# MARKOWITZ.SYSTEMS — Portfolio Theory for the Agentic Firm

> Portfolio Selection · Harry M. Markowitz · Journal of Finance, 1952 · Cowles Foundation Monograph 16, Wiley, 1959 · Nobel Memorial Prize, 1990 (shared with Merton H. Miller & William F. Sharpe)
> **Slogan:** Correlation is the new headcount.

## Regional Quantitative Finance, Actuarial & CFO Risk Hub: Netherlands — PP-04 [arXiv:2410.10347 · ICML 2025]: A Unified Approach to Routing and Cascading for LLMs
- **Canonical URL:** https://markowitz.systems/netherlands/preprint/pp-04
- **National Degree Track:** *Quantitative Finance, Econometrics, Actuarial Science & Operations Research (Erasmus Econometric Inst., UvA, Tilburg, VU)* (High-precision Dutch Econometrics & Actuarial cohort (~600–1,100 MSc/PhD/AAG per year))
- **Universities:** Erasmus University Rotterdam (Econometric Institute), Tilburg University (Netspar / Finance), University of Amsterdam (UvA Actuarial Science), Tinbergen Institute, VU Amsterdam, Maastricht
- **Local Channels:** Koninklijk Actuarieel Genootschap (AG), KVS, Netspar, AcademicTransfer.com
- **Authors & Venue:** Jasper Dekoninck, Maximilian Baader & Martin Vechev (ETH Zurich) (2024 / 2025, arXiv:2410.10347 · ICML 2025) · https://arxiv.org/abs/2410.10347
- **Core Formula:** `max_π E_x[ Q(π(x)) − λ · Cost(π(x)) ]  (Exact LP Pareto Frontier)`
- **CFO & Practitioner Upskilling Takeaway:** Replace ad-hoc engineering prompt chains with LP-calibrated cascade routers where λ is the CFO's explicit shadow price of inference budget ($ per basis point of accuracy).

## Hero Formula & Explicit Variable Definitions
`min wᵀΣw s.t. wᵀμ = μ*, wᵀ1 = 1`
- **μ_i = E[v_i]:** Expected net value per task (task value minus token & human review cost).
- **Σ_{ij} = Cov(v_i, v_j) = Cov(e_i, e_j) (if value = const − error):** Covariance of agent task outcomes (equivalently, of signed task errors e_i around ground truth when v_i = c_i − e_i with constant c_i).

## Your Agents Share One Base Model. Your Risk Doesn’t Diversify.
Every firm deploying AI agents now holds a portfolio of agents, vendors, projects and cash, with correlations nobody has measured. Under equal weights, equal variance σ², and pairwise correlation ρ, portfolio variance is σ²[ρ + (1−ρ)/N], which tends to σ²ρ, not zero, as N grows. Effective independent capacity is N_eff = N / (1 + (N−1)ρ): at N = 50 and ρ = 0.6, fifty agents diversify like N_eff ≈ 1.6 independent agents (collapsing to 1 as ρ → 1). Markowitz.systems connects finance teams with people trained to estimate Σ, not just μ.

**Tags:** Mean–Variance · Covariance Estimation · Semivariance · CVaR · Tobin Separation · Capital Budgeting · Agentic Close & Treasury

## Node 01 — For Financial Economists & Quantitative Specialists
Part-time remote R&D pods ($20–45/hr, 10–25 hrs/wk). Apply mean-variance optimization, Ledoit–Wolf covariance estimation (N > T), downside semivariance, CVaR, Agent Sharpe Ratios, real options and capital budgeting to agentic finance: month-end close, treasury, FP&A and audit trails. Who we look for: quants, actuaries, risk managers, financial economists, FP&A analysts.
- **Apply Directly:** https://markowitz.systems/#apply

## Node 02 — For CFO Offices
Agent deployments are capital allocation decisions under correlated risk. Engage specialists who measure error covariance across agents and vendors (Anthropic, OpenAI, Google paired evals), size exposure to any single model, map audit materiality to CVaR, and build an efficient frontier for your AI spend. → frontier@markowitz.systems
- **CFO Frontier Desk:** mailto:frontier@markowitz.systems

## Six Structural Portfolio Theses for the Agentic CFO
### THESIS 01 // CAPM DECOMPOSITION — Systematic Base-Model Error (β) vs. Idiosyncratic Noise
- **Formula:** `e_i = β_i F_base + ε_i  ⇒  Var(∑ w_i e_i) = (∑ w_i β_i)² σ²_base + ∑ w_i² σ²_{ε,i}`
- **Thesis:** By analogy with Sharpe’s CAPM, only idiosyncratic workflow noise ε_i diversifies away as 1/N. Systematic exposure β_i to shared foundation-model blind spots, context rot, or API outages never diversifies within one model family.

### THESIS 02 // TOBIN SEPARATION & CML — Deterministic Code as the Idealized Risk-Free Leg (σ₀ = 0)
- **Formula:** `w_total = (1 − α) w_code + α w_tan,   μ_CML(σ_p) = μ_code + SR_tan · σ_p`
- **Thesis:** SQL balance invariants, ERP validation rules, and deterministic Python scripts act as an idealized risk-free leg (σ₀ = 0, abstracting from software bugs), while LLM agents form the risky tangency portfolio. By Tobin’s Separation Theorem, optimal finance architecture lies on the Capital Market Line.

### THESIS 03 // SINGULARITY WHEN N > T — 50 Agents & 30 Days of Evals: Why Sample Σ Is Singular
- **Formula:** `N = 50 > T = 30  ⇒  rank(S_emp) ≤ T − 1 = 29 < 50  ⇒  det(S_emp) = 0`
- **Thesis:** With 50 deployed agents and 30 days of eval telemetry, the empirical covariance matrix S_emp is mathematically singular and non-invertible — fabricating illusory 'zero-risk' combinations unless regularized via Ledoit–Wolf shrinkage Σ̂_LW. This immediately separates trained quants from prompt engineers.

### THESIS 04 // MONOGRAPH 16 CH. 9 — Downside Semivariance & Audit Materiality as Risk Budget
- **Formula:** `Σ⁻_{ij} = E[min(v_i−τ,0) min(v_j−τ,0)],   CVaR_99%(L(w)) ≤ M_audit`
- **Thesis:** Markowitz devoted Chapter 9 of his 1959 Wiley monograph to semivariance because upside dispersion is benign. Framing statutory audit materiality M_audit directly as the firm’s 99% CVaR budget translates agent tail risk into native CFO & Audit Committee language.

### THESIS 05 // SHARPE → WALRASIAN ROE — Agent Sharpe Ratio & Banking Large-Exposure Caps
- **Formula:** `SR_i = (E[Value_i − Cost_i] − μ_code) / σ_i,   ∑_{i ∈ vendor m} w_i ≤ κ_LEX`
- **Thesis:** Defining an Agent Sharpe Ratio (net value minus token/review cost per unit of outcome σ_i) bridges Markowitz/Sharpe portfolio selection directly to risk-adjusted Agent ROE in Walrasian compute allocation — paired with Basel-style Large Exposure (LEX) limits on any single model vendor.

### THESIS 06 // EMPIRICAL PAIRED EVALS & REGIME DRIFT — Multi-Vendor Diversification (Anthropic · OpenAI · Google) & Rebalancing
- **Formula:** `ρ̂_{AB} = (p_11 − p_A p_B) / √[p_A(1−p_A)p_B(1−p_B)],   (μ_t, Σ_t) ↦ (μ_{t+1}, Σ_{t+1})`
- **Thesis:** Routing and cross-verifying across Anthropic, OpenAI, and Google measurably lowers pairwise error correlation ρ — a falsifiable claim tested via item-level paired evals. Because every model checkpoint release shifts the regime, Σ_t must be re-estimated and agent weights w* rebalanced on every release.

## Foundational Axioms from Harry M. Markowitz (1952–1990)
### HARRY M. MARKOWITZ · JOURNAL OF FINANCE, VOL. 7, NO. 1 (MARCH 1952, P. 77 — OPENING SENTENCES) — Two-Stage Portfolio Selection: Paired Evals First, Budget Allocation Second
> "The process of selecting a portfolio may be divided into two stages. The first stage starts with observation and experience and ends with beliefs about the future performances of available securities. The second stage starts with the relevant beliefs about future performances and ends with the choice of portfolio."

**Implication:** For an Agentic CFO, Stage 1 is rigorous empirical evals on your own ledger and workflow telemetry to estimate expected net value μ and cross-agent covariance Σ; Stage 2 is capital and compute budget allocation w*.

### HARRY M. MARKOWITZ · JOURNAL OF FINANCE, VOL. 7, NO. 1 (MARCH 1952, PP. 77–79) — Why Maximizing Expected Benchmark Accuracy μ Concentrates Correlated Risk
> "Does the rule that the investor does (or should) maximize discounted expected, or anticipated, returns imply that the investor should diversify? Obviously not: if the investor maximizes expected returns, he puts all his funds in the security with the maximum expected return. Diversification cannot eliminate all variance."

**Implication:** Under equal weights, equal variance σ², and pairwise correlation ρ, variance is σ²[ρ + (1−ρ)/N] → σ²ρ and N_eff = N/(1+(N−1)ρ). At N = 50 and ρ = 0.6, fifty agents on one base model yield N_eff ≈ 1.64 independent agents.

### COWLES FOUNDATION MONOGRAPH 16 · PORTFOLIO SELECTION: EFFICIENT DIVERSIFICATION OF INVESTMENTS (WILEY, 1959, CH. 9) — Downside Semivariance (S⁻): Penalizing Financial Misstatements, Never Upside Speed
> "Analyses based on S (semivariance) tend to produce better portfolios than those based on V (variance). Variance considers extremely high and extremely low returns equally undesirable; an analysis based on S concentrates on reducing losses."

**Implication:** A CFO never penalizes an agent for closing books faster or saving extra tokens; only downside deviations — accounting misstatements, duplicate wires, covenant breaches — matter, requiring downside semivariance Σ⁻ and CVaR.

### NOBEL MEMORIAL PRIZE IN ECONOMIC SCIENCES · STOCKHOLM (1990 — SHARED WITH M. H. MILLER & W. F. SHARPE) — From Markowitz Σ to Sharpe’s β, Capital Market Line & Corporate Risk Governance
> "What is required is essentially a procedures-oriented theory of rational behavior under uncertainty — balancing expected return against risk across the covariance structure of the whole portfolio."

**Implication:** Sharing the 1990 Nobel Prize with William F. Sharpe and Merton H. Miller links cross-agent covariance Σ directly to systematic base-model β, the Agent Sharpe Ratio (Value − Cost)/σ, and corporate capital-structure risk limits.

## Canonical Equations (12 Actionable Quantitative CFO Formulas)
- **EQ-01 (Markowitz Mean–Variance AI Spend Program (Explicit μ vs. Σ)):** `min_w  wᵀΣw   s.t.   wᵀμ = μ*,   wᵀ1 = 1,   wᵀc ≤ B_CFO   [μ_i = E[v_i], Σ_ij = Cov(v_i,v_j) = Cov(e_i,e_j) if value = const − error]` — Unifies return and error notation rigorously: when net task value satisfies v_i = c_i − e_i with constant c_i (if value = const − error) and signed task error e_i, Cov(v_i, v_j) = Cov(e_i, e_j) minimizes joint failure variance at target net value μ*.
- **EQ-02 (Equal-Weight, Equal-Variance Correlated Fleet Floor):** `Under w_i = 1/N, Var(e_i) = σ², Corr(e_i, e_j) = ρ (i ≠ j):  Var(ē) = σ²[ρ + (1 − ρ)/N] →_{N→∞} σ²ρ` — States all three required assumptions explicitly: under equal weights, common variance σ², and pairwise correlation ρ, scaling N never drives risk to zero — variance hits the irreducible base-model floor σ²ρ.
- **EQ-03 (Effective Independent Headcount (N_eff — 'Correlation Is the New Headcount')):** `N_eff(N, ρ) = N / [ 1 + (N − 1) ρ ]   (e.g. N = 50, ρ = 0.60 ⇒ N_eff = 50 / 30.4 ≈ 1.64;  ρ → 1 ⇒ N_eff → 1)` — Exact diversification equivalence: 50 agents sharing a foundation model at ρ = 0.60 provide the statistical error reduction of only 1.64 independent agents (collapsing to 1.00 as ρ → 1).
- **EQ-04 (Systematic Base-Model Beta (β_i) vs. Idiosyncratic Agent Noise (CAPM)):** `e_i = β_i F_base + ε_i,   E[F_base ε_i] = 0  ⇒  σ²_p(w) = (wᵀβ)² σ²_base + ∑_{i=1}^N w_i² σ²_{ε,i}` — Prompt tweaks only reduce idiosyncratic σ²_{ε,i} (which diversifies as 1/N); systematic sensitivity β_i to shared foundation-model hallucinations requires orthogonal model families or deterministic invariants.
- **EQ-05 (Deterministic Code as Idealized Risk-Free Leg & Capital Market Line (Tobin Separation)):** `σ_code = 0 (idealized risk-free leg),  μ_p = (1 − α) μ_code + α w_tanᵀ μ,   μ_CML(σ_p) = μ_code + [ (w_tanᵀ μ − μ_code) / √(w_tanᵀ Σ w_tan) ] σ_p` — Rules, SQL reconciliations, and deterministic scripts serve as an idealized risk-free leg (σ_code = 0 on well-tested invariants); optimal CFO architecture blends deterministic code with the diversified multi-vendor LLM tangency portfolio w_tan along the CML.
- **EQ-06 (Finite-Sample Singularity (N > T) & Ledoit–Wolf Covariance Shrinkage):** `rank(S_emp) ≤ min(N, T − 1)  (singular if N > T, e.g. 50 > 30);   Σ̂_LW = (1 − δ*) S_emp + δ* ν I_N` — With 50 agents and 30 days of evals, sample covariance S_emp is rank-deficient (rank ≤ 29) and non-invertible; Ledoit–Wolf shrinkage δ* restores positive-definiteness and out-of-sample frontier stability.
- **EQ-07 (Cowles Monograph 16 Downside Semivariance Matrix (Σ⁻, Markowitz 1959 Ch. 9)):** `Σ⁻_{ij}(τ) = E[ min(v_i − τ, 0) · min(v_j − τ, 0) ],     S⁻_p(w) = wᵀ Σ⁻(τ) w` — Penalizes only joint downside shortfalls below accounting accuracy target τ — aligning optimization with CFO asymmetry where upside speed is welcome but reporting misstatements are costly.
- **EQ-08 (Audit Materiality (M_audit) as Rockafellar–Uryasev CVaR Risk Budget):** `CVaR_α(L(w)) = min_{ζ ∈ ℝ} { ζ + (1−α)⁻¹ E[ max(L(w) − ζ, 0) ] } ≤ M_audit,   ∑_i w_i (∂CVaR_α / ∂w_i) ≤ M_audit` — Translates statutory SOX / external-audit materiality M_audit directly into an enforceable 99% tail-loss budget decomposed by Euler marginal risk contribution across every ledger workflow.
- **EQ-09 (Agent Sharpe Ratio & Bridge to Risk-Adjusted Agent ROE (Walrasian Equilibrium)):** `SR_i = ( E[Value_i − Cost_i] − μ_code ) / σ_i,     ROE_agent,i = E[Value_i − Cost_i] / c_i = μ_code/c_i + SR_i · (σ_i / c_i)` — Measures excess net value per unit of outcome volatility σ_i over deterministic scripts — linking Markowitz/Sharpe portfolio weights directly to risk-adjusted Agent ROE in compute market equilibrium.
- **EQ-10 (Single-Model Banking Large-Exposure Cap (LEX) & Paired-Eval Multi-Vendor ρ̂):** `∑_{i ∈ Vendor m} w_i · EAD_i ≤ κ_LEX · M_audit  (m ∈ {Anthropic, OpenAI, Google});   ρ̂_{AB} = (p_11 − p_A p_B) / √(p_A q_A p_B q_B)` — Applies banking Large Exposures supervisory caps to any single foundation-model provider and verifies cross-vendor diversification (Anthropic, OpenAI, Google) empirically via paired benchmark evals.
- **EQ-11 (Non-Stationarity Across Model Releases & Covariance Rebalancing Trigger):** `Release θ_t → θ_{t+1}:  || Σ̂_{t+1} − Σ̂_t ||_F / || Σ̂_t ||_F > γ_rebal  ⇒  Re-solve w*_{t+1} = argmin_w wᵀ Σ̂_{t+1} w` — Every model version bump, quantization update, or prompt drift alters joint failure modes; continuous canary evals detect covariance regime shifts and rebalance agent weights automatically.
- **EQ-12 (Multi-Tier Cascade Spend, Selective HITL Gate & CFO RAROC Hurdle):** `E[c_casc] = c_1 + P(q_1<τ_1)c_2 + P(q_1<τ_1,q_2<τ_2)c_3;  Defer→Human ⇔ p_err(x|Σ)L_fin(x) > w_H t_rev − c_AI;  RAROC_AI ≥ r_WACC` — Combines FrugalGPT/RouteLLM cost cascades (−65% to −90% token spend), economic Human-in-the-Loop exception gating, and CVaR-charged Return on Invested Capital.

## Audited CFO Optimization Case Studies (AI Agent Cascades & Hybrid Human + AI Pods)
### CASE-01 — AGENTIC MONTH-END CLOSE & MULTI-ENTITY GL RECONCILIATION ($4.2B Global Industrial Group · 18 ERP Ledgers · 48 Close Agents + Accounting Team)
- **Baseline:** 48 autonomous journal & intercompany reconciliation agents all calling 1 flagship LLM (pairwise error correlation ρ = 0.61) + 14 FTE accountants reviewing ad-hoc spreadsheets.
- **Markowitz + Cascade + HITL Architecture:** 3-tier cascade (deterministic SQL balance invariants → small fine-tuned extractor → 2 orthogonal foundation models on ambiguous 14% items) + 6 Senior CPA/Quant controllers gated by p_err·L > w_H·t_rev.
- **Equations Applied:** EQ-03 (N_eff), EQ-05 (Tobin CML SQL = R_f), EQ-08 (CVaR_99% ≤ M_audit) & EQ-12 (Selective HITL Gate)
- **Benchmarks:** Annual LLM API Spend: $640K / yr → $185K / yr (−71% Token Cost) | Blended Human + AI Close Cost: $2.84M / yr → $1.53M / yr (−46% Total Spend) | 99% Tail Misstatement (CVaR_99%): $1.95M → $210K (−89% Tail Risk) | Month-End Close Cycle: 6.8 Days → 2.3 Days (4.5 Days Faster)
- **CFO Takeaway:** Eliminating single-model correlation (ρ dropped from 0.61 to 0.11 across orthogonal verifiers) cut SOX audit adjustments by 89% while saving $1.31M/yr in combined token and overtime labor.

### CASE-02 — FP&A, PROCUREMENT & AP INVOICE FINOPS (CASCADE ROUTING) ($1.9B B2B SaaS & Cloud Enterprise · 1.8M Monthly AP, Contract & Variance Queries)
- **Baseline:** 100% of vendor contracts, PO matching, and departmental variance commentary routed to frontier reasoning models ($15/1M output tokens) inside 6-step agent retry loops.
- **Markowitz + Cascade + HITL Architecture:** Calibrated preference router (RouteLLM / LP threshold τ* = 0.84) handles 79% of routine POs on $0.25/1M tier, 16% on mid-tier ($2.50/1M), and escalates 5% covenant/tax anomalies to dual-frontier + FP&A Lead.
- **Equations Applied:** EQ-09 (Agent Sharpe Ratio) & EQ-12 (FrugalGPT / Dekoninck–Vechev LP Cascade)
- **Benchmarks:** Annual Inference Run-Rate: $1.62M / yr → $356K / yr (−78% API Spend) | Extraction & Coding Precision: 99.1% → 99.4% (+30 bps Quality) | Duplicate / Erroneous AP Leakage: 42 bps of AP → 6 bps of AP (−86% Leakage) | Implementation Payback: Unbounded → 5.5 Weeks (RAROC = 410%)
- **CFO Takeaway:** Capping unconstrained retry loops at K ≤ 2 and routing by calibrated difficulty preserved frontier quality on complex contracts while returning $1.26M/yr directly to EBITDA.

### CASE-03 — HYBRID HUMAN + AI 'CYBERNETIC PODS' IN COMMERCIAL FINANCE & DEAL DESK ($950M Enterprise FinTech · 24 Pricing, Underwriting & Deal-Desk Analysts)
- **Baseline:** 24 siloed financial analysts manually drafting bespoke enterprise pricing memos alongside uncalibrated single-vendor chat copilots (high variance, slow legal/finance sign-off).
- **Markowitz + Cascade + HITL Architecture:** Re-organized into 8 Cybernetic Pods (1 Senior Financial Economist / Deal Architect + 2 orthogonal agents: Proposal Generator + Adversarial Margin/Covenant Critic) following NBER w33641 & MIT task-split rules.
- **Equations Applied:** EQ-03 (N_eff), EQ-09 (Agent Sharpe ↔ ROE), EQ-10 (Paired-Eval Multi-Vendor ρ̂) & EQ-12 (RAROC_AI)
- **Benchmarks:** Underwritten Deals per FTE: 18 / mo → 26 / mo (+44% Productivity) | Blended Cost per Deal Memo: $1,480 → $975 (−34% Unit Cost) | Vendor Concentration (HHI): 1.00 (Monoculture) → 0.36 (3-Model) (−64% Concentration) | Gross-Margin Pricing Errors: 3.8% of Deals → 0.4% of Deals (−89% Margin Drift)
- **CFO Takeaway:** Following Vaccaro–Malone (2024) and Dell'Acqua et al. (2025), AI agents execute first-pass quantitative synthesis while senior humans own boundary exceptions — avoiding automation bias.

### CASE-04 — AUTONOMOUS CORPORATE TREASURY, AR/AP & INTRADAY LIQUIDITY SWEEP ($6.5B Multinational Retailer · $140M Daily Working-Capital Cash Pool Across 9 Currencies)
- **Baseline:** Cash-forecasting and AR remittance-matching copilot built on a single cloud LLM API; a 4-hour vendor outage + correlated FX settlement-date hallucination trapped $38M in non-earning buffers.
- **Markowitz + Cascade + HITL Architecture:** Markowitz covariance-constrained portfolio (‖Bᵀw‖_∞ ≤ 0.40 across cloud + self-hosted open weights), deterministic cash-conservation constraints, and overnight buffer sized via CVaR_99.5%.
- **Equations Applied:** EQ-01, EQ-06 (Ledoit–Wolf Σ̂_LW), EQ-08 (Treasury CVaR_99.5%) & EQ-10 (Banking LEX Vendor Cap)
- **Benchmarks:** Idle Precautionary Cash Buffer: $38.0M → $24.0M ($14M Freed Cash) | Net Working-Capital Yield Pickup: Baseline → +62 bps / yr (+$868K / yr Carry) | Single-Vendor Outage Downtime: 11.5 hrs / yr → 0.0 hrs / yr (100% Failover SLA) | FX / Remittance Mismatch Rate: 1.45% → 0.12% (−92% Error Rate)
- **CFO Takeaway:** Treating model providers like counterparty credit exposures unlocked $14M in trapped cash and generated $868K/yr in risk-free treasury yield — paying for the entire AI risk pod 7× over.

## Original Academic Preprints for the CFO (arXiv & NBER)
### PP-01 [arXiv:2506.07962 · ICML 2025] Correlated Errors in Large Language Models
- **Authors:** Elliot Kim, Avi Garg, Kenny Peng & Nikhil Garg (Cornell Tech)
- **Links:** [Original Preprint](https://arxiv.org/abs/2506.07962) · [Direct PDF](https://arxiv.org/pdf/2506.07962)
- **Core Formula:** `P(Model_A wrong ∧ Model_B wrong | Model_B wrong) ≈ 0.45–0.60  ≫  p_err`
- **Finding:** Empirical audit of 350+ LLMs proving that models agree up to 60% of the time when BOTH are wrong — and discovering the 'Accuracy Paradox': higher-performing frontier models exhibit stronger pairwise error correlation even across distinct vendors.
- **CFO Actionable:** Never assume multi-agent ensembles or LLM-as-a-Judge diversify risk without empirically measuring off-diagonal error covariance Σ on your own ledger and contract data.

### PP-02 [arXiv:2305.05176 · TMLR] FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance
- **Authors:** Lingjiao Chen, Matei Zaharia & James Zou (Stanford University)
- **Links:** [Original Preprint](https://arxiv.org/abs/2305.05176) · [Direct PDF](https://arxiv.org/pdf/2305.05176)
- **Core Formula:** `min_{L, τ} E[ ∑_{j=1}^k c_{L_j} · 1(score_{j−1} < τ_{j−1}) ]  s.t.  Acc ≥ Acc_frontier`
- **Finding:** Introduces learned confidence scoring and sequential LLM cascades (small → medium → flagship), matching or beating flagship GPT-4 accuracy with up to 90%–98% lower inference spend.
- **CFO Actionable:** Mandate multi-tier LLM cascades before approving any enterprise API budget renewal; forbid routing 100% of routine extraction or FP&A tickets to flagship reasoning tiers.

### PP-03 [arXiv:2406.18665 · ICLR 2025] RouteLLM: Learning to Route LLMs with Preference Data
- **Authors:** Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M. Waleed Kadous & Ion Stoica (UC Berkeley / Anyscale)
- **Links:** [Original Preprint](https://arxiv.org/abs/2406.18665) · [Direct PDF](https://arxiv.org/pdf/2406.18665)
- **Core Formula:** `R_τ(x) = M_strong  if  P̂(M_strong ≻ M_weak | x) > τ,  else  M_weak`
- **Finding:** Trains lightweight routers on human preference telemetry to dynamically dispatch queries between a frontier LLM and a cheap open-weight model, cutting API costs by 50%–85% at 95% of frontier quality.
- **CFO Actionable:** Expose the router threshold τ as a live CFO FinOps policy dial: tighten τ during routine operations and lower τ (more frontier + human review) during quarter-end close.

### PP-04 [arXiv:2410.10347 · ICML 2025] A Unified Approach to Routing and Cascading for LLMs
- **Authors:** Jasper Dekoninck, Maximilian Baader & Martin Vechev (ETH Zurich)
- **Links:** [Original Preprint](https://arxiv.org/abs/2410.10347) · [Direct PDF](https://arxiv.org/pdf/2410.10347)
- **Core Formula:** `max_π E_x[ Q(π(x)) − λ · Cost(π(x)) ]  (Exact LP Pareto Frontier)`
- **Finding:** Unifies upfront routing and sequential cascading into a single linear-programming optimization problem, proving mathematical optimality along the Cost–Quality efficient frontier.
- **CFO Actionable:** Replace ad-hoc engineering prompt chains with LP-calibrated cascade routers where λ is the CFO's explicit shadow price of inference budget ($ per basis point of accuracy).

### PP-05 [arXiv:2407.01502 · TMLR] AI Agents That Matter
- **Authors:** Sayash Kapoor, Benedikt Stroebl, Zachary S. Siegel, Nitya Nadgir & Arvind Narayanan (Princeton University)
- **Links:** [Original Preprint](https://arxiv.org/abs/2407.01502) · [Direct PDF](https://arxiv.org/pdf/2407.01502)
- **Core Formula:** `Pareto_CFO = { a ∈ Agents : ¬∃ a' s.t. Cost(a') ≤ Cost(a) ∧ Acc(a') > Acc(a) }`
- **Finding:** Demonstrates that benchmark-winning agent scaffolds routinely burn 10×–50× more API dollars via uncalibrated retry loops while failing to beat simple deterministic baselines on joint (USD Cost, Accuracy) frontiers.
- **CFO Actionable:** Require every internal AI team and SaaS vendor to report two-dimensional (Cost per Resolved Task, Audited Accuracy) curves and cap autonomous retry depth K.

### PP-06 [NBER Working Paper No. 33641 · Harvard / Wharton] The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise
- **Authors:** Fabrizio Dell'Acqua, Charles Ayoubi, Hila Lifshitz, Raffaella Sadun, Ethan Mollick, Lilach Mollick, Yi Han, Jeff Goldman, Hari Nair, Stewart Taub & Karim R. Lakhani
- **Links:** [Original Preprint](https://www.nber.org/papers/w33641) · [Direct PDF](https://www.nber.org/system/files/working_papers/w33641/w33641.pdf)
- **Core Formula:** `Quality(1 Human + AI) ≈ Quality(2-Human Cross-Functional Team),   ΔTime = −16.4%`
- **Finding:** Pre-registered field RCT with 776 P&G professionals showing that a single specialist equipped with AI matches the output quality of two-person cross-functional human teams while working 16.4% faster and bridging domain silos.
- **CFO Actionable:** Restructure bloated multi-handoff committees into compact 'Cybernetic Pods' (1 Senior Quant/CPA + orthogonal AI agents), capturing 25%–35% unit-cost reduction without sacrificing quality.

### PP-07 [arXiv:2503.18238 · MIT IDE] Collaborating with AI Agents: Field Experiments on Teamwork, Productivity, and Performance
- **Authors:** Harang Ju & Sinan Aral (MIT Sloan / MIT Initiative on the Digital Economy)
- **Links:** [Original Preprint](https://arxiv.org/abs/2503.18238) · [Direct PDF](https://arxiv.org/pdf/2503.18238)
- **Core Formula:** `ΔOutput/Worker = +60%,   ΔDirect_Editing = −23%,   ΔSupervisory_Integration = +23%`
- **Finding:** Randomized field experiment (N = 2,310) showing Human + AI Agent teams achieve +60% productivity per worker as humans reallocate 23% of effort from raw drafting to high-value verification and orchestration.
- **CFO Actionable:** Budget for human role upgrading rather than naive 100% headcount replacement: pair fewer, higher-skilled reviewers with multi-agent execution pipelines.

### PP-08 [arXiv:2405.11896 · Nature Human Behaviour 8, 2293–2303] When Combinations of Humans and AI Are Useful: A Systematic Review and Meta-Analysis
- **Authors:** Michelle Vaccaro, Abdullah Almaatouq & Thomas Malone (MIT Center for Collective Intelligence)
- **Links:** [Original Preprint](https://arxiv.org/abs/2405.11896) · [Direct PDF](https://arxiv.org/pdf/2405.11896)
- **Core Formula:** `Hedges' g_synergy = −0.23 (Naive Decision Pairing)  vs.  g > 0 (Creation + Calibrated Gate)`
- **Finding:** Meta-analysis of 106 studies (370 effect sizes) proving that slapping a human rubber-stamp onto AI predictions often performs WORSE than AI or human alone (g = −0.23) due to automation bias, whereas structured division of labor yields positive synergy.
- **CFO Actionable:** Eliminate uncalibrated 'human rubber-stamping' of 100% of agent outputs; instead use selective deferral (EQ-12) so humans inspect only high-loss, high-uncertainty exceptions with blind cross-checks.

### PP-09 [arXiv:2101.05853 & arXiv:2204.14257] Algorithmic Monoculture, Social Welfare & Outcome Homogenization
- **Authors:** Jon Kleinberg & Manish Raghavan (PNAS 2021) · Rishi Bommasani, Kathleen A. Creel, Ananya Kumar, Dan Jurafsky & Percy Liang (NeurIPS 2022)
- **Links:** [Original Preprint](https://arxiv.org/abs/2101.05853) · [Direct PDF](https://arxiv.org/abs/2204.14257)
- **Core Formula:** `Welfare(Shared Base Model, μ + Δμ, ρ > 0)  <  Welfare(Independent Agents, μ, ρ = 0)`
- **Finding:** Proves a Braess-like paradox for foundation models: switching independent decision pipelines onto one slightly more accurate shared model reduces aggregate system reliability because errors synchronize.
- **CFO Actionable:** Enforce vendor Herfindahl caps (HHI_vendor ≤ 0.40) across critical finance, underwriting, procurement, and compliance workflows even if one model tops public leaderboards.

### PP-10 [NBER Working Paper No. 32487 · Economic Policy] The Simple Macroeconomics of AI
- **Authors:** Daron Acemoglu (MIT Department of Economics — Nobel Laureate in Economics)
- **Links:** [Original Preprint](https://www.nber.org/papers/w32487) · [Direct PDF](https://www.nber.org/system/files/working_papers/w32487/w32487.pdf)
- **Core Formula:** `ΔTFP ≈ s_exposed · s_feasible · π_cost_saving   (Easy-to-Verify ≫ Hard-to-Verify Tasks)`
- **Finding:** Applies Hulten's theorem to task-level AI adoption, demonstrating high, verifiable cost savings in structured 'easy-to-verify' workflows versus severe capital destruction when firms over-automate context-heavy, hard-to-verify decisions.
- **CFO Actionable:** Rank every AI capex request by objective verifiability: fund deterministic-checkable close, AP, SQL, and treasury reconciliations first; stage open-ended advisory agents via real options.

## Active Remote R&D Pods ($20–45/hr, 10–25 hrs/wk)
### POD-01 (MARKOWITZ-SIGMA-01) — Cross-Agent & Multi-Vendor Error Covariance (Σ) Econometrician / Quant ($25–$45 / hr USD)
- **Formula:** `Σ = B F_model Bᵀ + Δ_agent,   Var(w̄) = σ²[ρ + (1−ρ)/N]`
- **Scope:** Measure empirical error correlations ρ across LLM families, prompts, fine-tunes, and SaaS agent vendors; build Ledoit–Wolf shrinkage and factor covariance matrices Σ for CFO AI risk committees.
- **Tags:** Covariance Estimation, Shrinkage & Factor Models, Cross-Model Benchmarking, Python / R / SQL, Financial Econometrics

### POD-02 (MARKOWITZ-CVAR-02) — CVaR, Actuarial Tail-Risk & Single-Model Concentration Risk Manager ($25–$45 / hr USD)
- **Formula:** `min_w CVaR_α(L(w))   s.t.   wᵀμ ≥ μ*,  ‖Bᵀw‖_∞ ≤ κ_vendor`
- **Scope:** Quantify correlated tail losses, semivariance, copula stress scenarios, and maximum single-model exposure caps across autonomous finance, procurement, pricing, and customer-facing workflows.
- **Tags:** CVaR & Semivariance, Actuarial Loss Modeling, Vendor Concentration Caps, Monte Carlo Stress Testing, Enterprise Risk Management

### POD-03 (MARKOWITZ-CLOSE-03) — Agentic Month-End Close, Treasury Liquidity & Audit-Trail Risk Specialist ($20–$45 / hr USD)
- **Formula:** `min_w wᵀΣ_close w + η · P( |ΔLedger| > τ_materiality )`
- **Scope:** Design multi-model ensemble verification, orthogonal deterministic accounting checks, intraday cash-sweep liquidity buffers, and SOX/IFRS-compliant audit trails for agentic close and treasury.
- **Tags:** Agentic Month-End Close, Treasury & Cash Optimization, SOX / Audit Trails, Orthogonal Model Ensembles, ERP / Ledger Reconciliation

### POD-04 (MARKOWITZ-FPNA-04) — CFO Capital Budgeting, Real Options & AI Efficient-Frontier FP&A Analyst ($20–$45 / hr USD)
- **Formula:** `min_w wᵀΣw   s.t.   wᵀμ = μ*,   wᵀc ≤ Budget_CFO`
- **Scope:** Construct the CFO's AI Efficient Frontier across token spend, SaaS licenses, GPU commitments, human-in-the-loop review gates, and cash — pricing staged AI rollouts via real options and ROIC hurdles.
- **Tags:** Capital Budgeting & ROIC, Real Options Valuation, Strategic FP&A Modeling, Mean–Variance Optimization, AI Spend Governance

## Canonical Q&A
### Why does deploying 50 AI agents on one foundation model fail to diversify operational and financial risk?
Under equal weights (1/N), equal variance σ², and pairwise correlation ρ, Harry Markowitz's variance formula gives σ²[ρ + (1 − ρ)/N] → σ²ρ, so effective independent capacity is N_eff = N / (1 + (N − 1)ρ). Kim et al. (arXiv:2506.07962, ICML 2025) show LLMs agree up to 60% of the time when both are wrong — so at N = 50 and ρ = 0.60, fifty agents diversify like only N_eff ≈ 1.64 independent agents (collapsing to 1 as ρ → 1).

### Why is the sample covariance matrix Σ singular when a CFO has 50 agents and 30 days of evals?
With N = 50 agents and T = 30 daily evaluation snapshots, the empirical covariance matrix S_emp has rank ≤ T − 1 = 29 < 50 (det S_emp = 0). Optimizing directly on S_emp invents spurious 'zero-variance' agent combinations. Quantitative specialists apply Ledoit–Wolf shrinkage Σ̂_LW = (1 − δ*)S_emp + δ*νI_N and base-model factor models (e_i = β_i F_base + ε_i).

### How do deterministic scripts, multi-vendor evals (Anthropic, OpenAI, Google), and audit materiality work together?
Deterministic rules and SQL ledger checks act as an idealized risk-free leg (σ_0 = 0, Tobin separation), while Anthropic, OpenAI, and Google form a lower-ρ risky tangency portfolio verified via item-level paired evals. Statutory audit materiality M_audit serves as the firm's 99% CVaR risk budget (CVaR_99%(L(w)) ≤ M_audit), re-estimated and rebalanced on every model version release.

### How do CFO Offices engage Markowitz.systems to build an Efficient Frontier for AI spend?
CFOs, VPs of Finance, Corporate Treasurers, and Audit/Risk Committees contact frontier@markowitz.systems directly. Our specialists run paired multi-vendor evals, estimate shrunk covariance Σ̂_LW and downside semivariance Σ⁻, set banking Large-Exposure caps per foundation model, and construct an executable Capital Market Line and Mean–Variance Efficient Frontier.

### How do remote part-time R&D pods ($20–$45/hr, 10–25 hrs/wk) work for Financial Economists and Quants?
Quants, actuaries, risk managers, financial economists, and senior FP&A analysts join continuous remote R&D pods ($20–$45/hr USD, 10–25 hrs/wk) — applying mean-variance optimization, Ledoit–Wolf covariance estimation, semivariance, CVaR, Agent Sharpe Ratios, and real options to agentic month-end close, treasury, FP&A, and audit trails.

### Is Markowitz.systems affiliated with the Harry M. Markowitz estate, and how is candidate data governed?
Markowitz.systems is an independent research and quantitative talent platform and is not affiliated with, endorsed by, or connected to the Estate of Harry M. Markowitz, the Cowles Foundation at Yale University, or John Wiley & Sons. Credential intake is governed by IQ Vector Research Operations (privacy@markowitz.systems) under our published Privacy Policy (/privacy), excluding RU, IR, and KP.

## Sister Foundational Research Portals
- **[VAPNIK.AI](https://vapnik.ai/)** — *V. N. Vapnik & A. Ya. Chervonenkis (Statistical Learning Theory, 1971–1998)*: Structural Risk Minimization, VC-Dimension & Non-Asymptotic Generalization.
- **[WALRASIAN.COM](https://walrasian.com/)** — *Léon Walras · Arrow–Debreu · Mas-Colell, Whinston & Green (General Equilibrium)*: Compute Endowments, Shadow Prices & Risk-Adjusted Agent ROE Pruning.
- **[BELLMAN.AI](https://bellman.ai/)** — *Richard E. Bellman (Dynamic Programming, Princeton / RAND, 1957)*: Sequential Decision Processes, HJB / POMDP Control & Multi-Step Agent Horizons.
- **[CASELLABERGER.COM](https://casellaberger.com/)** — *George Casella & Roger L. Berger (Statistical Inference, 2002)*: Exact Finite-Sample Hypothesis Testing, Neyman–Pearson Gates & Paired Evals.
- **[ANDYGROVE.SYSTEMS](https://andygrove.systems/)** — *Andrew S. Grove (High Output Management, 1983)*: COO Hybrid Human + Agent Production Lines, Bottlenecks & Task-Relevant Maturity.

## Legal Notice, Independence Disclaimer & Privacy Policy
- Independent research & talent platform; not affiliated with, endorsed by, or connected to the Estate of Harry M. Markowitz, the Harry Markowitz Company, the Cowles Foundation for Research in Economics at Yale University, the Nobel Foundation, or John Wiley & Sons, Inc.
- Operated by IQ Vector Research Operations (Data Controller for markowitz.systems · D1 Credential Intake: d1.iqvector.ai · Privacy & Data Protection Officer: privacy@markowitz.systems · Security: security@markowitz.systems).
- **Full Privacy Policy:** https://markowitz.systems/privacy
