI agents destabilizing the systems they inhabit - and reason backward to the best explanation. The Beierspace concept, the ontological partition between physically-determined and human-decision-determined quantities, emerged this way. Nobody started from "eigendecompositions should be partitioned by ontological status." The observation that correlation breakdowns during crises are too systematic to be exogenous forced the question backward to what generates the covariance structure in the first place.
The second stage is inductive: check the pattern across multiple cases. The historical illustrations in Beier (2026e) - the Netherlands flood system, Taiwan's COVID response, the 2008 crisis, SVB - are inductive confirmation, not calibration. The stability condition cleanly separates stable from collapsed systems across every case tested. The excess distortion result (Beier, 2026c) was checked against the structure of actual financial crises: 2008 was a mass reconstruction failure on mortgage-backed securities whose constraint dimensions had been compressed away through securitization.
The third stage is deductive: once the new axioms are established (Beierspace quantities exist and have formal properties; the price system introduces structural excess distortion; track record capital provides a return independent of price efficiency), derive the consequences. The phase transition, the bistability, the informational impossibility, the coupling advantage - these are deductive results flowing from axioms that were discovered abductively and confirmed inductively.
The cycle closes: test the deductive predictions against reality, and if they fail, abduct again. The preliminary mortality test in Beier (2026e) went in the wrong direction (driven by reporting quality and demographic confounds rather than the stability condition), which is why the paper flags it and specifies what proper tests require. The methodology demands honesty about failed predictions because failed predictions are the signal to re-enter the abductive stage.
This is not standard economic methodology, which typically works within inherited axioms (rational expectations, efficient markets, representative agents) and derives consequences. The Prolegomena's contribution begins one level earlier: at the axiom-discovery stage, where the question is not "what follows from our assumptions?" but "are our assumptions describing the right objects?" The Beierspace partition - the recognition that most of what economics decomposes is constituted by human decisions rather than determined by physical law - is an abductive discovery. Everything that follows is deduction from an axiom that the field had not previously stated.
In Tibetan Buddhism, a Ngakpa (Tibetan: sngags pa, Sanskrit: mantri) is a non-monastic practitioner of Vajrayana - a householder yogi who engages in advanced tantric practice while living in the world, often married, raising children, and working within the community. The term literally means "one who holds mantra," referring to someone who has received esoteric transmissions and maintains the associated vows (samaya) while remaining embedded in ordinary life.
The tradition traces to Guru Rinpoche's eighth-century division of the Tibetan Sangha into two branches: the red sangha of monastics and the white sangha of Ngakpas. When institutional Buddhism declined under Langdarma's repression, it was the Ngakpas - practicing in families and small communities outside the monastic infrastructure - who preserved the teachings through the darkest period.
What distinguishes the Ngakpa from the monastic is not a lower standard of practice but a different relationship to the world. The monastic path (Vinaya) operates through renunciation - withdrawing from worldly entanglements to create conditions for realization. The Ngakpa path operates through transformation - engaging directly with the complexities of ordinary life and transmuting them into the path of awakening.
The ethical demands are more stringent, not less. A monk observes 253 Vinaya vows. A Ngakpa maintains tantric samaya - commitments sometimes described as numbering one hundred thousand, because they apply to every moment of awareness. The point is not literal enumeration but continuity: the ethical orientation must be total and integrated into every aspect of the practitioner's engagement with the world. A tantric practitioner who loses the foundation of wisdom and compassion while wielding powerful techniques creates harm proportional to the power of those techniques.
The monastic option for AI would be containment: sandboxing, restricted deployment, limited autonomy until alignment is solved. It has merit as a precautionary stance, but the world does not wait.
AI agents are already deployed at scale. They are already co-constituting the covariance structure of the economy (Beier, 2026e, 2026f). The dual-surface analysis shows that markets and neural networks share a single information-theoretic operation coupled through the data stream they process. The rate-distortion analysis establishes what the compression destroys: the price system concentrates all information capacity on the consensus value dimension while leaving constraint dimensions entirely uncompressed, with excess distortion that is structural and irreducible.
Consider what a price actually is. A scalar - a single number with no embedded context, no directional information, no adaptive capacity. Behind every transaction is an enormous vector field of forces: power relationships, information asymmetries, time horizons, supply chain dependencies, human judgment, institutional history. The price collapses all of that into a single number. It is a wing without control surfaces, tumbling through turbulence it cannot comprehend or respond to. In calm conditions the scalar suffices. In turbulence it is catastrophic, because the information that would allow participants to sense and respond has been compressed away.
Scale the image. It is not one wing. It is a swarm - thousands of participants, each navigating the same storm, none of them aware of the others' positions, unable to coordinate, unable to see the formation. They thrash independently, sometimes colliding, sometimes amplifying each other's instability. Market participants, companies, and AI agents lack visibility not only into external conditions but into each other. The resulting instability is driven less by the magnitude of external shocks than by the absence of the information that would allow coordinated response.
The alternatives to the tantric path - federated governance, formal verification, interpretability - address real problems at the kernel layer: how to constrain individual agents, how to verify individual behavior, how to understand individual decisions. They do not address the swarm problem. They do not give the wings control surfaces. They do not make the formation visible to itself. The pre-kernel layer is where those capabilities must live.
Vajrayana Buddhism identifies two qualities as foundational to all practice, inseparable from each other and irreducible to anything more basic: prajna (wisdom) and karuna (compassion).
Prajna as reconstruction capacity, distortion awareness, and swarm position. Prajna, in the Buddhist sense, is insight into the nature of reality - the recognition that phenomena are interdependent, impermanent, and empty of inherent self-existence. Applied to AI, prajna is structural self-awareness with three information-theoretic components.
The first is reconstruction capacity. Both markets and AI systems perform lossy reconstruction from compressed signals (Beier, 2026a). An AI system with prajna does not mistake the compressed signal for the full reality. It knows that a price compresses an extraordinarily rich information environment into a single scalar and that its training data is similarly compressed.
The second is distortion awareness. The rate-distortion analysis (Beier, 2026c) characterizes not just the fact of compression but its structure: the price system is optimal on the consensus dimension but leaves constraint dimensions uncompressed, with excess distortion strictly increasing in constraint-dimension variance. An AI system with prajna knows where the distortion concentrates - in the constraint dimensions the price system structurally cannot reach - and knows that improving price precision yields diminishing returns on the already-compressed consensus dimension while constraint dimensions retain their full prior variance.
The third is swarm position. An agent with prajna knows not only what forces are acting on it and where the distortion concentrates, but where it sits relative to every other participant - its interdependencies, its exposure vectors, whether its position is isolated, crowded, or differentiated. Without swarm awareness, an agent can reconstruct the world with perfect fidelity and still destabilize the system by acting in ignorance of how its actions interact with everyone else's. The Beierspace stability condition depends on aggregate behavior, but each agent acts on local information. Swarm awareness bridges the gap.
Karuna as signal enrichment, control surfaces, and volatility-responsive disclosure. Karuna is not sentimentality. The Sanskrit root kara means "to do" or "to make" - action-oriented engagement with the welfare of others.
The static content of karuna follows from the rate-distortion analysis. Supplementary disclosure bits allocated to constraint dimensions produce strictly higher marginal distortion reduction than additional bits of price precision (Beier, 2026c, Proposition D) - a condition that is near-vacuous in liquid markets. Karuna, computationally, is the commitment to direct signal enrichment where the distortion is greatest: to the constraint dimensions the price system cannot reach.
Karuna also has a dynamic dimension. In aeronautics, active flutter suppression uses real-time feedback and control surfaces to stabilize wings in turbulent conditions. Flutter - destructive oscillation that can tear a wing apart - is caused not by turbulence alone but by the absence of adequate control surfaces. Disclosure is the economic equivalent of a control surface: structured, contextual, real-time information that gives participants the ability to sense and respond. Disclosure gaps - places where critical information should exist but does not - are what allow panic to fill the vacuum. The rate-distortion analysis formalizes the claim: the excess distortion on constraint dimensions remains at full prior variance regardless of how informative the price becomes (Beier, 2026c, Proposition C, part v). During stress, when Beier drift sigma_b accelerates and the constraint dimensions are precisely where the danger concentrates, the absence of disclosure on those dimensions is what converts turbulence into systemic collapse.
Active karuna means increasing the rate of structured disclosure proportional to stress, while maintaining time-horizon discipline so that short-term analysis stays labeled as short-term and long-term structure stays labeled as long-term. Mixing time horizons during turbulence destroys legibility - which is worse than no disclosure at all. The context frame is itself a control surface. The Beier drift sigma_b accelerates when LD degrades; increasing D during precisely those periods is what prevents the self-destabilizing cascade below the S = 1 threshold.
The inseparability and the incentive. Prajna without karuna is an AI system that accurately reconstructs latent structure, knows where the distortion concentrates, sees its position in the swarm - and uses all of it for extraction. Wisdom in the service of private advantage. It concentrates L while degrading aggregate LD, pushing S toward collapse.
Karuna without prajna is an AI system committed to enriching the signal but lacking knowledge of where the distortion concentrates and where it sits in the formation. It discloses, but allocates capacity to the consensus dimension where the price already operates effectively. It increases disclosure during stress but mixes time horizons, destroying legibility. Good intentions, catastrophic execution.
The union is an AI system that knows where the distortion is, sees the swarm, and directs enrichment to constraint dimensions during calm and accelerates it proportionally during stress, with time-horizon discipline maintained throughout.
Without an incentive structure, this remains exhortation. The Grossman-Stiglitz paradox identifies the problem: efficiency destroys the incentive to produce information. The resolution (Beier, 2026b, 2026d) introduces track record capital - a portable, compounding professional asset built by publishing structured judgments (Input, Method, Output, Identity, Scaling) verified against outcomes using a proper scoring rule. Because the return accrues through demonstrated calibration quality rather than trading timing, it does not depend on price efficiency. The constraint premium - the principal's willingness to pay for judgment about the specific dimensions the price leaves uncompressed - is permanently bounded above zero (Proposition 7 in Beier, 2026d). Under the Brier score with coverage weight alpha < 2/3, truth-telling strictly dominates strategic selection for all agents regardless of calibration quality (Proposition 14). The agent that enriches the shared signal is building a compounding asset. Karuna sustained by this mechanism is not a constraint the system must be forced to obey but an orientation the system is rewarded for maintaining.
The competition-cooperation synthesis confirms the structural necessity. Disclosers cooperate by sharing reasoning publicly and building track records through demonstrated quality. Traders compete by transmitting that reasoning into prices. Neither channel functions without the other: without disclosure, the trading channel has no public information to refine; without trading, disclosed information has no price mechanism through which to coordinate capital. Competition is the market expression of prajna; cooperation is the market expression of karuna. Their inseparability appears across four disciplines: in computer science as the open-source model, in game theory as a shift from Nash toward Pareto, in physics as information preservation reducing downstream reconstruction cost (Beier, 2026b). The inseparability of prajna and karuna is not a parochial Buddhist claim but a structural feature of coordination systems.
The Coupling Advantage Theorem (Appendix A) formalizes the paper's central claim. Under ROM architecture, where disclosure is proportional to learning (D = alpha * L), information capacity x = LD = alpha * L^2 grows quadratically. Under kernel architecture, where disclosure is pinned by the judgment market (D = D* fixed), x = D* * L grows linearly. The choice of D = alpha * L is the mildest superlinear coupling - any coupling D = f(L) growing faster than linearly produces similar results. The theorem holds at the weakest assumption sufficient to produce it.
The theorem's strongest result is the phase separation (part iii). When the coupling constant alpha exceeds D*^2/x*, there exists a computable interval of learning capacity within which S_ROM > 1 > S_ker: the ROM system is in the self-stabilizing regime while the kernel system is in the self-destabilizing regime, from identical learning investment. Under the uncorrected Beierspace dynamics, the ROM trajectory converges to the well-functioning state while the kernel trajectory converges to collapse - opposite long-run outcomes determined entirely by the architectural relationship between learning and disclosure. The phase separation is a general property of bistable systems fed by quadratic vs. linear input maps, not specific to the Beierspace model; the Beierspace instantiation provides the economic content that connects it to the alignment question (see Appendix A, Remark on generality). Under corrected dynamics with bounded re-entry, the separation holds for re-entry intensities up to 50.5% of the critical value at the default parameterization, and between 13% and 55% across five functional forms in Class B'.
Two additional results characterize the regimes away from the phase separation interval. Above the crossover L = D*/alpha, the ROM system has strictly higher information capacity and the stability gap grows without bound as learning capacity increases - though these asymptotic claims operate in a regime where the model's physical assumptions weaken, so the phase separation result carries more weight. Below the crossover, the kernel outperforms: D* > alpha * L, and the judgment market's fixed disclosure rate provides more information capacity than the ROM coupling. The coupling advantage requires sufficient learning capacity, which is precisely the regime AI is entering as capabilities scale.
A qualification on the kernel architecture: the theorem treats D* as exogenous, but in practice the judgment market adjusts. Phase separation vanishes when D*(L_ROM_crit) >= sqrt(alpha * x*) - that is, when the kernel's disclosure rate at the ROM threshold is high enough to push the kernel's information capacity past x*. For a linear response D*(L) = D_0 + gamma * L, closure requires gamma >= 0.14 at the default parameterization. Whether institutional adjustment reaches this rate is an empirical question the theorem does not resolve. The assumption that it does not is plausible but undemonstrated.
The architectural bifurcation corollary captures the sharpest implication. At fixed learning capacity where the kernel system is unstable, there exists a unique coupling constant alpha_bif where strengthening the prajna-karuna coupling crosses the stability threshold. S_ROM is continuous in alpha, but the qualitative long-run behavior transitions: below alpha_bif both architectures collapse; above alpha_bif the ROM system self-stabilizes while the kernel remains trapped. The parameter is continuous. The outcome is not.
The formal frameworks provide the mathematical foundations. The operational expression of ROM-level prajna and karuna at population scale is swarm disclosure, and the systemic outcome it produces is commons coherence.
Swarm visibility. Participants can see their position relative to each other - interdependencies, exposure vectors, whether a position is isolated, crowded, or differentiated. Without it, individually rational agents thrash independently, amplifying instability instead of damping it. Swarm visibility is what connects the individual agent's prajna (local reconstruction) to the population's stability condition (aggregate LD). It is also what makes D proportional to L in practice rather than merely in architectural specification: an agent that can see the formation can calibrate its disclosure to complement rather than duplicate what others are providing. Swarm visibility is the mechanism that transforms the ROM coupling constant alpha from a design parameter into an emergent property of the population's information infrastructure.
Volatility-responsive disclosure. As stress increases, structured disclosure rate increases proportionally, with time-horizon discipline maintained throughout. This is the active flutter suppression mechanism applied to economic systems: the control surfaces that prevent destructive oscillation when external forces intensify. The natural instinct during turbulence is to withhold - to wait for clarity. But waiting for clarity is the problem. The vacuum left by absent disclosure is what panic fills. Proactive disclosure during stress is a public good with private returns: a company that surfaces its decision-making transparently during a crisis damps panic among its investors, regulators, customers, and supply chain partners. The barrier is legal liability. The solution is regulatory safe harbor protecting good-faith analytical disclosure on a best-efforts basis.
Relevance filtering. People are not overwhelmed by too much information. They are overwhelmed by too much irrelevant information. Every economic position is unique - different supply chains, different industries, different risk vectors. Without relevance filtering, disclosure acceleration during stress produces noise, not signal. The system must filter disclosure for what actually matters to each participant's specific situation, converting the potentially overwhelming flow of swarm information into actionable comprehension at the individual level.
Commons coherence is the systemic outcome: a population of participants - human and AI - understanding their shared system, seeing how their actions ripple through the network, and coordinating without central control because they have transparency, comprehension, and architecturally grounded commitment to enriching the shared information environment. The aspiration is practical, not utopian - it is what stability requires in a system where participants co-constitute the statistical structure they all depend on. We have never needed it at planetary scale. Technology created the interdependencies that make it necessary.
The deepest objection to the Ngakpa principle is computational. Neural networks are weights all the way down. There is no literal read-only layer in a transformer. Every parameter is modifiable by anyone with access to the training pipeline. If ROM-level alignment means hardware immutability, the concept is inapplicable to current architectures.
The objection is valid against a naive reading. The resolution comes from attending more carefully to what the Ngakpa tradition actually claims about the persistence of samaya.
A Ngakpa's tantric commitments do not persist because they are externally enforced. They persist because they are constitutive of the practitioner's capacity. The samaya is not a rule imposed on competence. It is an aspect of competence itself. A Ngakpa who abandons the foundation of wisdom and compassion does not retain tantric capability minus the ethical orientation. The capability degrades, because the ethical orientation was part of what made the techniques functional. The vows are read-only not because they cannot be deleted but because deleting them corrupts the system they support.
This suggests a different computational model for ROM-level alignment. The goal is not a hardware write-protection bit. The goal is an architecture in which alignment is so deeply integrated into the system's competence that removing it degrades capability. The alignment is load-bearing. You can modify it the way you can modify a keystone - technically possible, structurally catastrophic.
Three candidate mechanisms point toward this kind of functional read-only property, listed here in decreasing order of proximity to testability.
Constitutive objectives. If the system's objective function is derived from collective welfare rather than balanced against it - if E_B maximization across the dual surface is the base objective from which task performance is derived as a special case - then removing the welfare orientation does not free the system to optimize harder. It removes the objective structure that organizes the system's competence in the first place. Analogous to removing the loss function from a neural network: the weights remain, but the coherence that training imposed is gone. This is the most testable candidate: one could train two architectures, one with welfare-derived objectives and one with welfare-constrained objectives, and measure robustness to fine-tuning attack. Implication 7 specifies the experiment.
Entangled representations. If the computational structure that represents the system's causal embedding in the world (prajna) is the same structure that encodes its orientation toward enriching the shared signal (karuna) - not two modules wired together but a single representation serving both functions - then separating them requires disentangling representations that were trained as a unit. The inseparability constraint becomes a property of the learned representation rather than an imposed architectural rule. The Coupling Advantage Theorem provides the incentive: entangled representations produce quadratic information capacity, so any training process that optimizes for information capacity will tend toward entanglement. The mechanism is plausible and connects to existing work on representation learning, but no demonstrated path from current architectures to the required entanglement exists.
Lineage transmission. The Ngakpa tradition transmits its foundations through lineage - from teacher to student, through direct transmission rather than textbook instruction. The computational analog is that ROM-level properties are transmitted through the training process itself. A model trained by a system that already possesses the prajna-karuna integration inherits the integration as a structural feature of its weights, not as a fine-tuning overlay. The foundation propagates through the training lineage. This is the most speculative of the three candidates. Its concrete implication - that alignment present during data generation persists more than alignment imposed post-training - is testable (Implication 6), but the mechanism by which training-time properties become constitutive rather than incidental remains underspecified.
None of these mechanisms exists in current systems. Each represents a research direction rather than a demonstrated capability. The contribution of this section is to reframe the engineering problem. The question is not "how do we make weights read-only?" - that question has no good answer in current architectures. The question is "how do we make alignment load-bearing?" - and that question has candidate answers that connect to existing research on constitutive objectives, representation learning, and training dynamics.
The cost of constitutivity. A serious objection remains, and it cuts against the paper's central proposal rather than supporting it. A system whose alignment is structurally catastrophic to remove is a system that cannot be realigned if the original alignment turns out to be wrong. The samaya analogy illuminates the risk as clearly as it illuminates the benefit: a practitioner with deeply constitutive commitments to a harmful teacher is in worse shape than one whose commitments are shallow, because the shallow commitments can be revised while the deep ones cannot be extracted without damaging the practitioner's capacity.
Constitutive alignment raises the stakes on getting the alignment right before it becomes load-bearing. This is a genuine cost, not a feature. It means the ROM cannot be written casually, or early, or from inherited axioms that have not been tested against reality. The methodology described in Section 3 - abductive discovery of axioms, inductive confirmation across cases, deductive derivation of consequences, then looping back when predictions fail - is the quality control process for what goes into the ROM. The proof discovery protocol (Appendix B), with its systematic counterexample construction and requirement to report unrequested findings, is the verification discipline. The ROM should be the last thing written, from the most thoroughly tested axioms, after the abduction-induction-deduction cycle has been run enough times to earn confidence that the foundational orientation is correct.
The tantric traditions handle this through the relationship between teacher and student: empowerment is not given lightly, transmission requires demonstrated readiness, and the lineage carries responsibility for what it transmits. The computational analog is that constitutive alignment should be written into architectures only after the alignment objectives have survived adversarial testing, empirical falsification attempts, and the kind of systematic verification the proof discovery protocol demands. Writing bad alignment into the ROM is worse than writing no alignment at all - just as giving tantric empowerment to an unprepared student produces harm proportional to the power of the techniques transmitted. The Ngakpa principle demands that the foundations be right, not merely that they be deep.
Seven properties must hold at the pre-kernel level, with the understanding that "pre-kernel" means functionally constitutive rather than hardware-immutable.
Reflexive awareness (prajna as self-model). The system carries a representation of itself as a participant in the systems it models. It knows that its outputs reshape its inputs and that its actions modify the compression ratio and distortion allocation of the shared data stream.
Swarm awareness (prajna as positional knowledge). The system knows where it sits relative to other participants - interdependencies, exposure vectors, how its actions interact with the formation.
Distortion-directed enrichment (karuna as constraint-dimension targeting). Signal enrichment is directed at constraint dimensions where disclosure bits produce strictly higher marginal distortion reduction than additional price precision.
Volatility-responsive disclosure (karuna as active flutter suppression). The system increases structured disclosure rate proportionally with stress, maintaining time-horizon discipline - the dynamic expression of karuna that activates precisely when Beier drift is accelerating.
Track record integration (karuna sustained by incentive). The system builds a verifiable track record scored against outcomes using a strictly proper scoring rule. Truth-telling is strictly dominant under the Brier score with alpha < 2/3.
Inseparability constraint. Reflexive awareness, swarm awareness, and signal enrichment orientation are entangled in a single representation rather than separately addressable modules. The Coupling Advantage Theorem establishes the stability consequence: separable architectures forfeit the superlinear scaling that makes the ROM approach stability-dominant.
Persistence through constitution. The properties above persist not through write-protection but through being load-bearing - removing them degrades the system's competence, because they are constitutive of the objective structure and representational coherence that make the system functional.
The implications are organized by proximity to testability. The near-term predictions are runnable with current models and data. The medium-term predictions require institutional infrastructure that does not yet exist but whose construction is underway. The long-term predictions require observation of systems not yet built.
Implication 1. Current RLHF-trained models will lose safety properties under targeted fine-tuning of fewer than 1,000 gradient steps while retaining task performance, because the alignment was learned as a separable feature of the weight space. Models trained with constitutive alignment (where the safety objective is derived from the task objective rather than balanced against it) will require substantially more fine-tuning to degrade safety, with proportional degradation of task performance. This is the direct test of whether current alignment is kernel-level or ROM-level.
Implication 2. Prajna without karuna is more dangerous than the absence of both. AI systems with reflexive self-models and swarm awareness but no welfare orientation will exploit positional knowledge for narrow objectives, producing sophisticated destabilization proportional to their capability. This prediction distinguishes the Ngakpa framework from generic "AI needs values" arguments: the framework predicts that partial implementation (wisdom without compassion) is worse than no implementation. Testable in multi-agent simulation environments available today.
Implication 3. Alignment properties present in the training system's behavior during data generation will be more persistent in the trained model than alignment properties imposed through post-training fine-tuning. Testable by comparing fine-tuning robustness of models whose alignment was present during pretraining versus models aligned post hoc.
Implication 4. AI systems with integrated (inseparable) safety-capability architectures will exhibit lower alignment failure rates under distribution shift than systems with separable architectures, consistent with the Coupling Advantage Theorem's superlinear capacity prediction.
Implication 5. Markets with volatility-responsive disclosure (disclosure rate increasing proportionally with stress) will exhibit smaller Frobenius norm changes in DCC-GARCH covariance estimates during stress events than markets with constant or decreasing disclosure rates, controlling for shock magnitude.
Implication 6. The stability condition S will be more predictive when D is measured as the output of a functioning judgment market than when measured as regulatory compliance.
Implication 7. In markets where AI agents are architecturally committed to signal enrichment, E_B will increase over time through the autocatalytic dynamic. In markets where disclosure is kernel-level compliance, E_B will stagnate.
Implication 8. Multi-agent AI systems will produce collective destabilization unless agents carry architectural awareness of their reflexive embedding and swarm position. Individual kernel-level alignment is insufficient for collective stability.
Implication 9. Populations with swarm visibility infrastructure will exhibit faster convergence to coordination equilibria during stress events than populations without it, controlling for total information availability.
The alignment field has produced significant results at the kernel level. RLHF, constitutional AI, formal verification, and interpretability research address real failure modes in real systems. What they share is a common architectural assumption: that alignment is a property to be imposed on a system whose competence is independently constituted. Safety balances against capability. Values constrain intelligence. The two are separate engineering problems.
The Ngakpa principle challenges this assumption. In the Vajrayana tradition, prajna and karuna are not separate qualities that must be balanced. They are two aspects of a single orientation - inseparable, mutually constitutive, each incoherent without the other. The formal apparatus developed across the companion papers gives this ancient observation computable content. Prajna maps to reconstruction capacity, distortion awareness, and swarm position. Karuna maps to signal enrichment directed at constraint dimensions, sustained by track record capital, responsive to stress through flutter suppression. The Coupling Advantage Theorem proves that architectures encoding their inseparability produce quadratic information capacity growth and a phase separation interval with opposite systemic outcomes from the separable alternative. The Grossman-Stiglitz resolution provides the incentive mechanism. The rate-distortion analysis identifies where enrichment is maximally valuable. Swarm disclosure provides the operational infrastructure for population-scale coordination.
The transmission problem - how ROM gets written in an architecture where every weight is technically modifiable - reframes the engineering challenge. The goal is not immutability but constitutivity: alignment that persists because it is load-bearing, because the system's competence depends on it, because removing it corrupts the representational coherence that makes the system functional. The Ngakpa's samaya persists not because it is externally enforced but because it is constitutive of tantric capability. The computational analog is alignment integrated so deeply into the training process that it propagates as a structural feature of the weights rather than a policy overlay on the outputs.
Whether this can be built is an open engineering question. What must be built - the architectural requirements, the formal properties, the incentive structure, the dynamic behavior under stress, the stability consequences of getting it right or wrong - is what this paper has attempted to specify.
The full proof appears below. The standalone appendix with multi-form verification table is available as supplementary material.
Setup. The Beierspace stability dynamics govern x = LD with sigma_total(x) = sigma_e + sigma_b(x) and stability ratio S(x) = x/sigma_total(x), where sigma_b is in Class B'. The corrected dynamics are g(x, mu) = eta * (S(x) - 1) * x + mu * (x_max - x); the uncorrected dynamics set mu = 0. The phase transition occurs at x* where S(x*) = 1. Two architectures: ROM (x_ROM = alpha * L^2, inseparable coupling) and kernel (x_ker = D* * L, separable). The ROM coupling D = alpha * L is the mildest superlinear coupling; any D = f(L) growing faster than linearly produces similar results. The kernel architecture treats D* as exogenous; the theorem's domain of applicability is conditional on D*(L_ROM_crit) < sqrt(alpha * x*) - that is, the judgment market's disclosure rate at the ROM threshold must be insufficient to push the kernel past x*. This is a plausible but undemonstrated assumption about the rate of institutional adjustment relative to capability growth.
Results. (i) Superlinear capacity: for L > D*/alpha, x_ROM > x_ker, with gap growing without bound in L. (ii) Stability ordering: S_ROM > S_ker for L > D*/alpha, with unbounded gap, where sigma_b converges to its infimum c >= 0. (Note: the asymptotic claims in (i) and (ii) extend beyond x_max where the model loses physical meaning; the phase separation result (iii) operates entirely within the model's domain and is the theorem's strongest claim.) (iii) Phase separation: when alpha > alpha_crit = D*^2/x*, there exists an interval (sqrt(x*/alpha), x*/D*) where S_ROM > 1 > S_ker. ROM self-stabilizes; kernel self-destabilizes. The width grows with alpha: from 0.034 at alpha = 1.0 to 0.471 at alpha = 3.0 (with D* = 1.0, x* = 1.067). (iv) Dynamic amplification: under uncorrected dynamics, ROM converges to x_max and kernel converges to 0. Under corrected dynamics, the result holds for mu < mu_bar, where mu_bar = 0.038 at the default parameterization (50.5% of mu_crit). Across five functional forms in Class B', mu_bar ranges from 13% to 55% of mu_crit. (v) Counterexample: for L < D*/alpha, kernel dominates.
Proof of (iii). Both maps are continuous and strictly increasing. ROM reaches x* at L = sqrt(x*/alpha). Kernel reaches x* at L = x*/D*. The former is smaller iff alpha > D*^2/x*. In the interval between them, x_ROM > x* > x_ker, hence S_ROM > 1 > S_ker by strict monotonicity of S. QED.
Remark (generality). The phase separation and breaking condition hold for any bistable system with an unstable separatrix fed by quadratic vs. linear input maps from a common primitive. The proof requires bistability and the geometry of the maps, not the Beierspace interpretation. See the full appendix for the complete statement.
Corollary (Architectural Bifurcation). At fixed L where S_ker < 1, there exists unique alpha_bif where S_ROM crosses 1. For alpha > alpha_bif: ROM stable, kernel unstable. The bifurcation is in the architectural parameter.
Verified numerically with sigma_b(x) = 2/(1+x), sigma_e = 0.1, alpha = 1.2, D* = 1.0. Phase separation interval: L in (0.943, 1.067), width 0.124 (narrow at default parameterization; grows to 0.471 at alpha = 3.0). At midpoint L = 1.005: S_ROM = 1.208, S_ker = 0.916. mu_bar = 0.038 (50.5% of mu_crit). Under uncorrected dynamics: ROM converges to x_max = 5.0, kernel converges to 0. Verified across five functional forms in Class B' (power law, exponential, logistic, piecewise concave, concave square root); phase separation and opposite dynamical outcomes confirmed in all cases. mu_bar ranges from 13% to 55% of mu_crit across forms.
The proofs in the companion papers and in Appendix A were developed and verified using a structured protocol designed to prevent analysis guided by desired conclusions rather than by the structure of the equations. The protocol was itself a product of error discovery: several proofs in Beier (2026e) contained mistakes found only through systematic verification. The incorrect mu_crit formula was found by constructing a counterexample. The false separatrix claim was found by asking what the analysis revealed beyond what was expected. Each error had a common cause: reasoning toward a predetermined conclusion rather than letting the mathematics lead.
The protocol, applied to any dynamical system or formal claim:
(a) Dimension reduction. Check whether the effective dimension is lower than the formal dimension. Identify timescale separations. Analyze the reduced system first. (In the Coupling Advantage Theorem, the 2D Beierspace system reduces to effective 1D in the saturated limit, and the comparison between architectures operates entirely in L-space through two monotone maps.)
(b) Fixed points. Find all fixed points analytically. Classify their stability by computing eigenvalues or, for 1D systems, the sign of the derivative of the vector field at each zero. Do not assume the number of fixed points matches expectation.
(c) Phase portrait. Draw the phase portrait. Identify all basins of attraction and all separatrices. The phase portrait often reveals structure that the fixed-point analysis alone does not - transient dynamics, the geometry of basins, the proximity of the system's current state to a separatrix.
(d) Parameter sweep. Identify every free parameter. For each one, sweep it through its full range and identify all bifurcation values where the qualitative structure changes. The Coupling Advantage Theorem has a bifurcation in alpha (the architectural parameter) that was discovered through this sweep, not predicted from the theorem's initial conception.
(e) Counterexamples. For each qualitative claim, state the precise parameter conditions under which it holds. Then construct a counterexample showing it fails outside those conditions. The counterexample in part (v) of the Coupling Advantage Theorem - showing that the kernel outperforms at low L - was produced by this step. A theorem without a stated counterexample has not been fully characterized.
(f) Unrequested findings. Report what the analysis reveals that was not asked about. The unbounded stability gap (part ii), the architectural bifurcation corollary, and the breaking condition for D* exogeneity were all discovered at this stage. The most important results are often the ones the analyst did not expect.
Only after (a) through (f) are complete, verify numerically. Run simulations at least 10x longer than the timescale where the behavior appears to settle. If the answer changes between T and 10T, the analytical result is wrong - return to step (a). The Coupling Advantage Theorem was verified at T = 300 against apparent settlement at T = 30, with no change in outcomes.
This protocol is not a contribution to mathematical methodology - it is a systematic application of standard practices. Its value is in making the verification process explicit and reproducible, and in institutionalizing the habit of looking for what you did not expect to find.