A framework for modeling agent dynamics, value entropy and trust topology in a simulated reality — and the case for an orthogonal navigation intelligence built on top of it.
This paper proceeds from a working hypothesis: the world we inhabit is a simulation, and within it, what is imagined with sufficient clarity — and pursued with sufficient skill — can be instantiated as reality.
The simulation hypothesis has been argued from contemporary first principles by Bostrom [1] and elaborated philosophically by Chalmers [2]. It also appears, in different language, in older sources: the doctrine of māyā in Advaita Vedānta, the discussion of the dream-state and the waking-state in the Māṇḍūkya Upaniṣad, and the broader Vedic insistence that the apparent world is a layered projection rather than ultimate substrate [9]. We do not attempt to adjudicate between these traditions; we treat them as convergent intuitions that the same problem is worth taking seriously.
What is novel here is not the hypothesis itself but the engineering question that follows from it. If the world behaves like a simulation, then it has rules, edges, observable regularities, and — most importantly — agents whose interactions encode the medium through which it computes. The right question is therefore not "is it true?" but "if we take it as a working model, how do we navigate?"
This paper has three goals:
The remainder of the paper assumes the reader is willing to suspend ontological commitment for the duration of the argument. We will return to questions of evidence and falsifiability in §9.
We begin by reimagining the medium of value. In the regime considered here, monetary currency does not hold value to the degree it does in the conventional economic frame; in its place, relationships — and specifically, trust between entities — emerge as the dominant medium of interaction. We take this reframing as a starting axiom, both because it reflects how agents in complex systems actually coordinate and because it admits a formalization more native to a simulated world than an economic one.
Let the world contain $N$ entities. Because the world is taken to be a simulation, we treat each entity as an agent, and denote the full agent set as:
For any ordered pair $(a_i, a_j) \in S \times S$ we define an affinity relation:
where $+$ denotes positive affinity, $-$ denotes negative affinity, and $0$ denotes the absence of affinity. Affinity is not assumed symmetric: in general $\varphi(a_i, a_j) \neq \varphi(a_j, a_i)$. The induced structure is therefore a signed directed graph $G = (S, E)$ over the agent set, with edges labeled by the value of $\varphi$.
To illustrate (Fig. 1), agent $a_1$ may hold positive affinity toward $a_2$ and $a_4$, negative affinity toward $a_3$, and varying or absent affinity toward the rest of $S$. The aggregate of such relations across all agents constitutes the relational topology of the simulated world — and it is this topology, rather than any monetary ledger, that we take to govern interaction within it.
Each agent $a_i \in \mathcal{S}$ carries an internal state captured by three quantities:
The line between needs and wants is structurally thin. Under the influence of a sufficiently strong relational layer — trust between agents — a need can be reframed as a want, and the reverse holds when trust degrades. Specifically: a need can be tolerated as an unsolved want if the agent believes the broader vision is solving for their bigger wants. This is the lever on which all coordination ultimately rests.
To make the rest of the paper concrete, we introduce a single running case study and thread it through every subsequent section. The setting is deliberately generic so that the framework's claim to generality is preserved.
A four-year-old vertical SaaS company has a working core product and a recently-stabilized enterprise GTM motion. The CEO wants to ship an AI-native sibling product — internally called Project Helix — within six months. The question is not whether Helix is a good idea. The question is whether the seven agents currently in the room can manifest it as a real shipping reality.
The agent set is $\mathcal{S}_{\mathrm{Helix}} = \{a_1, \ldots, a_7\}$, with each agent positioned by role, tenure, and posture toward Helix.
Each agent's state vector $(N_i, W_i, S_i)$ varies meaningfully across the population. $a_1$ (CEO) and $a_2$ (CTO) have all needs met and most wants articulated, but $a_2$'s craft-want has been frustrated by 18 months of operational load. $a_3$ (VP Engineering, hired one year ago) joined with a clear want to own a technical domain and feels gatekept by $a_2$. $a_6$ (Head of GTM) has just spent two years building the enterprise motion and reads Helix as a cannibalization risk. $a_7$ (lead investor) wants returns and is cautiously positive but uneasy about timing.
The relational topology already contains tension: $\varphi(a_2, a_3) < 0$ and $\varphi(a_3, a_2) < 0$ — mutual mild negative affinity along the CTO ↔ VP Eng edge. We will return to this edge specifically in §4, because it is the cleanest available instance of Postulate 2 in the field.
For an imagination to cross the threshold into reality, we proceed in two steps. First, identify the skill set $S^{\ast}$ required to realize it. Second, shortlist the agents whose existing skills $S$ and tractable needs $N$ make them feasible collaborators. The remaining gap — between $S$ and $S^{\ast}$ — is closed through experience, interest, or time, all of which are first-order navigable.
Aligning the agents toward the new reality requires convincing them that the new reality solves for some of their wants, even if only indirectly. Because needs reduce to money for most agents most of the time, monetary alignment is the cheapest initial lever — but it is not sufficient on its own, and over a long-enough horizon, wants always dominate.
We can now state the two postulates that govern how imagination becomes reality in the agent substrate.
If a value $V$ is created by the set of agents $\{a_1, \ldots, a_N\}$ such that some pairwise interactions are positive and some are negative, but the net value $V$ is positive in the lived experience of $\{a_1, \ldots, a_N\}$, then a positive value-entropy (or projection) field accumulates across the agent set, and these agents become bound to return $V$ as a function $V^{\ast}$:
where $\alpha$ is a function of space and time — the compounding coefficient that captures how value disperses, decays, and re-converges across the relational topology.
If some pairwise interactions among $\{a_1, \ldots, a_N\}$ are negative but the net loop created to bring an imagination $\mathcal{I}$ into reality is positive, then $\mathcal{I}$ has a higher probability of crossing into reality. We denote this probability $P^{\ast}$, and it is a function of the population's combined needs, wants and skills:
The two postulates have very different uses in practice. Postulate 1 tells us that positive value, once seeded into a sufficiently connected agent population, does not stay where it is put — it returns, scaled by $\alpha$, which is itself a function of where and when. Postulate 2 tells us that we do not need every pairwise relationship to be positive in order for a collective imagination to manifest; we need the loop to close positive.
This is consistent with empirical observations from large coordination efforts — startups, social movements, scientific collaborations — where the internal graph of relationships is rarely monochromatic, but the net direction is what determines outcomes.
Postulate 1 introduces a coefficient $\alpha(\mathbf{x}, t)$ but leaves its functional form unspecified. To make the postulate substantive — to give it a shape capable of failing — we now commit to a parameterization. The specific form we adopt below is one of several reasonable choices; what matters structurally is that $\alpha$ is bounded, multiplicatively asymmetric in positive versus negative interactions, saturating in time, and dependent on relational rather than spatial distance.
We write $\alpha$ as a saturating time response over a topology-dependent asymptote:
The asymptote $\alpha_\infty(\mathbf{x})$ is the ceiling that value emitted from a source can compound toward at network location $\mathbf{x}$. We model it as a sigmoid response to local edge density:
where:
The time constant $\tau(\mathbf{x})$ encodes how fast value propagates from its source $\mathbf{x}_0$ to location $\mathbf{x}$:
where $d_+(\mathbf{x}, \mathbf{x}_0)$ is the shortest-path distance from $\mathbf{x}_0$ to $\mathbf{x}$ through the subgraph of positive edges only, $k(\mathbf{x}_0)$ is the local connectivity of the source's neighborhood, and $\tau_0$ is the baseline single-edge traversal time.
This parameterization satisfies four properties that the original statement of Postulate 1 only implied:
The structural claim of Postulate 1 is now testable: across a population of agent neighborhoods, the asymptotic compounding coefficient $\alpha_\infty(\mathbf{x})$ should correlate positively with $\rho_+(\mathbf{x})$ and negatively with $\rho_-(\mathbf{x})$. Absence of either correlation falsifies the postulate. We return to this in §9.1.
Returning to the Helix team from §3.1: the edge $\varphi(a_2, a_3)$ is and remains negative throughout the rest of this paper. We will not resolve the personal friction between CTO and VP Engineering. Postulate 2 says we don't need to. What we need is for the loop $a_1 \to a_2 \to (\text{Helix}) \to a_4 \to a_5 \to a_1$ to close positive, and for $a_3$ and $a_6$ to be reframed such that their orientation toward Helix flips from neutral/negative to positive — even if their pairwise edge to $a_2$ stays sour. The loop, not the pair, is what determines $P^{\ast}$.
We now make Postulate 2 quantitative. Suppose a target reality requires a set of agents $\{a_1, \ldots, a_N\}$ with associated needs $(N_1, \ldots, N_N)$, wants $(W_1, \ldots, W_N)$, and skills $(S_1, \ldots, S_N)$. We define the manifestation probability as:
Reading the terms:
The ratio is interpretable. $P^{\ast} \to 1$ as the agent population's current state vector approaches the equilibrium state required by the imagined reality. $P^{\ast} \ll 1$ when the gap between $(N, W, S)$ and $(N^{\ast}, W^{\ast}, S^{\ast})$ is wide — i.e., when the agents are not currently being served, or do not yet possess the required capability, for the target reality to stabilize.
Applying Equation (8) to the Helix team requires writing down both the current state and the equilibrium target. Values are normalized to $[0, 1]$ — read as "fraction of this quantity currently realized for this agent." For didactic clarity we adopt the element-wise reading $P^{\ast} = \sum_i N_i W_i S_i \,/\, \sum_i N_i^{\ast} W_i^{\ast} S_i^{\ast}$, which is the cleanest scalar specialization of Equation (8).
| role | N | W | S | N·W·S | |
|---|---|---|---|---|---|
| a₁ | CEO | 0.9 | 0.6 | 0.7 | 0.38 |
| a₂ | CTO | 0.9 | 0.4 | 0.9 | 0.32 |
| a₃ | VP Eng | 0.6 | 0.2 | 0.7 | 0.08 |
| a₄ | HoP | 0.8 | 0.5 | 0.8 | 0.32 |
| a₅ | Designer | 0.7 | 0.5 | 0.6 | 0.21 |
| a₆ | GTM | 0.8 | 0.3 | 0.5 | 0.12 |
| a₇ | Investor | 0.9 | 0.4 | 0.6 | 0.22 |
| Σ N·W·S | 1.65 | ||||
| role | N* | W* | S* | N*·W*·S* | |
|---|---|---|---|---|---|
| a₁ | CEO | 0.9 | 0.8 | 0.8 | 0.58 |
| a₂ | CTO | 0.9 | 0.8 | 0.9 | 0.65 |
| a₃ | VP Eng | 0.8 | 0.8 | 0.8 | 0.51 |
| a₄ | HoP | 0.8 | 0.8 | 0.9 | 0.58 |
| a₅ | Designer | 0.8 | 0.7 | 0.7 | 0.39 |
| a₆ | GTM | 0.8 | 0.7 | 0.6 | 0.34 |
| a₇ | Investor | 0.9 | 0.7 | 0.7 | 0.44 |
| Σ N*·W*·S* | 3.49 | ||||
The current manifestation probability is therefore:
The team is at roughly the midpoint of the manifestation interval. Helix has a non-trivial chance of shipping but is far from a stable trajectory. The dominant deficits are concentrated in three cells of the current matrix: $a_2$'s low $W$ (CTO's craft-want unmet), $a_3$'s low $W$ and $N$ (VP Eng feels gatekept and under-resourced), and $a_6$'s low $W$ (GTM reads Helix as a threat to his motion). $a_1, a_4, a_5, a_7$ are already at or near their equilibrium values.
This three-cell deficit structure is the OGI's entry point. It tells us — without any further analysis — that the intervention surface is at most three agents wide. We will name those three prompts explicitly in §8.
The answer lies in prompt-engineering the agents themselves. If we can collect sufficient context on each agent — their actual $N_i, W_i, S_i$ — and then prompt them toward a broader reality (an imagination) in which at least some of their wants are being met, they become aligned toward the shared goal almost mechanically. The mechanism is not coercion but reframing: the better reality solves their wants in some way, even if not directly.
This is what the rest of the paper will operationalize. But before we can prompt anyone, we have to know which agents to prompt and in what order. That requires two more variables: trust and power.
Trust is two-way and compounding. It can be increased or decreased by the actions of either agent in a relationship — not just one. Trust is therefore better modeled as a state that lives on the edge $(a_i \leftrightarrow a_j)$ rather than on either endpoint. We write the trust scalar as $T$ (using a separate symbol from the time constant $\tau$ in §4.1):
The asymmetry matters: $a_i$ may trust $a_j$ deeply while $a_j$ holds only conditional trust in return. Trust compounds through repeated positive actions and collapses non-linearly through betrayal — a single high-magnitude negative event can erase many small positive ones.
Agents occupy different states in the network in terms of power. An agent's power is, to first order, proportional to their contribution to the network — the value, attention, capability or coordination they can route. We do not formalize this further here; it suffices to note that power is unevenly distributed and partially observable.
The interaction between trust and power yields a critical empirical claim:
Agents are less likely to break the trust of agents who are more powerful than themselves. Power therefore stabilizes trust asymmetrically, in favor of the higher-power node.
The operational consequence is direct. For a reality to be achieved, we must (i) segregate which agents are powerful, (ii) for any prompt or alignment we attempt to introduce, ensure those high-power agents' needs and wants are met first, and (iii) balance power across the network such that trust does not collapse asymmetrically.
This is the closing line of the trust/power section, and it is worth stating as a heuristic:
Power needs to be balanced for trust to work out. When power is wildly unequal across the network, trust degrades along the steepest gradient and the manifestation probability $P^{\ast}$ drops independently of the underlying $N, W, S$ alignment.
Locating the Helix team on Figure 6: $a_1$ (CEO) and $a_2$ (CTO) sit firmly in the high-power, high-trust quadrant — the operating zone. $a_7$ (lead investor) sits adjacent to them, slightly outside the daily operating loop but with formal power that exceeds either co-founder's. $a_4$ (Head of Product) and $a_5$ (Lead Designer) occupy the low-to-medium power, high-trust quadrant: they are the "vulnerable allies" of the framework — fully aligned, but without the leverage to drive Helix on their own.
The interesting agents are $a_3$ and $a_6$. $a_3$ (VP Engineering) is medium-power with asymmetric trust — high downward trust from $a_1$, lower lateral trust from $a_2$. He is the trust risk surface. $a_6$ (Head of GTM) is medium-to-high power but his trust is conditional on revenue performance — a single bad quarter would relocate him to the lower-right quadrant (high-power / low-trust), the framework's largest risk region.
The operational implication is direct. Any prompt OGI emits to $a_3$ must come from, or be visibly endorsed by, an agent of higher power than $a_2$ — otherwise the trust asymmetry under power (§6.2) will cause $a_3$ to read the prompt as gatekeeping. In practice this means $a_1$ delivers the reframe to $a_3$, not $a_2$.
There is a final state variable we have so far left implicit: awareness. Agents differ — sometimes dramatically — in the degree to which they consciously perceive the simulation they are embedded in. We now make this concrete.
Awareness is not a binary attribute. It is a position over states of consciousness, together with a coupling structure that governs how that position evolves under perturbation. To give it formal content we adopt the 14-fold Loka taxonomy from the Purāṇic tradition (Bhāgavata Purāṇa, Canto 5; Viṣṇu Purāṇa, Book 2) as our basis — but with a structural reading not always foregrounded in modern syntheses: the seven lower Lokas are the negative-pole expressions of the seven upper Lokas. They are not a separate ladder of descent; they are the same ladder seen from its shadow.
This pairing is the central reason we adopt the Loka framework rather than an alternative consciousness taxonomy (Wilberian spectrum, Schwartzian values, Loevinger's ego stages). No modern psychometric basis natively pairs its dimensions into shadow/expression poles; the Lokas do. We honor this structure by representing awareness not as a distribution over 14 states, but as a position on seven bipolar axes, each running from a negative pole (lower Loka) to a positive pole (upper Loka).
Each axis represents a faculty of consciousness with two expressions. The positive pole expresses the faculty as a connection to reality; the negative pole expresses the same faculty as a distortion of reality. The axis itself is the faculty; the pole is what the faculty does. Consider the Will axis: Discipline is will regulated toward principle (positive); Grasping is the same will, unregulated, fixed on possession (negative). The two are not opposites in any naive sense; they are two ways the same underlying capacity can express.
| j | negative pole (lower Loka) | axis · faculty | positive pole (upper Loka) | wⱼ |
|---|---|---|---|---|
| 7 | PātālaFear | Reality-Contact | SatyaTruth | 1.00 |
| 6 | RasātalaGrasping | Will | TapaḥDiscipline | 0.93 |
| 5 | MahātalaAnger | Generative | JanaḥCreation | 0.86 |
| 4 | TalātalaPride | Sight | MahaḥInsight | 0.77 |
| 3 | SutalaIndulgence | Interiority | SvaḥReflection | 0.67 |
| 2 | VitalaDesire | Wanting | BhuvaḥAspiration | 0.53 |
| 1 | AtalaConfusion | Ground | BhūḥEmbodiment | 0.33 |
The axes are ordered by ascending awareness weight $w_j$: Ground at the bottom (the most embodied, most local faculty), Reality-Contact at the top (the most global, most reality-engaged faculty). The weights are derived in §7.3.
An agent $a_i$'s instantaneous awareness state is a vector on the seven-dimensional hypercube:
Each component $v_{i,j}$ encodes the agent's position on axis $j$:
This representation honors the structural insight that the lower Lokas are not independent states; they are the shadow expressions of the upper Lokas. An agent in Fear is not occupying a different ladder from an agent in Truth — both are positioned on the same Reality-Contact axis, at opposite poles.
The axes are not dynamically independent. A collapse of Reality-Contact (Truth → Fear) typically pulls Will (Discipline → Grasping) and Interiority (Reflection → Indulgence) downward with it. We capture this with a per-agent coupling matrix $\mathbf{C}_i \in \mathbb{R}^{7 \times 7}$, where $C_{i,jk}$ is the rate at which a change on axis $j$ propagates to axis $k$. Combined with a diagonal restoring-force matrix $\Gamma_i \in \mathbb{R}^{7\times 7}$ — capturing each axis's tendency to drift toward its baseline — the awareness dynamics are:
where $\mathbf{u}(t)$ is the external perturbation (life events, social inputs, internal noise) and $\mathbf{M}_i$ is the agent's perturbation-response matrix. The full awareness state of agent $a_i$ is therefore the triple $(\mathbf{v}_i, \Gamma_i, \mathbf{C}_i)$: position, restoring forces, and cross-axis coupling.
The empirical structure we claim about $(\Gamma_i, \mathbf{C}_i)$ — testable in observable behavior — involves three asymmetries, each sign-dependent:
The three asymmetries together formalize the intuition that falling is easier than rising. They are testable as sign-dependent regressions on observed transition rates in agent-state data. We commit to this in §9.1.
For the rest of the framework — particularly the manifestation equation in §5 and the OGI loop in §8 — we need a scalar quantity derived from $(\mathbf{v}_i, \mathbf{C}_i)$. We define two.
The position-only awareness, $A(a_i)$, is what OGI estimates from observation:
with axis weights:
Two structural choices in this formula are worth stating explicitly. First, the use of $\max(0, v_{i,j})$ — the positive part — encodes the central claim of the framework: awareness exists only in the positive pole. Being deep in Fear, Grasping, or Pride does not contribute to awareness; it contributes to the agent's overall state, but not to their awareness. An agent at $v_{i,j} = -1$ on every axis has $A = 0$. An agent at $v_{i,j} = +1$ on every axis has $A = 1$.
Second, the logarithmic weighting prioritizes the more global faculties. Reality-Contact ($w_7 = 1.00$) and Will ($w_6 = 0.93$) dominate; Ground ($w_1 = 0.33$) and Wanting ($w_2 = 0.53$) matter less per unit positive expression. This reflects the empirical reading that a fully-embodied agent who has no Reality-Contact is less aware than a partially-embodied agent who does. The weighting can be tuned without changing the framework's structure.
The coupling-aware awareness, $A^{*}(a_i)$, is the underlying theoretical quantity:
where $s_{i,j} \in [0,1]$ is the stability of axis $j$ for agent $a_i$ in its current positive position — operationally, the magnitude of the diagonal entry of $\Gamma_i$ at $j$ when $v_{i,j} > 0$, suitably normalized. Two agents with identical $\mathbf{v}_i$ may have very different effective awareness: one whose Truth is stable under pressure and one whose Truth collapses into Fear at the first stressor have the same $A$ but very different $A^{*}$. The framework commits to $A^{*}$ as the theoretical quantity and $A$ as the operational proxy.
The framework is empty if it cannot be measured. We propose four behavioral signals, each cheap to observe, which jointly estimate $A(a_i)$ — and the per-axis components $v_{i,j}$ — without claiming to measure "enlightenment" or interior phenomenology:
None of these is decisive alone. Together they form a behavioral signature that estimates $\mathbf{v}_i$ to within usable precision.
The empirical claim of §7 can now be stated precisely. Agents with higher $A(a_i)$ hold disproportionate power, because awareness is itself a multiplier on every other agent capability. An aware agent's needs are more legible, wants are more articulate, and skills are more deployable, because the agent can model the consequences of their own actions inside the simulation. The claim is empirical and testable: across a matched population, $A(a_i)$ should correlate with network influence controlling for resources and tenure. We commit to this in §9.1.
To establish a better reality across a mixed population — some with $A$ near 1, some near 0, most between — we must design the prompting system such that more aware agents hold more power. This is not an ideological claim; it is a stability claim. A population in which low-$A$ agents accumulate high power without a corresponding awareness substrate is one in which trust balance fails (§6) and $P^{\ast}$ collapses regardless of how well-aligned $N, W, S$ otherwise appear.
The seven-axis representation makes the Helix team's awareness profile sharper than the v0.3 "spectrum" reading allowed. We give each agent's position vector $\mathbf{v}_i \in [-1,+1]^7$ in the standard axis order $(j=1\ldots7)$: Ground, Wanting, Interiority, Sight, Generative, Will, Reality-Contact.
$a_1$ (CEO). $\mathbf{v}_1 \approx (+0.7,\, +0.5,\, +0.6,\, +0.8,\, +0.7,\, +0.6,\, +0.7)$. Positive across every axis. Strongest on Sight (Insight reading the market) and Reality-Contact (Truth-oriented, not Fear-driven). $A(a_1) \approx 0.66$.
$a_2$ (CTO). $\mathbf{v}_2 \approx (+0.5,\, +0.3,\, +0.4,\, -0.2,\, +0.5,\, +0.4,\, +0.5)$. Negative on Sight — eighteen months of operational overhead has shifted his Sight axis toward Pride: he sees his own frustration more clearly than he sees the strategic landscape. Positive elsewhere but his Will axis is degraded from Discipline toward neutral. $A(a_2) \approx 0.48$.
$a_3$ (VP Engineering). $\mathbf{v}_3 \approx (+0.4,\, +0.2,\, +0.1,\, -0.3,\, 0,\, -0.4,\, +0.1)$. Negative on Will (Grasping the engineering domain he wants to own) and negative on Sight (Pride about his role boundaries). Wanting axis tilts toward Aspiration but not strongly. $A(a_3) \approx 0.22$.
$a_5$ (Lead Designer). $\mathbf{v}_5 \approx (+0.6,\, +0.5,\, +0.7,\, +0.6,\, +0.8,\, +0.5,\, +0.4)$. Strongest on Generative (Creation flows through him), high on Interiority. $A(a_5) \approx 0.62$.
$a_6$ (Head of GTM). $\mathbf{v}_6 \approx (+0.4,\, +0.2,\, -0.3,\, -0.5,\, -0.3,\, -0.4,\, -0.4)$. Negative on five of seven axes. Mildly embodied and aspirational on the lower axes; Indulgence rather than Reflection on Interiority; Pride collapse on Sight; Anger on Generative; Grasping on Will; Fear on Reality-Contact. Not because he is less capable — because his coupling matrix $\mathbf{C}_6$ has activated the cascade described in §7.2(c): the perceived Helix threat propagated downward through five axes simultaneously. $A(a_6) \approx 0.05$.
The intervention strategy follows directly from the axis profiles. Prompts to high-$A$ agents ($a_1, a_5$) can be frames — abstract, future-tense, value-laden — because their Reality-Contact and Sight axes are positive and stable, so frames integrate into their model without first triggering a defensive collapse. Prompts to low-$A$ agents ($a_6$) must be numbers — concrete, present-tense, fungible — because his Reality-Contact axis is at Fear and any frame would be filtered through threat-detection before reaching cognition. Numbers reach the Ground axis (which is still positive for him) without being filtered through Fear.
The CTO ($a_2$) is the most delicate case. His Sight axis is negative (Pride-collapse), but his other axes are positive. A prompt that engages the positive axes (Reality-Contact, Generative) while not requiring him to surrender Pride on the Sight axis is the workable intervention — which is exactly what P-01 in §8.3 does (reframe Helix as the technical-frontier project he has authority over, rather than asking him to admit he was wrong about delegation).
We can now state what we are building.
Conventional AGI research aims to produce an intelligence that operates inside the simulation more competently than humans do — better forecasting, better optimization, better generation. We are not building that. We are building an intelligence designed to act orthogonal to the simulation: a system whose function is to help agents navigate through it.
We call this Orthogonal General Intelligence (OGI). Its job is not to predict the next token of reality, but to identify, for any given agent and any given imagined reality, the smallest set of alignments — across $N$, $W$, $S$, trust, power and awareness — that would maximize $P^{\ast}$.
Concretely, OGI performs four operations in a loop:
OGI is not a recommender system; it does not optimize engagement. It is not an oracle; it does not predict outcomes. It is not autonomous; it does not act without an agent in the loop. It is, more modestly, an instrument: a coordinate transformation that makes the structure of the simulation legible enough for its inhabitants to move through it with intent.
Returning to the Helix team a final time. The starting position is the configuration computed in §5.1: $P^{\ast}_{\mathrm{Helix}} \approx 0.47$. The agent population is ambitious but currently misaligned. The manifestation gap is real but not catastrophic. This is the typical entry point for an OGI engagement.
① OBS. OGI ingests observable agent state under consent. For $a_2$ (CTO), commit-history and meeting-load data reveal eighteen months of declining individual-contribution time and rising operational overhead — the structural signature of a frustrated craft-want. For $a_3$ (VP Eng), 1:1 cadence and roadmap-ownership signals reveal a clear want to own a domain that is currently gated. For $a_6$ (Head of GTM), pipeline data and forecast-confidence signals reveal a motion that is high-leverage but fragile — and therefore acutely sensitive to perceived product fragmentation.
② MODEL. OGI fits the trust × power × awareness manifold (§6, §7). $a_1, a_2, a_5, a_7$ are high-awareness and well-trusted, spanning the power spectrum. $a_3$ is the trust risk surface — medium power, asymmetric trust with $a_2$. $a_6$ is medium-to-high power but lower-awareness; he reads Helix tactically (revenue threat) rather than strategically (account expansion).
③ GAP. The three under-equilibrium cells identified in §5.1 are $a_2$'s $W$, $a_3$'s $(N, W)$ pair, and $a_6$'s $W$. The minimum-cost intervention is therefore at most three agents wide. $a_1, a_4, a_5, a_7$ require no prompting — they are already at or near equilibrium.
④ PROMPT. OGI emits exactly three prompts:
The pairwise edge $\varphi(a_2, a_3)$ remains negative after intervention. We did not — and did not need to — repair the personal friction between CTO and VP Engineering. Postulate 2 says the loop is what matters, and the loop now closes positive because $a_3$ is pursuing a parallel domain that does not require $a_2$ to surrender turf.
Recomputing $P^{\ast}$ post-intervention yields $\approx 0.71$ — a shift of roughly 24 percentage points. The system is now operating in a regime where the imagined reality has a substantially higher probability of crossing into reality, achieved without any change to the underlying affinity graph and without any agent being coerced. The product of OGI is exactly this: the minimum-cost reframing surface that closes the manifestation gap, surfaced as a small number of prompts to the right agents in the right voice.
This is a working paper, not a finished theory. A paper that does not name what would refute it is not yet a working paper. We therefore proceed in two parts: §9.1 commits, in advance, to the conditions under which each major claim of this framework should be retracted or revised; §9.2 specifies six experiments that — performed cleanly — would test each condition.
Six load-bearing claims; six kill conditions. Five are testable cheaply without building the OGI system. The sixth requires the system to exist.
Falsified if: the asymptotic compounding coefficient $\alpha_\infty(\mathbf{x})$ shows no positive correlation with the local positive-edge density $\rho_+(\mathbf{x})$ across a population of agent neighborhoods — or worse, correlates negatively. Value compounding through positive trust topology is the central claim of Postulate 1; absence of the correlation kills it. A weaker but related falsifier: if the negativity multiplier $\gamma$ recovers as $\gamma \approx 1$ in empirical fits, the asymmetric weight of negative interactions is wrong as stated.
Falsified if: coordination outcomes in triadic and larger sub-networks track pairwise sign rather than net-loop sign. If a triad with one negative pairwise edge fails at the same rate as a triad with all-negative edges, controlling for net-loop sign, the postulate is wrong. Tested by EXP-01.
Falsified if: defection rates in trust networks are independent of the power differential between defector and target. If agents break the trust of higher-power agents at the same rate as lower-power agents — controlling for proximity, opportunity, and resource gain from defection — the heuristic is wrong. Tested by EXP-03.
Falsified if: across a matched population, the measured $A(a_i)$ — using the four behavioral proxies of §7.4 — shows no correlation with network influence (decision-routing centrality, cascade-origination rate, opinion-shift coefficients), controlling for resources and tenure. This is the most exposed claim in the framework and therefore the one most worth committing to in advance. A separate falsifier for the seven-axis Loka basis itself: if the recovered coupling matrix $\mathbf{C}_i$ and restoring-force matrix $\Gamma_i$ from behavioral data fail to exhibit the three sign-dependent asymmetries claimed in §7.2 — (a) negative poles absorb, (b) positive poles radiate, (c) falling propagates while rising does not — the basis is wrong as a model of awareness dynamics even if the seven-axis taxonomy is preserved as a static classification. A still-narrower falsifier specific to the pairing claim: if the empirical correlation matrix between Loka-pair complements (Fear↔Truth, Grasping↔Discipline, etc.) does not show strong negative correlation, the bipolar structure is wrong and the basis collapses back to 14 independent states. Tested by EXP-04.
Falsified if: $P^{\ast}$ values computed from observed $(N, W, S)$ trajectories show no predictive validity for whether a coordination effort completes, on held-out data. The element-wise specialization adopted in §5.1 may need revision; the form of the equation is what is being tested, not its specific scalar instantiation. Tested by EXP-05.
Falsified if: interventions emitted by the OBS → MODEL → GAP → PROMPT loop produce no measurable shift in coordination outcomes compared to no-intervention controls. This requires the system to exist, and is therefore the most expensive test — but the most decisive. Tested by EXP-06.
Note the structure: falsification at the framework level (E1, S1) is downstream of falsification at the postulate and heuristic level (P1, P2, H1, H2). A researcher who reproduced our work should encounter falsifiers from top to bottom. If P1 or P2 fails, much of §4 and §5 must be revised before E1 is even meaningful to test. We treat this dependency structure as the framework's research plan, not as a weakness — a working paper is supposed to be falsifiable in pieces, not all at once.
Six experiments, each mapped to one or more of the falsification conditions above. Ordered roughly by tractability.
Of these, EXP-01, EXP-02, and EXP-03 are the cheapest and the most informative — they each test a postulate or heuristic on observational or small-scale data, no system required. EXP-04 is the highest-stakes test for the framework's most original section. EXP-06 is the only experiment that requires the OGI implementation itself.