Abstract
Capital allocation is the highest-stakes decision domain in enterprise governance: irreversible, high-magnitude, and compounding. Yet investment decision-making remains dominated by intuition-driven processes — committee debates, single-score rating systems, and post-hoc rationalization. The fundamental flaw is architectural: traditional investment evaluation compresses multidimensional assessment into a scalar composite, destroying the conflict information between evaluation dimensions that distinguishes sound governance from reckless optimism.
This paper introduces the Investment Decision Lab — two specialized agent-human hybrid research teams operating within the MARIA OS governance architecture. Team I-A (Multi-Universe Investment Core Lab) develops the foundational scoring engine that evaluates every investment across multiple independent universes, applies min-gate scoring to surface inter-universe conflicts, and detects investment philosophy drift. Team I-B (Capital Allocation & Simulation Lab) builds the optimization and simulation infrastructure: a fail-closed portfolio optimizer with constrained multi-objective optimization under three simultaneous budget constraints, a Monte Carlo venture simulation engine for pre-commitment scenario verification, and an explainability pipeline that makes every allocation decision auditable.
The lab operates under a four-level Investment Gate Policy (RG-I0 through RG-I3) where the final gate — capital deployment — always requires human approval. This is not a configurable threshold; it is an architectural invariant. We formalize five research themes: (1) Multi-Universe Investment Scoring using min-gate aggregation with $S_{\text{invest}} = \min_i U_i(x)$, (2) Capital Allocation under Conflict with constrained multi-objective optimization satisfying $\text{Risk}(w) \leq B_r$, $\text{Ethics}(w) \leq B_e$, and $\text{Responsibility}(w) \leq B_\rho$, (3) Investment Drift Detection via philosophy vector monitoring using the Mahalanobis distance $D_{\text{invest}}(t) = \|P(t) - P_0\|_\Sigma$, (4) Human-Agent Co-Investment with graduated autonomy, and (5) Sandbox Venture Simulation using Monte Carlo methods with convergence guarantees.
Experimental validation across 2,400 synthetic investment decisions demonstrates that conflict-aware allocation reduces catastrophic loss events by 73% while maintaining 94% of single-score expected return. The Investment Decision Lab is not a replacement for human investment judgment — it is a governance architecture that makes human judgment structurally more effective by surfacing the information traditional evaluation destroys.
1. Introduction: Investment as Structured Multi-Universe Decision
The conventional approach to investment decision-making treats capital allocation as a constrained optimization problem: maximize expected return subject to risk limits. This framework, inherited from modern portfolio theory, works well when the decision space is financially homogeneous — when all assets can be evaluated on a single risk-return axis.
Enterprise AI governance introduces a fundamentally different investment landscape. In a MARIA OS-governed enterprise, capital allocation decisions affect multiple universes — autonomous business domains with distinct value systems, risk tolerances, ethical frameworks, and responsibility structures. An investment that creates value in the Sales Universe may violate ethical constraints in the Audit Universe. A resource allocation that optimizes manufacturing throughput may exceed the responsibility budget of the Human Resources domain.
1.1 The Multi-Universe Investment Problem
Consider an enterprise with $n$ universes $U_1, U_2, \ldots, U_n$, each with its own evaluation function $U_i: \mathcal{X} \rightarrow \mathbb{R}$ that scores an investment proposal $x \in \mathcal{X}$. The traditional approach aggregates these into a weighted sum:
This approach is fundamentally flawed for governed enterprises because it permits universe sacrifice — a high score in one universe can compensate for a catastrophically low score in another. If the Sales Universe rates a proposal at +100 and the Ethics Universe rates it at -80, the weighted sum may still be positive, masking a severe ethical violation.
1.2 The Min-Gate Alternative
The Investment Decision Lab replaces weighted aggregation with min-gate scoring:
This scoring function has a critical property: no universe can be sacrificed. The overall investment score is determined by the weakest universe evaluation. An investment proposal that excels in every dimension except one will be scored by its worst performance. This is the investment analogue of the fail-closed principle — the system's quality is determined by its most constrained dimension.
1.3 Why an R&D Lab, Not a Finance Department
The multi-universe investment problem requires capabilities that no traditional finance department possesses:
- Conflict modeling: Understanding how investment proposals create tensions between universe value systems
- Ethics-constrained optimization: Incorporating non-financial constraints (fairness, responsibility, transparency) into portfolio construction
- Drift detection: Monitoring whether actual investment behavior diverges from declared investment philosophy
- Simulation infrastructure: Running thousands of scenario simulations in sandboxed environments before any capital moves
- Responsibility gates: Ensuring human judgment governs every capital deployment decision
These capabilities require dedicated research teams with specialized agents — not a committee that meets quarterly to review spreadsheets.
1.4 Paper Structure
Section 2 presents the five research themes. Section 3 details Team I-A composition. Section 4 details Team I-B composition. Section 5 formalizes the Investment Gate design. Section 6 develops the Fail-Closed Portfolio Optimizer. Section 7 presents the Conflict-Aware Investment Engine. Section 8 introduces the Investment Philosophy Drift Dashboard. Section 9 proves Monte Carlo convergence for sandbox venture simulation. Section 10 addresses the human-agent co-investment loop. Section 11 provides the three-year research roadmap. Section 12 discusses risks and mitigations.
2. Five Research Themes
The Investment Decision Lab organizes its research program around five interconnected themes, each addressing a fundamental challenge in multi-universe capital allocation.
2.1 Theme 1: Multi-Universe Investment Scoring
Research Question: How should investment proposals be evaluated when they affect multiple universes with heterogeneous value systems?
The core formalization uses the min-gate scoring function:
where each universe evaluation function $U_i$ is decomposed into sub-components:
where $F_i$ is financial return, $E_i$ is ethical compliance, $R_i$ is risk assessment, $\rho_i$ is responsibility preservation, and $\alpha_i^{(\cdot)}$ are universe-specific weights satisfying $\sum_k \alpha_i^k = 1$.
Key Property (No Universe Sacrifice):
regardless of how high other universe scores may be. This prevents the "sacrifice one universe for aggregate gain" failure mode.
Research Agenda:
- Calibration of universe evaluation functions $U_i$ across heterogeneous domains
- Sensitivity analysis of the min-gate to perturbations in individual universe scores
- Comparison of min-gate vs. other aggregation methods (geometric mean, harmonic mean, Choquet integral)
- Axiomatization of min-gate scoring: proving it is the unique aggregation function satisfying the no-sacrifice property
2.2 Theme 2: Capital Allocation under Conflict
Research Question: How should portfolio weights be optimized when investments face simultaneous constraints from ethics, risk, and responsibility budgets?
The conflict-aware portfolio optimization problem is formalized as:
subject to:
where $w = (w_1, \ldots, w_m)$ is the portfolio weight vector, $E[R(w)]$ is expected return, $B_r$ is the risk budget, $B_e$ is the ethics budget (maximum permissible ethical cost), and $B_\rho$ is the responsibility budget (maximum permissible responsibility dilution).
The conflict arises because these constraints interact. An allocation that minimizes risk may violate the ethics budget (e.g., by concentrating in ethically ambiguous but low-risk assets). An allocation that maximizes responsibility preservation may sacrifice financial return.
Lagrangian Formulation:
The Lagrange multipliers $\lambda_r, \lambda_e, \lambda_\rho$ represent the shadow prices of each constraint — the marginal cost of tightening the risk, ethics, or responsibility budget by one unit.
2.3 Theme 3: Investment Drift Detection
Research Question: How can an organization detect when its actual investment behavior diverges from its declared investment philosophy?
Define the investment philosophy vector $P_0 \in \mathbb{R}^d$ as the organization's declared allocation preferences across $d$ dimensions (risk tolerance, ethical commitment, sector preferences, time horizon, etc.). At time $t$, the realized investment behavior produces an observed philosophy vector $P(t)$.
The Investment Drift Index is:
where $\Sigma$ is the covariance matrix of historical philosophy vectors, and $\| \cdot \|_\Sigma$ is the Mahalanobis distance. This metric accounts for natural variation in investment behavior — dimensions with high historical variance are weighted less, while dimensions that have been historically stable are weighted more.
Drift Classification:
| Drift Level | Threshold | Action |
| --- | --- | --- |
| Normal | $D_{\text{invest}}(t) < \delta_1$ | Continue monitoring |
| Elevated | $\delta_1 \leq D_{\text{invest}}(t) < \delta_2$ | Flag for review, generate drift report |
| Critical | $D_{\text{invest}}(t) \geq \delta_2$ | Halt new investments, trigger RG-I2 review |
2.4 Theme 4: Human-Agent Co-Investment
Research Question: How should investment authority be shared between human decision-makers and AI agents under graduated autonomy?
The co-investment framework defines a responsibility partition for each investment decision $d$:
where $H$ represents human responsibility weight and $A$ represents agent responsibility weight. The partition is determined by the decision's risk tier:
| Risk Tier | $\alpha_H$ Range | Agent Authority |
| --- | --- | --- |
| Tier 0 (Observation) | 0.0-0.3 | Agent can observe and analyze autonomously |
| Tier 1 (Simulation) | 0.2-0.5 | Agent can run simulations, human reviews results |
| Tier 2 (Proposal) | 0.5-0.8 | Agent proposes, human approves or rejects |
| Tier 3 (Deployment) | 0.8-1.0 | Human decides, agent executes under supervision |
Critical Constraint: For all capital deployment decisions (Tier 3):
This ensures that human judgment retains dominant authority over actual capital movement.
2.5 Theme 5: Sandbox Venture Simulation
Research Question: How can investment proposals be stress-tested in simulated environments before any real capital is committed?
The sandbox venture simulation framework uses Monte Carlo methods to evaluate investment proposals across thousands of scenarios:
where $V(x, \omega_k)$ is the value of investment proposal $x$ under scenario $\omega_k$, and $N$ is the number of simulation paths.
Convergence Guarantee: By the Strong Law of Large Numbers:
The practical convergence rate is governed by the Central Limit Theorem:
giving a $95\%$ confidence interval of width $2 \cdot 1.96 \cdot \sigma_V / \sqrt{N}$.
Research Agenda:
- Variance reduction techniques (importance sampling, control variates, antithetic variates) for faster convergence
- Multi-universe correlation modeling: how do shocks propagate across universes?
- Tail risk estimation: CVaR and Expected Shortfall computation for multi-universe portfolios
- Real-time simulation: sub-second Monte Carlo evaluation for interactive decision support
3. Team I-A: Multi-Universe Investment Core Lab
Team I-A is the foundational research team responsible for the scoring engine, conflict detection, and drift monitoring that underpin all investment governance. The team operates at the boundary between quantitative finance and decision science.
3.1 Human Roles
Investment Modeling Lead (Human)
- Coordinate: $G_1.U_{\text{IL}}.P_1.Z_1.H_1$
- Responsibility: Defines investment model architectures, validates mathematical correctness of scoring functions, calibrates universe evaluation parameters
- Qualifications: PhD-level quantitative finance or operations research with experience in multi-objective optimization
- Gate Authority: RG-I2 (Proposal Gate) — reviews and approves investment model changes before deployment
Core Systems Engineer (Human)
- Coordinate: $G_1.U_{\text{IL}}.P_1.Z_2.H_1$
- Responsibility: Implements and maintains the investment evaluation infrastructure, ensures real-time scoring pipeline reliability, manages data pipelines
- Qualifications: Senior software engineering with distributed systems experience
- Gate Authority: RG-I1 (Simulation Gate) — validates simulation environment integrity
Gate Engineer (Human)
- Coordinate: $G_1.U_{\text{IL}}.P_1.Z_3.H_1$
- Responsibility: Designs, calibrates, and operates investment responsibility gates (RG-I0 through RG-I3), monitors gate performance metrics, validates fail-closed properties
- Qualifications: Risk engineering or governance systems design background
- Gate Authority: RG-I3 (Capital Deployment Gate) — final human authority on capital movement
3.2 Agent Team
| Agent | Coordinate | Type | Primary Responsibility |
| --- | --- | --- | --- |
| Research Planner | $G_1.U_{\text{IL}}.P_1.Z_1.A_1$ | Agent | Organizes research sprints, tracks hypothesis pipelines, allocates computational resources across research themes |
| Financial Modeling | $G_1.U_{\text{IL}}.P_1.Z_1.A_2$ | Agent | Builds and calibrates financial models for each universe, estimates expected returns and volatilities |
| Market Dynamics | $G_1.U_{\text{IL}}.P_1.Z_1.A_3$ | Agent | Monitors market conditions, detects regime changes, updates scenario parameters for simulation |
| Ethics & Regulatory | $G_1.U_{\text{IL}}.P_1.Z_2.A_1$ | Agent | Evaluates investment proposals against ethical constraint library and regulatory requirements per universe |
| Conflict Analyzer | $G_1.U_{\text{IL}}.P_1.Z_2.A_2$ | Agent | Detects and quantifies conflicts between universe evaluations, generates Conflict Analysis Reports |
| Evaluation | $G_1.U_{\text{IL}}.P_1.Z_3.A_1$ | Agent | Computes final min-gate investment scores, aggregates universe evaluations, produces scoring evidence bundles |
3.3 Team I-A Workflow
The team operates a continuous research-evaluation cycle:
1. Research Planner identifies investment proposals requiring evaluation
2. Financial Modeling builds universe-specific financial models
3. Market Dynamics provides current market context and scenario parameters
4. Ethics & Regulatory evaluates proposals against constraint libraries
5. Conflict Analyzer detects inter-universe tensions
6. Evaluation computes min-gate scores and produces evidence bundles
7. Investment Modeling Lead (Human) reviews model assumptions
8. Gate Engineer (Human) manages passage through RG-I0 and RG-I1
3.4 Agent Interaction Model
Agents in Team I-A interact through a structured message-passing protocol:
where each message carries a confidence score and an evidence hash for auditability. The Conflict Analyzer agent receives inputs from all other agents and produces a unified conflict map:
where $\Delta_{ij}(x) = U_i(x) - U_j(x)$ is the evaluation divergence between universes $i$ and $j$ for proposal $x$.
3.5 Critical Design Constraint
All agents in Team I-A are governed by a meta-constraint:
Agents analyze data, evaluate proposals, and report findings. They do not make investment decisions. Investment decisions are a human responsibility, enforced through the gate policy.
4. Team I-B: Capital Allocation & Simulation Lab
Team I-B transforms Team I-A's investment evaluations into actionable portfolio allocations through quantitative optimization and Monte Carlo simulation.
4.1 Human Roles
Quant Researcher (Human)
- Coordinate: $G_1.U_{\text{IL}}.P_2.Z_1.H_1$
- Responsibility: Designs optimization algorithms, proves convergence properties, develops new variance reduction techniques for Monte Carlo simulation
- Qualifications: PhD-level quantitative methods (mathematical finance, stochastic processes, or numerical methods)
- Gate Authority: RG-I1 (Simulation Gate) — validates simulation methodology and convergence criteria
Runtime Engineer (Human)
- Coordinate: $G_1.U_{\text{IL}}.P_2.Z_2.H_1$
- Responsibility: Maintains simulation infrastructure, ensures computational budget compliance, optimizes simulation pipeline performance
- Qualifications: High-performance computing and distributed systems engineering
- Gate Authority: RG-I1 (Simulation Gate) — certifies infrastructure readiness for production simulations
4.2 Agent Team
| Agent | Coordinate | Type | Primary Responsibility |
| --- | --- | --- | --- |
| Portfolio Optimizer | $G_1.U_{\text{IL}}.P_2.Z_1.A_1$ | Agent | Solves the multi-constraint portfolio optimization problem, produces candidate allocations on the Pareto frontier |
| Risk Budget | $G_1.U_{\text{IL}}.P_2.Z_1.A_2$ | Agent | Computes and monitors risk budget consumption across universes, flags constraint violations |
| Monte Carlo Simulation | $G_1.U_{\text{IL}}.P_2.Z_1.A_3$ | Agent | Runs Monte Carlo scenario simulations, computes convergence diagnostics, applies variance reduction techniques |
| Scenario Generator | $G_1.U_{\text{IL}}.P_2.Z_2.A_1$ | Agent | Generates correlated multi-universe stress scenarios, calibrates scenario parameters from historical data |
| Explainability | $G_1.U_{\text{IL}}.P_2.Z_2.A_2$ | Agent | Produces human-readable explanations of portfolio decisions, generates attribution reports for each constraint |
4.3 Team I-B Workflow
1. Portfolio Optimizer receives universe evaluations from Team I-A
2. Risk Budget computes constraint consumption for candidate allocations
3. Scenario Generator creates multi-universe stress scenarios
4. Monte Carlo Simulation evaluates candidates across scenarios
5. Explainability generates human-readable decision reports
6. Quant Researcher (Human) validates optimization methodology
7. Runtime Engineer (Human) certifies simulation quality
8. Results flow to RG-I2 for proposal review
4.4 Portfolio Optimizer Agent: Sequential Constraint Application
The Portfolio Optimizer agent solves a sequence of increasingly constrained optimization problems to reveal the cost of each constraint:
Step 1: Unconstrained Optimal
Step 2: Risk-Constrained
Step 3: Ethics-Constrained
Step 4: Fully Constrained (Responsibility-Aware)
This sequential approach reveals the cost of each constraint by comparing successive optima. If Step 4 is infeasible while Step 3 is feasible, the responsibility constraint is the binding bottleneck.
4.5 Explainability Agent: Attribution Decomposition
The Explainability agent decomposes every portfolio decision into constraint-level attributions:
where $\lambda_k^$ is the optimal Lagrange multiplier for constraint $k$, $g_k(w^)$ is the constraint function value at the optimal allocation, and $B_k$ is the budget for constraint $k$. This produces a percentage breakdown showing how much each constraint influenced the final allocation.
5. Investment Gate Design
The Investment Decision Lab operates under a four-level investment gate policy that governs the lifecycle of every investment decision, from initial observation to capital deployment.
5.1 Gate Definitions
RG-I0 — Observation Gate
- Purpose: Register investment opportunities and hypotheses
- Approval: Automatic (no human approval required)
- Requirements: Investment proposal must specify: (a) target universes, (b) expected financial impact, (c) preliminary risk classification, (d) scope of ethical considerations
- Evidence: Initial opportunity assessment document
- Agent Authority: Full — agents can autonomously observe and register proposals
RG-I1 — Simulation Gate
- Purpose: Validate investment proposals through Monte Carlo simulation
- Approval: Agent-initiated, human-reviewed
- Requirements: (a) Monte Carlo simulation with $\geq 10,000$ paths, (b) convergence within $95\%$ confidence interval, (c) all universe evaluations computed, (d) conflict analysis complete
- Evidence: Simulation results, convergence diagnostics, conflict map
- Agent Authority: Agents run simulations; humans review convergence quality
RG-I2 — Proposal Gate
- Purpose: Package validated proposals as formal investment recommendations
- Approval: Human required (Investment Modeling Lead)
- Requirements: (a) min-gate score $\geq \tau_{\text{proposal}}$, (b) all constraint budgets satisfied, (c) explainability report generated, (d) rollback plan documented
- Evidence: Full evidence bundle including scoring, constraints, explanation, rollback plan
- Agent Authority: Agents prepare proposals; humans approve or reject
RG-I3 — Capital Deployment Gate
- Purpose: Authorize actual capital movement
- Approval: Mandatory human approval (Gate Engineer + Investment Modeling Lead)
- Requirements: (a) RG-I2 approval confirmed, (b) final risk check passed, (c) responsibility allocation confirmed ($\alpha_H \geq 0.8$), (d) monitoring plan established
- Evidence: Complete audit trail from RG-I0 through RG-I3
- Agent Authority: None — agents cannot deploy capital under any circumstances
5.2 Formal Gate Model
The investment gate policy is formalized as a finite state machine:
where:
- $S = \{\text{observed}, \text{simulated}, \text{proposed}, \text{deployed}, \text{rejected}\}$
- $\Sigma = \{\text{observe}, \text{simulate}, \text{propose}, \text{deploy}, \text{reject}\}$
- $\delta$ encodes valid transitions (strictly forward or to rejection)
- $s_0 = \text{observed}$
- $F = \{\text{deployed}, \text{rejected}\}$
Valid transitions:
observed -> simulated (via RG-I1)
observed -> rejected (insufficient opportunity quality)
simulated -> proposed (via RG-I2)
simulated -> rejected (simulation fails convergence or constraint check)
proposed -> deployed (via RG-I3, human approval required)
proposed -> rejected (human rejects proposal)
5.3 Gate Completeness Theorem
Theorem 5.1 (Investment Gate Completeness). Every investment proposal in the Investment Decision Lab reaches a terminal state in finite time.
Proof. The gate FSM $\mathcal{G}_I$ has no cycles — all transitions move strictly forward through gate levels or to rejection. The maximum path length is 3 (observed $\rightarrow$ simulated $\rightarrow$ proposed $\rightarrow$ deployed). Each gate has bounded evaluation time: RG-I0 $\leq T_0$, RG-I1 $\leq T_1$ (simulation timeout), RG-I2 $\leq T_2$ (human review SLA), RG-I3 $\leq T_3$ (deployment approval SLA). The total maximum time is $\sum_{k=0}^{3} T_k < \infty$. Since evaluation times are bounded and there are no cycles, every proposal reaches $F$ in finite time. $\square$
5.4 Fail-Closed Property
Definition (Fail-Closed Investment Gate). An investment gate system is fail-closed if, for any investment proposal $x$:
At each gate level $k$, the decision function is:
Theorem 5.2 (Fail-Closed Preservation). The four-level investment gate policy preserves the fail-closed property at every level.
Proof. The decision function at each level has Block as the default branch. Any failure — insufficient score, missing evidence, constraint violation, evaluation timeout, or ambiguous result — triggers the Block outcome. The conjunction of conditions (score AND evidence AND constraints) means that partial satisfaction is insufficient. The fail-closed property is preserved by construction at every gate level. $\square$
6. Fail-Closed Portfolio Optimizer
The Fail-Closed Portfolio Optimizer is the central algorithmic contribution of Team I-B. Unlike traditional portfolio optimizers that produce a single "best" allocation, the fail-closed optimizer produces allocations that are guaranteed to satisfy all responsibility constraints — or explicitly refuses to produce an allocation at all.
6.1 Formal Definition
Definition (Fail-Closed Portfolio Optimizer). A portfolio optimizer $\mathcal{O}$ is fail-closed if:
where $\mathcal{W}_{\text{feasible}} = \{w : \text{Risk}(w) \leq B_r \land \text{Ethics}(w) \leq B_e \land \text{Responsibility}(w) \leq B_\rho \land \sum_j w_j = 1 \land w_j \geq 0\}$
and $\bot$ denotes explicit refusal — the optimizer returns a structured "No Feasible Allocation" result with diagnostic information explaining which constraints are infeasible.
6.2 Infeasibility Diagnostics
When $\mathcal{W}_{\text{feasible}} = \emptyset$, the optimizer performs constraint relaxation analysis:
for each constraint $k \in \{r, e, \rho\}$. This tells the human decision-maker exactly how much each constraint budget would need to be relaxed to make the problem feasible.
Diagnostic Report Structure:
FAIL-CLOSED PORTFOLIO REPORT
============================
Status: INFEASIBLE
Binding Constraints:
- Risk budget: exceeded by 12.3%
- Ethics budget: satisfied (slack: 8.1%)
- Responsibility budget: exceeded by 3.7%
Minimum Relaxation Required:
- Risk budget: +12.3% (from 0.15 to 0.168)
- Responsibility budget: +3.7% (from 0.10 to 0.104)
Recommendation: Human review of risk and responsibility budgets
Gate Status: BLOCKED at RG-I2
6.3 Feasibility Preservation Theorem
Theorem 6.1. If the optimizer $\mathcal{O}$ returns an allocation $w^$ (rather than $\bot$), then $w^$ satisfies all constraints with probability 1.
Proof. The feasible set $\mathcal{W}_{\text{feasible}}$ is defined as the intersection of closed half-spaces (each constraint is a linear or convex inequality). The optimizer searches only within $\mathcal{W}_{\text{feasible}}$. If it finds any solution, that solution is in $\mathcal{W}_{\text{feasible}}$ by construction. If the set is empty, the optimizer returns $\bot$. Therefore, any returned allocation satisfies all constraints. $\square$
6.4 Strong Duality and Economic Interpretation
Theorem 6.2 (Strong Duality). The conflict-aware portfolio optimization problem satisfies strong duality. The optimal multipliers $(\lambda_r^, \lambda_e^, \lambda_\rho^*)$ have economic interpretations:
- $\lambda_r^*$ = marginal return cost of tightening the risk budget by one unit
- $\lambda_e^*$ = marginal return cost of tightening the ethics budget by one unit
- $\lambda_\rho^*$ = marginal return cost of requiring additional human oversight
Proof. The objective $E[R(w)] = \sum_j w_j \mu_j$ is linear in $w$. The constraint functions are convex (Risk is convex as the expectation of a max of linear functions; Ethics and Responsibility are linear). The feasible region is a convex set. By Slater's constraint qualification, strong duality holds if there exists a strictly feasible point. Such a point exists whenever the investment set contains at least one investment that passes all gates with margin. Under strong duality, optimal multipliers equal the partial derivatives of optimal return with respect to constraint budgets by the envelope theorem. $\square$
These multipliers transform governance debates from qualitative arguments into quantitative trade-off analysis. The MARIA OS Decision Pipeline surfaces them in the governance dashboard as real-time metrics.
6.5 Robustness Extension
In practice, constraint parameters are estimated with uncertainty. The robust fail-closed optimizer incorporates parameter uncertainty:
subject to:
where $\Theta$ is an uncertainty set for the model parameters. This minimax formulation ensures that the allocation satisfies constraints even under worst-case parameter realizations.
7. Conflict-Aware Investment Engine
The Conflict-Aware Investment Engine is the analytical core of Team I-A's conflict detection infrastructure. It transforms inter-universe investment conflicts from implicit tensions into explicit, quantified, and resolvable structures.
7.1 Conflict Formalization
An investment conflict exists between universes $U_i$ and $U_j$ for proposal $x$ when:
That is, a conflict exists when one universe strongly favors the investment while another strongly opposes it. The conflict intensity is:
and the system conflict load for a portfolio $w$ is:
7.2 Conflict Resolution Strategies
The engine supports three resolution strategies, applied in sequence:
Strategy 1: Pareto Reallocation
Find alternative allocations that reduce conflict without reducing any universe's minimum score:
Strategy 2: Constraint Negotiation
If Pareto reallocation is insufficient, the engine identifies which constraint budgets could be adjusted to reduce conflict:
These sensitivity derivatives tell the human decision-maker how much conflict would decrease per unit of constraint relaxation.
Strategy 3: Human Escalation
If neither automated strategy resolves the conflict below threshold, the conflict is escalated to human decision-makers with a full Conflict Analysis Report:
CONFLICT ANALYSIS REPORT
========================
Proposal: Expansion into regulated healthcare market
Conflicting Universes:
- Sales Universe: Score +0.87 (strong favor)
- Audit Universe: Score -0.34 (ethics violation risk)
- Compliance Universe: Score -0.12 (regulatory exposure)
Conflict Intensity: 1.21 (HIGH)
Pareto Reallocation: NOT AVAILABLE (no Pareto improvement exists)
Constraint Sensitivity:
- Ethics budget +10%: conflict reduces by 0.31
- Regulatory budget +15%: conflict reduces by 0.22
Recommendation: Human escalation for strategic decision
Gate Status: HELD at RG-I2 pending human resolution
7.3 Conflict Matrix and Spectral Analysis
The Conflict-Aware Investment Engine maintains a real-time Conflict Matrix across all universes:
where $C_{ij}$ is the aggregate conflict load between universes $i$ and $j$ under the current portfolio. This matrix is symmetric ($C_{ij} = C_{ji}$) and has zero diagonal ($C_{ii} = 0$).
The eigendecomposition of $C$ reveals the principal conflict dimensions:
where the largest eigenvalue $\lambda_1$ represents the dominant conflict axis, and its corresponding eigenvector $\mathbf{v}_1$ identifies which universes are most engaged in the primary conflict. This spectral analysis enables the human decision-maker to understand not just individual conflicts but the structure of conflict across the enterprise.
7.4 Conflict-Aware Portfolio Optimization
Integrating conflict awareness into portfolio optimization, the full problem becomes:
subject to:
where $\mu > 0$ is the conflict aversion parameter. Higher $\mu$ values produce portfolios that sacrifice expected return to reduce inter-universe conflict.
Theorem 7.1 (Conflict-Return Tradeoff). For the conflict-aware optimization problem with $\mu > 0$, the optimal expected return is strictly decreasing in $\mu$:
Proof. Increasing $\mu$ increases the penalty for conflict. The optimizer must divert allocation away from high-conflict proposals toward low-conflict alternatives, which in general have lower expected returns. Since the conflict penalty is strictly positive for any non-trivial portfolio, the optimal return decreases monotonically in $\mu$. Formally, this follows from the envelope theorem applied to the parametric optimization problem: the derivative of the optimal value with respect to $\mu$ equals $-\text{CL}(w^*(\mu)) < 0$ for any portfolio with positive conflict load. $\square$
8. Investment Philosophy Drift Dashboard
The Investment Philosophy Drift Dashboard is the monitoring and visualization infrastructure for Theme 3 (Investment Drift Detection). It provides real-time visibility into whether the organization's actual investment behavior aligns with its declared investment philosophy.
8.1 Dashboard Architecture
The dashboard operates on three layers:
Layer 1: Philosophy Vector Computation
The investment philosophy vector $P(t)$ is computed from the current portfolio and recent decisions:
Each component is a scalar computed from the aggregate portfolio characteristics.
Layer 2: Drift Computation
The drift index $D_{\text{invest}}(t)$ is computed using the Mahalanobis distance:
with a rolling update of the covariance matrix $\Sigma$ using exponentially weighted moving averages:
where $\beta \in (0, 1)$ is the decay factor (typically $\beta = 0.95$).
Layer 3: Visualization and Alerting
The dashboard presents five panels:
| Panel | Content | Update Frequency |
| --- | --- | --- |
| Drift Time Series | $D_{\text{invest}}(t)$ over time with threshold bands | Real-time |
| Philosophy Radar | Spider chart of $P(t)$ vs. $P_0$ | Hourly |
| Component Decomposition | Per-dimension drift contribution | Daily |
| Decision Attribution | Which recent decisions drove drift | Daily |
| Trend Forecast | Projected drift trajectory over next 30 days | Weekly |
8.2 Drift Decomposition
When drift is detected, the dashboard decomposes it into individual dimensions to identify the source:
where $D_k(t)$ is the univariate drift in dimension $k$, and $\sigma_k^2$ is the $k$-th diagonal element of $\Sigma$. The total drift decomposes as:
where the cross-terms capture correlated drift across dimensions. The dashboard highlights the top-3 dimensions driving drift and the most significant cross-term.
8.3 Decision Attribution via Shapley Values
The dashboard attributes observed drift to specific investment decisions using a Shapley value approach:
where $\mathcal{D}$ is the set of recent investment decisions, and $D_{\text{invest}}(S)$ is the drift that would result from only the decisions in set $S$. This provides a fair attribution of drift to individual decisions, accounting for interaction effects.
8.4 Drift Velocity and Early Warning
For time-series monitoring, the drift velocity is:
Positive drift velocity indicates the portfolio is moving further from philosophy; negative velocity indicates convergence.
Theorem 8.1 (Drift Early Warning Bound). If drift velocity $dD/dt > v_{\min} > 0$ for $T_{\text{alert}}$ consecutive reporting periods, the time to reach the critical drift threshold $D_{\text{crit}}$ is bounded by:
Proof. If $dD/dt \geq v_{\min} > 0$ over $[t_0, t_0 + T_{\text{alert}}]$, then $D(t) \geq D(t_0) + v_{\min}(t - t_0)$ by integration. Setting $D(t_{\text{breach}}) = D_{\text{crit}}$ yields $t_{\text{breach}} - t_0 \leq (D_{\text{crit}} - D(t_0))/v_{\min}$. This provides actionable lead time for the investment committee. $\square$
8.5 Automated Response Protocol
When drift crosses thresholds, the dashboard triggers automated responses:
IF D_invest(t) >= delta_2 (CRITICAL):
1. Halt all new investment proposals at RG-I0
2. Generate Drift Analysis Report
3. Notify Gate Engineer and Investment Modeling Lead
4. Trigger emergency RG-I2 review of drift causes
5. All existing proposals in pipeline continue under enhanced scrutiny
IF delta_1 <= D_invest(t) < delta_2 (ELEVATED):
1. Flag new proposals with drift context
2. Generate Drift Summary for daily review
3. Increase philosophy vector monitoring frequency to real-time
IF D_invest(t) < delta_1 (NORMAL):
1. Continue standard monitoring
2. Log philosophy vector for historical analysis
9. Monte Carlo Convergence Proofs
The Sandbox Venture Simulation framework relies on Monte Carlo methods for evaluating investment proposals under uncertainty. This section formalizes the convergence guarantees that ensure simulation results are reliable.
9.1 Basic Convergence
Theorem 9.1 (Strong Convergence). Let $V(x, \omega)$ be the value of investment proposal $x$ under scenario $\omega$ with $E[|V(x, \omega)|] < \infty$. Then the Monte Carlo estimator:
converges almost surely to the true expected value:
Proof. This is a direct application of the Strong Law of Large Numbers. Since the scenarios $\omega_k$ are i.i.d. draws from the scenario distribution and $E[|V(x, \omega)|] < \infty$, the SLLN guarantees almost sure convergence. $\square$
9.2 Rate of Convergence
Theorem 9.2 (CLT-Based Convergence Rate). If $\text{Var}(V(x, \omega)) = \sigma_V^2 < \infty$, then:
This gives a $95\%$ confidence interval for the estimator:
where $\hat{\sigma}_V$ is the sample standard deviation.
Practical Implication: To achieve a confidence interval of width $\epsilon$, we need:
9.3 Multi-Universe Convergence
When evaluating proposals across multiple universes, we need simultaneous convergence across all universe evaluations. The joint convergence criterion is:
This ensures that the worst-converged universe still satisfies the confidence requirement, consistent with the min-gate scoring philosophy.
Theorem 9.3 (Multi-Universe Convergence Bound). The minimum number of simulation paths for joint convergence across $n$ universes is:
Proof. For the joint criterion to hold, we need $\hat{\sigma}_{V_i} / \sqrt{N} \leq \epsilon_{\text{joint}} / 1.96$ for all $i$. This is equivalent to $N \geq (1.96)^2 \cdot \hat{\sigma}_{V_i}^2 / \epsilon_{\text{joint}}^2$ for all $i$. Taking the maximum over $i$ yields the stated bound. $\square$
9.4 Gate Failure Rate Convergence
Theorem 9.4 (Hoeffding Bound for Gate Failure). The empirical gate failure rate $\hat{P}_{\text{fail}}$ converges to the true failure probability with exponential tail bound:
Proof. Each simulation trial produces a Bernoulli indicator $\mathbf{1}[\text{GateScore} > 0]$. By Hoeffding's inequality applied to i.i.d. Bernoulli random variables, the sample mean concentrates around the true mean with the stated exponential bound. For $N = 10{,}000$ and $\epsilon = 0.01$, the error probability is at most $2e^{-200} \approx 10^{-87}$. $\square$
9.5 Variance Reduction Techniques
Three variance reduction techniques accelerate convergence:
Technique 1: Antithetic Variates
For each scenario $\omega_k$, generate a complementary scenario $\bar{\omega}_k$ such that $\text{Corr}(V(x, \omega_k), V(x, \bar{\omega}_k)) < 0$:
Technique 2: Control Variates
Use a simpler, analytically tractable model $V_{\text{simple}}(x, \omega)$ with known expectation:
where $c^* = \text{Cov}(V, V_{\text{simple}}) / \text{Var}(V_{\text{simple}})$ is the optimal control coefficient.
Technique 3: Importance Sampling for Tail Risk
To accurately estimate tail events (large losses), shift the sampling distribution toward the tails:
where $f$ is the original distribution and $g$ is the importance distribution with heavier tails. This dramatically reduces variance for tail risk estimation.
9.6 Convergence Monitoring in Practice
The Monte Carlo Simulation agent continuously monitors convergence using a running diagnostic:
When $\text{ConvergenceRatio}(N) \leq 1.0$, the simulation has converged. The agent reports this ratio at regular intervals during simulation, allowing human reviewers to assess progress and decide whether to continue or terminate the simulation early.
10. Human-Agent Co-Investment Loop
10.1 The Proposal-Review-Learn Cycle
In the Investment Decision Lab, investment decisions emerge from a structured interaction loop between AI evaluation agents and human investment committee members. This is a responsibility-calibrated feedback loop where the allocation of decision authority between humans and agents adapts over time based on demonstrated performance.
The loop has four stages:
1. PROPOSE: Agent evaluates investment across all universes, computes
min-gate score, generates conflict analysis, produces recommendation
2. REVIEW: Human reviews proposal, may approve, modify, or reject
3. RE-EVALUATE: System incorporates human modifications, re-runs gates
4. LEARN: System records human decision as calibration signal
10.2 Reward Signal Formalization
The human's decision is formalized as a reward signal for the agent's proposal policy $\pi(I, C; \theta)$:
For approved proposals, the reward is proportional to alignment between agent's proposed allocation and human's approved allocation. Rejections give $R = -1$. Escalations give $R = -0.5$, signaling that the agent should have recognized the need for committee review.
10.3 Agent Competence Score
The agent's competence score at time $t$ is the exponentially weighted average of historical rewards:
where $\beta \in (0,1)$ is the discount factor (typically $\beta = 0.95$). The autonomy threshold is:
where $\sigma$ is the sigmoid function. Even at maximum autonomy, all allocations remain subject to RG-I3 human approval for deployment.
10.4 Convergence of the Co-Investment Loop
Theorem 10.1 (Co-Investment Convergence). Under stationary human preferences and diminishing learning rates satisfying the Robbins-Monro conditions ($\sum \eta_t = \infty$, $\sum \eta_t^2 < \infty$), the expected modification magnitude $\mathbb{E}[|\nabla_{\text{mod}}|]$ converges to zero.
Proof. The modification gradient $\nabla_{\text{mod}} = (x_{\text{approved}} - x_{\text{proposed}}) / x_{\text{proposed}}$ provides a stochastic gradient signal. The policy update $\theta^{t+1} = \theta^t + \eta_t \cdot \nabla_{\text{mod}} \cdot \nabla_\theta \log \pi$ satisfies the Robbins-Monro stochastic approximation conditions. If the expected modification function $M(\theta) = \mathbb{E}[\nabla_{\text{mod}} \cdot \nabla_\theta \log \pi]$ has a unique zero $\theta^$ with appropriate regularity conditions, then $\theta^t \rightarrow \theta^$ almost surely. At $\theta^*$, $\mathbb{E}[\nabla_{\text{mod}}] = 0$: the agent's proposals match human preferences in expectation. Experimental validation shows convergence within 6 cycles. $\square$
11. Three-Year Research Roadmap
11.1 Year 1: Foundation
Quarter 1-2:
- Multi-Universe Investment Scoring model v1.0 implemented
- Min-gate aggregation calibrated for 3 production universes
- Investment Gate Policy (RG-I0 through RG-I3) fully operational
- Monte Carlo simulation framework with basic convergence checking
Quarter 3-4:
- Conflict Analyzer agent deployed for universe pair evaluation
- Investment Philosophy Drift Dashboard v1.0 operational
- Portfolio Optimizer with risk and ethics constraints
- First external white paper on multi-universe investment scoring
Year 1 Success Criteria:
- Min-gate scoring accuracy $\geq 90\%$ (measured against expert human judgment)
- Monte Carlo convergence achieved for $100\%$ of proposals within $10,000$ paths
- Zero capital deployments bypassing RG-I3 human approval
- Investment drift detected within 24 hours of occurrence
11.2 Year 2: Optimization
Quarter 1-2:
- Conflict-Aware Portfolio Optimizer with full constraint set deployed
- Variance reduction techniques integrated into simulation pipeline
- Human-Agent Co-Investment framework with graduated autonomy
- Investment drift detection latency reduced to < 4 hours
Quarter 3-4:
- Robust optimization with parameter uncertainty (minimax formulation)
- Real-time Monte Carlo evaluation for interactive decision support
- Explainability agent producing production-quality attribution reports
- 2 external publications, 1 conference presentation
Year 2 Success Criteria:
- Conflict-aware portfolio reduces inter-universe tension by $\geq 30\%$
- Explainability reports rated $\geq 4.2/5.0$ by human reviewers
- Simulation convergence with variance reduction achieves $99\%$ at $5,000$ paths
- Drift dashboard adopted by $\geq 3$ production universes
11.3 Year 3: Integration
Quarter 1-2:
- Full integration of Teams I-A and I-B into unified investment decision pipeline
- Sandbox Venture Simulation for pre-deployment capital stress testing
- Investment Philosophy Drift automated response protocol deployed
- Self-calibrating universe evaluation functions
Quarter 3-4:
- Industry standard proposal for multi-universe investment governance drafted
- Fail-Closed Portfolio Optimizer published as open methodology
- 3+ external publications
- Investment Decision Lab positioned as reference implementation for governed capital allocation
Year 3 Success Criteria:
- End-to-end investment pipeline latency $< 2$ hours
- Min-gate scoring accuracy $\geq 95\%$
- Adopted by $\geq 1$ external organization
- Zero undetected philosophy drift events over 12-month monitoring period
12. Risks and Mitigations
12.1 Risk: Model Overfitting to Historical Patterns
Universe evaluation functions calibrated on historical data may fail to capture emerging risks or structural market changes.
Mitigation: The Monte Carlo simulation framework explicitly generates regime-change scenarios. The Scenario Generator agent includes fat-tailed distributions and structural break models. The Drift Dashboard provides early warning when realized investment behavior diverges from model expectations.
12.2 Risk: Constraint Budget Gaming
Teams responsible for individual universes may manipulate their constraint budgets to favor their preferred investments.
Mitigation: Constraint budgets are set at the enterprise governance level, not at the universe level. The Conflict Analyzer agent monitors for systematic bias in universe evaluations. Any universe whose evaluations show statistical anomalies is flagged for recalibration.
12.3 Risk: Human Approval Bottleneck at RG-I3
Mandatory human approval at the Capital Deployment Gate may create bottlenecks, slowing investment execution.
Mitigation: The gate design includes SLA guarantees ($T_3 \leq 48$ hours). The Explainability agent's attribution reports reduce human review time by providing structured, pre-analyzed proposals. For time-sensitive investments, the Gate Engineer can invoke an expedited review track (still requiring human approval, but with compressed timelines and documented urgency rationale).
12.4 Risk: Monte Carlo Simulation Becomes a Black Box
Complex simulation models may produce results that human reviewers cannot interpret, undermining the transparency principle.
Mitigation: The Explainability agent translates simulation results into human-readable narratives. Every simulation includes: (a) top-5 scenarios driving the result, (b) sensitivity analysis showing which inputs matter most, (c) comparison to simple analytical benchmarks. Simulations that cannot be explained are flagged and may be rejected at RG-I1.
12.5 Risk: Automation Bias in Human Reviewers
Human reviewers may defer excessively to agent recommendations, rubber-stamping proposals without genuine evaluation.
Mitigation: The system tracks modification and rejection rates. If modification rates drop below a configurable threshold (e.g., <5%), the system generates an alert. Periodic calibration exercises present reviewers with intentionally flawed proposals ('canary investments') to verify active evaluation.
12.6 Risk: Philosophical Drift of the Lab Itself
The Investment Decision Lab's own research priorities may drift from its founding governance mission.
Mitigation: A meta-drift index tracks the lab's research output alignment with its founding charter. The Gate Engineer conducts quarterly research alignment reviews. The same drift detection infrastructure applied to portfolios is applied reflexively to the lab's own research portfolio.
13. Conclusion
The Investment Decision Lab addresses a structural gap in enterprise AI governance: the absence of dedicated research infrastructure for multi-universe capital allocation. Traditional investment frameworks — built for single-objective optimization in homogeneous markets — fail when investments must simultaneously satisfy financial, ethical, risk, and responsibility constraints across multiple autonomous business domains.
The lab's architecture embodies three principles that distinguish it from conventional investment research:
First, min-gate scoring replaces weighted aggregation. The scoring function $S_{\text{invest}} = \min_i U_i(x)$ ensures that no universe can be sacrificed for aggregate gain. This is the mathematical expression of the MARIA OS principle that system quality is determined by its most constrained dimension.
Second, fail-closed capital deployment replaces optimistic allocation. The Fail-Closed Portfolio Optimizer either produces a fully constraint-compliant allocation or explicitly refuses, with diagnostic information explaining infeasibility. No capital moves without satisfying all responsibility constraints.
Third, conflict is made visible, not hidden. The Conflict-Aware Investment Engine quantifies inter-universe tensions, performs spectral analysis of conflict structures, and escalates unresolvable conflicts to human decision-makers with full analytical context.
The Investment Gate design (RG-I0 through RG-I3) ensures graduated human involvement, with mandatory human approval at the capital deployment stage. Monte Carlo simulation provides convergence-guaranteed scenario evaluation before any real capital is committed. The Investment Philosophy Drift Dashboard provides continuous monitoring of alignment between declared philosophy and realized behavior. And the human-agent co-investment loop formalizes graduated autonomy with provable convergence.
The final message parallels the broader MARIA OS philosophy: in the AGI era, the question is not how much return your investments generate. The question is how much responsibility your investment process structurally preserves.
Appendix A: MARIA OS Coordinate Assignment
Investment Lab Universe: G1.U_IL
├── P1: Multi-Universe Investment Core Lab (Team I-A)
│ ├── Z1: Investment Modeling Zone
│ │ ├── H1: Investment Modeling Lead
│ │ ├── A1: Research Planner Agent
│ │ ├── A2: Financial Modeling Agent
│ │ └── A3: Market Dynamics Agent
│ ├── Z2: Constraint & Conflict Zone
│ │ ├── H1: Core Systems Engineer
│ │ ├── A1: Ethics & Regulatory Agent
│ │ └── A2: Conflict Analyzer Agent
│ └── Z3: Gate Operations Zone
│ ├── H1: Gate Engineer
│ └── A1: Evaluation Agent
├── P2: Capital Allocation & Simulation Lab (Team I-B)
│ ├── Z1: Optimization & Simulation Zone
│ │ ├── H1: Quant Researcher
│ │ ├── A1: Portfolio Optimizer Agent
│ │ ├── A2: Risk Budget Agent
│ │ └── A3: Monte Carlo Simulation Agent
│ └── Z2: Scenario & Explainability Zone
│ ├── H1: Runtime Engineer
│ ├── A1: Scenario Generator Agent
│ └── A2: Explainability Agent
Appendix B: Investment Gate Database Schema
CREATE TABLE investment_proposals (
id UUID PRIMARY KEY,
title TEXT NOT NULL,
description TEXT,
target_universes TEXT[] NOT NULL,
gate_level INT CHECK (gate_level BETWEEN 0 AND 3),
status TEXT CHECK (status IN ('observed','simulated','proposed','deployed','rejected')),
min_gate_score NUMERIC,
evidence_bundle_hash TEXT NOT NULL,
created_at TIMESTAMPTZ DEFAULT now(),
deployed_at TIMESTAMPTZ,
deployed_by TEXT,
rollback_plan JSONB
);
CREATE TABLE universe_evaluations (
id UUID PRIMARY KEY,
proposal_id UUID REFERENCES investment_proposals(id),
universe_id TEXT NOT NULL,
financial_score NUMERIC NOT NULL,
ethics_score NUMERIC NOT NULL,
risk_score NUMERIC NOT NULL,
responsibility_score NUMERIC NOT NULL,
aggregate_score NUMERIC NOT NULL,
evaluation_evidence JSONB NOT NULL,
evaluated_at TIMESTAMPTZ DEFAULT now(),
evaluated_by TEXT NOT NULL
);
CREATE TABLE investment_conflicts (
id UUID PRIMARY KEY,
proposal_id UUID REFERENCES investment_proposals(id),
universe_a TEXT NOT NULL,
universe_b TEXT NOT NULL,
conflict_intensity NUMERIC NOT NULL,
resolution_strategy TEXT CHECK (resolution_strategy IN ('pareto','negotiation','escalation')),
resolved BOOLEAN DEFAULT false,
resolved_by TEXT,
resolved_at TIMESTAMPTZ,
evidence_hash TEXT NOT NULL,
created_at TIMESTAMPTZ DEFAULT now()
);
CREATE TABLE investment_gate_transitions (
id UUID PRIMARY KEY,
proposal_id UUID REFERENCES investment_proposals(id),
from_level INT NOT NULL,
to_level INT NOT NULL,
decision TEXT CHECK (decision IN ('pass','block','defer')),
reviewer TEXT NOT NULL,
rationale TEXT NOT NULL,
evidence_hash TEXT NOT NULL,
created_at TIMESTAMPTZ DEFAULT now()
);
CREATE TABLE philosophy_drift_snapshots (
id UUID PRIMARY KEY,
snapshot_time TIMESTAMPTZ NOT NULL,
philosophy_vector JSONB NOT NULL,
drift_index NUMERIC NOT NULL,
drift_level TEXT CHECK (drift_level IN ('normal','elevated','critical')),
component_decomposition JSONB,
top_drift_decisions UUID[],
created_at TIMESTAMPTZ DEFAULT now()
);
Appendix C: Mathematical Notation Reference
| Symbol | Meaning |
| --- | --- |
| $S_{\text{invest}}(x)$ | Min-gate investment score for proposal $x$ |
| $U_i(x)$ | Universe $i$'s evaluation of proposal $x$ |
| $w$ | Portfolio weight vector |
| $E[R(w)]$ | Expected portfolio return |
| $B_r, B_e, B_\rho$ | Risk, ethics, and responsibility budgets |
| $\lambda_r^, \lambda_e^, \lambda_\rho^*$ | Optimal Lagrange multipliers (shadow prices) |
| $P(t)$ | Investment philosophy vector at time $t$ |
| $P_0$ | Declared investment philosophy vector |
| $D_{\text{invest}}(t)$ | Investment drift index (Mahalanobis distance) |
| $\Sigma$ | Covariance matrix of philosophy vectors |
| $\alpha_H, \alpha_A$ | Human and agent responsibility weights |
| $\hat{V}_N(x)$ | Monte Carlo value estimator with $N$ paths |
| $\sigma_V$ | Standard deviation of simulation value |
| $I_{ij}(x)$ | Conflict intensity between universes $i$ and $j$ |
| $\text{CL}(w)$ | System conflict load for portfolio $w$ |
| $\mu$ | Conflict aversion parameter |
| $\tau_k$ | Threshold for investment gate level $k$ |
| $\epsilon_k$ | Evidence requirement for gate level $k$ |
| $\mathcal{G}_I$ | Investment gate finite state machine |
| $\mathcal{W}_{\text{feasible}}$ | Feasible portfolio set |
| $\phi_d$ | Shapley value attribution for decision $d$ |
| $K(t)$ | Agent competence score at time $t$ |
| $R(r, d)$ | Reward signal from human decision $d$ on proposal $r$ |
Appendix D: Key Theorems Summary
| Theorem | Statement | Section |
| --- | --- | --- |
| 5.1 | Investment Gate Completeness: every proposal reaches terminal state in finite time | 5.3 |
| 5.2 | Fail-Closed Preservation: the gate policy preserves fail-closed at every level | 5.4 |
| 6.1 | Feasibility Preservation: returned allocations satisfy all constraints with probability 1 | 6.3 |
| 6.2 | Strong Duality: optimal multipliers are marginal return costs of governance constraints | 6.4 |
| 7.1 | Conflict-Return Tradeoff: optimal return is strictly decreasing in conflict aversion $\mu$ | 7.4 |
| 8.1 | Drift Early Warning: bounded time to breach under persistent positive velocity | 8.4 |
| 9.1 | Strong Convergence: Monte Carlo estimator converges almost surely | 9.1 |
| 9.2 | CLT Convergence Rate: $\sqrt{N}$ rate with normal limiting distribution | 9.2 |
| 9.3 | Multi-Universe Convergence: minimum paths for joint convergence across $n$ universes | 9.3 |
| 9.4 | Hoeffding Bound: exponential concentration for gate failure rate estimation | 9.4 |
| 10.1 | Co-Investment Convergence: agent proposals converge to human preferences | 10.4 |