Research

The Autonomous White-Collar Worker

Measuring the Economic Frontier of Claude Code, Computer-Using Agents, and Autonomous Knowledge Work

Builds an experimental laboratory in which a controller launches independent Claude Code sessions (claude-opus-5-5) on four Mac nodes to complete ten validated professional assignments in ten occupations under five randomised levels of computer access: over 150 pilot runs, scenario-weighted acceptance is 0.00 without a shell and 0.83–0.90 with one, the shell premium is +0.87 [0.67, 1.00] while the preregistered full-computer-use contrast is a precise null (+0.03 [0.00, 0.10]), and the marginal cost per accepted deliverable is 0.87–1.11 USD; the human arms were designed but not executed and the benchmark is agent-authored.

RAID: A Metal-Accelerated C++ Framework for Reproducible High-Frequency Financial Econometrics

Batched, Fused and Streaming Estimation on Apple Silicon without Sacrificing Inference

Asks whether realized measures, HAR and GARCH estimation can move to Apple's FP32-only GPUs without losing accuracy, inference or reproducibility: a C++20 library with compensated double-float Metal kernels matches FP64 within 10⁻⁶, bitwise reproducibly, and measured break-even points on one year of 1-minute bars for 1,000 US equities show where the GPU wins and where the CPU does.

The Invisible Balance Sheet

Can Intraday Financial Markets Detect Corporate Deterioration Before Accounting Disclosures?

Asks, on a strictly point-in-time panel of 3,796 U.S. firms and 102,522 firm-quarters built from 6.7 billion one-minute bars, whether 1-minute market data forecast adverse operating-income surprises better than daily data: fundamentals do most of the work, daily bars add −0.0068 in log loss and one-minute bars add nothing (+0.0013, +0.0016), a precise null across horizons, targets, placebos and unseen firms.

The Economic Geography of Digital Retail Supply: Product Diversity, Regional Specialization, and the Hidden Structure of Quebec's Online Economy

Evidence from 390,108 product listings of 1,569 Quebec online stores across 17 administrative regions (Fabri-Ka snapshot, 1 November 2026)

Tests, under a preregistered design with merchant-reshuffling nulls and three adversarial referee rounds, whether online catalogues of 1,569 Quebec stores reveal regional specialization: the signal is about who sells, not how much is listed (Cramér's V 0.112 vs 0.086 under merchant-equal weights), catalogues are extremely concentrated (Gini 0.85) and the complexity indicator fails validation and is withheld.

The Hidden Architecture of Grocery Prices: Promotions, Retail Competition, and the Cost of Living in Quebec

Evidence from 300,613 posted-price events across 54 Quebec grocery feeds, August–November 2026

Assembles an auditable panel of 124,579 grocery listings from 54 Quebec feeds (300,613 price events) and shows that identical products differ by a median 0.154 log points across banners, that promotions account for half of all price changes and double measured volatility, that sales return exactly to their reference price, and that naive indices are off by 10–30 points for banners with assortment turnover; publishes a five-component Quebec Grocery Price and Affordability Index.

The Hidden Dynamics of Used-Car Markets: Price Discovery, Inventory Liquidity, and Strategic Dealer Behavior

Evidence from a Daily Panel of 273,265 Quebec Listings

Follows 273,265 Quebec dealer listings about seven times a day for eleven weeks: depreciation is nonlinear, one listing in five is marked down (median 2.8%) with a hazard that doubles over three weeks, same-day repricing co-moves among neighbours without any lagged reaction, regional premia are within 4% of Montréal, and a five-component Automobile Market Liquidity Index explains 56% of the common variation in regional turnover.

The Social Compression Hypothesis

A Pre-Registered Test of Coordination Demand, Institutional Capacity, and Instability in Historical and Modern Societies, with a Formal Framework and Agent-Based Experiments

Reads institutions as compression devices that lower the communication cost of coordination, proves five theorems and runs 6,048 agent-based experiments, then tests the empirical corollary on Seshat (626 polities), the Ethnographic Atlas and a 14,637 country-year panel under a frozen pre-registration: the demand–capacity mismatch is associated with instability (0.20 and 0.39 log-odds per SD) but adds nothing to scale or capacity alone, so the universal theory is not supported.

Beyond the Square-Root Barrier: Automated Discovery of Algebraic and Algorithmic Shortcuts in Bitcoin Cryptography

A hypothesis-driven, reproducible study of secp256k1, ECDSA, BIP340 Schnorr and SHA-256 with machine-checked identities and controlled negative results

Runs a six-stage autonomous discovery pipeline (25 hypotheses, exact symbolic checks, 6,240 synthetic ECDLP instances, SAT cryptanalysis of reduced SHA-256, Lean 4 proofs) against Bitcoin's primitives and finds no shortcut: the secp256k1 structure is fully explained by complex multiplication by Z[ω], Pollard rho matches 1.254√n, batch Schnorr verification saturates at 3.1×, and sixteen reproductions, six observations and three rejected conjectures are released with full provenance.

TITAN-MoE: Topology-Aware and Predictive Expert Orchestration for Distributed Mixture-of-Experts Inference on Apple Silicon

A Measured Systems Study on a Heterogeneous Mac Cluster

Builds and measures TITAN-MoE, an MLX-native mixture-of-experts runtime with bounded expert residency, NVMe loading, causal prefetch and remote execution on a 21-node Mac cluster; bit-identical to mlx-lm at 489 vs 494 tok/s, but prefetch, kernel fusion and the gigabit remote tier do not pay (preliminary draft)

Zero-Forcing Bounds for Generalized Petersen Graphs: Computational Discovery and Certified Proofs for P(n,5)

Determines the zero forcing number of the generalized Petersen graph P(n,5) for every n ≥ 11 and proves that Z(P(n,5)) = 12 for all n ≥ 23, using the first-force boundary method, 129 drat-trim-verified DRAT certificates and exhaustive enumeration of 432 billion subsets.

PRISME: Intervention-Supervised Sparse Workspace Editing for Small Language Models

Design, controlled evaluation, and negative and positive evidence at the 100M-parameter scale

Implements, trains and tests PRISME, a sparse 16-slot latent workspace with a counterfactually supervised editing controller, against matched baselines; a preregistered negative result: sparse edits cut a frozen backbone's loss by 2.8%, a global-recurrence adapter does as well, and the learned controller is worse than no routing.

Beyond Text: Cache Compression and Knowledge Transfer in Multi-Agent Systems

A Reproducible Study of Fidelity, Communication Cost, and Repeated Handoffs

Measures 22 ways of passing information between language-model agents (text, summaries, structured evidence, quantized and evicted KV caches) on Qwen2.5-3B and Llama-3.2-3B over 13,280 trials with measured bytes and latency: lossless representations tie in quality, and lossy ones are model- and regime-dependent.

Protocol-Guided Autonomous Research with Claude Code

Scientific Verification, Reproducibility, and the Cost of Delegation: A Controlled Study

Four-arm controlled study of 288 live Claude Code runs (claude-opus-5-5) on 24 bounded scientific tasks scored by hidden deterministic oracles: verified completion is 0.903 under an open instruction and 0.917 under a full protocol with executable gates and review, a null +0.014 (95% interval [−0.028, +0.056]) at two to three times the cost.

Beyond One-Step Distributed Maximum Likelihood: Local Bias Cancellation, Global Synchronization, and the Price of Inexact Curvature

A joint order–round frontier, its sharpness in an exact model, and measured communication costs

Distributed maximum likelihood with order-r split-sample jackknife debiasing on K workers followed by q synchronized Newton rounds: first-order equivalence to the centralized estimator survives far more workers than the classical square-root limit (a half-block jackknife plus one step tolerates K up to N to the 7/8), with sharpness in the exponential model, Monte Carlo and cluster benchmarks.

The Invisible Hand of Machines: Price Discovery in Markets Populated by LLM Agents

Randomized matched-seed experiments in an event-sourced limit-order-book simulator assign 0–100 percent of trader slots to a small reasoning model (gpt-5.4-nano): quoted spreads fall from 19 to 6 basis points and no bubble occurs in 112 live runs, but returns become predictable (lag-one autocorrelation 0.02 to 0.32).

Can Autonomous AI Agents Discover the Fundamental Value of Real Estate? Evidence from Quebec Transactions and Multi-Agent Deliberation

Preregistered experiment on 631,866 registered Quebec sales (2021–2026) testing whether debate among twelve role-specialised LLM valuation agents adds information: debate reaches a mean absolute log error of 0.158 against 0.159 for the same agents kept silent, and every preregistered contrast includes zero.

When AI Agents Become Economic Agents: An Agentic General Equilibrium Theory

Delegating decisions to agents built on shared foundation models turns idiosyncratic error into common model noise: closed-form loadings in the delegation share and number of models, a zero private–social wedge under constant returns, and order-flow correlation rising from 0.11 to 0.97 in an agentic market laboratory.

From Labor to Capital: Artificial General Intelligence, General Equilibrium, and Asset Valuation

A general-equilibrium extension of Restrepo (2025) in which compute is a priced, depreciating, energy-using capital good alongside structures and land; calibrated to the United States and solved to 2300, the labor share falls to 0.283 by 2050 while compute's market value never exceeds one fifth of GDP.

Beyond Hedonic Models: Spatiotemporal, Hierarchical and Probabilistic Architectures for Residential Real Estate Price Forecasting

A pseudo-real-time evaluation on 632,000 Quebec transactions and 82 local market indices, 2021–2026

A strict pseudo-real-time test of five forecasting architectures and an integrated stack against 14 benchmarks on 632,006 Quebec sales (2021–2026): gradient boosting is the most accurate model for future sale prices at every horizon, and no architecture improves on momentum and autoregression for index changes.

SHARP-HPI: A Spatially Hierarchical, Adaptive and Robust Framework for Quality-Adjusted Housing Price Indices

Monthly municipal indices for Quebec from 770,000 registered sales, 2021–2026

A two-stage hedonic and state-space index estimator with empirical-Bayes pooling across cells, applied to 769,995 registered Quebec sales (2021–2026); in Monte Carlo it reduces the RMSE of municipal monthly changes to 1.58 log points against 2.91 for per-cell state-space estimation.

The Global Evidence on Housing Price Determinants: A Meta-Analysis of Hedonic Pricing Models, 1970–2026

A PRISMA-style meta-analysis of the global hedonic housing literature, 1970–2026: 103,011 records screened to 242 eligible studies from 55 countries and 18,831 coefficient rows, giving a pooled living-area elasticity of 0.65 and heterogeneity beyond sampling error for every implicit price.

From Listing Images to Property Valuation: A Multimodal AI Framework for Automated Building Characterization and the Cost Approach

Evidence from Québec single-family listings matched to the assessment roll

Vision–language models read 14,245 photographs and the text of 1,200 Québec house listings matched to the assessment roll to specify the improvements for a cost approach: the indication's ratio to asking price moves from 0.75 to 0.95, a modest gain next to a 29.9% unit-cost-anchor effect.

The Endogenous Housing Liquidity Premium: Search Frictions, Seller Urgency, and Price Discovery

A continuous-time search model with free entry of buyers, continuously distributed seller carrying costs and Nash bargaining derives the urgency discount, sale hazards, price dispersion and the cash–mortgage gap in eight propositions; in the illustrative baseline the most urgent seller concedes 5.6%. No data are used.

Housing, Leverage, and Intergenerational Wealth Inequality: A Dynamic General Equilibrium Theory

A closed-form three-stage benchmark and a twelve-period overlapping-generations model with nominal mortgages, gifts and estates: transfers create a 0.95 income-year owner–renter wealth gap among equal-income households, prices absorb 64–66% of the down-payment effect, and a 10% inflation surprise lowers the real house price by 0.45%.

Does the Sales Comparison Approach Actually Work? A Large-Scale Point-in-Time Validation of Comparable Selection, Price Adjustments, and Value Reconciliation

Evidence from 391,000 Single-Family Transactions in Quebec, 2021–2026

A point-in-time backtest of 2,256 configurations of the sales comparison approach on 390,996 Quebec single-family sales: adjustments cut the five-comparable median error from 18.1% to 13.0%, and the best transparent rule reaches 11.1% out of sample against 9.0% for gradient boosting.

Faster, Wider, More Fragile? The Evolution of Intraday Shock Propagation in U.S. Equity Markets, 1992–2026

Using 7.69 billion one-minute bars on every listed and delisted U.S. stock from 1992 to 2026, isolated ±6σ shocks raise market-wide tail activity by at most about 2% of its base rate in every decade, while transmission to the twenty closest correlation neighbours roughly tripled.

Hedonic House Pricing with Language-Model and Vision Features: Evidence from Quebec Real-Estate Listings

A large language model and a vision–language model convert the descriptions and façade photographs of 19,123 Quebec house listings into schema-constrained hedonic regressors, raising in-sample R² from 0.717 to 0.755 (3.0 of the 3.8 points from the photograph) and out-of-sample R² for every estimator.

VQuant: An Agentic Large Language Model Architecture for Automated Quantitative Financial Analysis and Document Generation

Reasoning-Augmented Tool Orchestration, In-Loop Numerical Computation, and Multi-Format Reporting

A design-science description of VQuant, a deployed platform in which a reasoning LLM orchestrates 245 schema-defined tools and a Python sandbox so that every number in a financial report comes from live data or auditable code; audited against the source code, with no empirical evaluation yet.

Forecasting Realized Volatility Across Asset Classes: HAR, GARCH, and Machine Learning in a Unified Multi-Asset Horse Race, 2013–2026

Twenty-seven realized-variance forecasts (HAR, GARCH, penalized regressions, tree ensembles, LSTM, Transformer, combinations) raced on 25 instruments from four asset classes over 2013–2026 under one protocol: log-linear models win, and ridge on an extended predictor set beats HAR on 24 of 25 assets (working draft).

Half a Million Prices, Twenty Models

A Systematic Assessment of Hedonic Specifications and Estimation Methods for the Quebec Housing Market, 2021–2026

A controlled horse race of twenty hedonic models on 514,212 roll-matched Quebec sales (2021–2026) under one retransformation rule: spatial controls move median error by ten points, functional form by two, and boosting's 3.2-point lead under random validation shrinks to an insignificant 0.5 point in a forward-in-time holdout.

The Assessment Gap in Quebec

Vertical and Horizontal Inequity in Municipal Property Valuation: Evidence from 522,769 Sales Matched to the Assessment Roll

The first province-wide ratio study for a Canadian province matches 522,769 Quebec sales (2021–2026) to 625 municipal assessment rolls; the measured regressivity depends on outlier treatment, with a within-market elasticity of −0.34 under a lenient trim but −0.09 under the IAAO rule.

A Grand Hedonic Model of the Canadian Housing Market

Decomposing the Value of Structure and Location across 142,121 MLS Listings with High-Dimensional Neighbourhood Fixed Effects

A national semi-log hedonic model on 142,121 Canadian MLS listings with 1,155 absorbed neighbourhood fixed effects raises explained log-price variance from 44% to 78%; a Shapley–Owen decomposition attributes about 34 points to structure and 36 to location, and out-of-sample MdAPE is 15.6%.

Prediction Markets as Arrow–Debreu Securities: Evidence from Polymarket

State Prices, No-Arbitrage Bounds and Macroeconomic News in Blockchain-Based Contingent Claims

Polymarket negative-risk markets on the June 2026 FOMC decision (146 daily closes) and the 2024 presidential election (304 closes) are read as Arrow–Debreu state prices: the complete set averages 1.017 and 1.015, within executable bounds, with no daily lead–lag against Treasuries, equities, the VIX or Bitcoin.

The Options-Implied Information Content for Cross-Asset Return and Volatility Prediction: Evidence from 3.8 Billion Option Contracts

On a 2010–2025 panel of 69 liquid US underlyings built from 3.83 billion option records, the implied volatility surface explains only 0.3% of weekly stock returns but raises the HAR-RV R² for weekly realized variance from 0.421 to 0.515; the earlier weekly return results are withdrawn.

Intraday Volatility Transmission Across Asset Classes: A Connectedness Analysis of Equities, Foreign Exchange, Commodities, and Cryptocurrencies

Diebold–Yilmaz connectedness of daily realized volatility built from 5-minute bars for ten instruments on 2,514 common trading days (2016–2026): cross-asset shocks explain 54.8% of forecast-error variance, U.S. equity futures are the only robust net transmitters, and bitcoin becomes a net receiver in the spot-ETF era.

Airbnb, Residential Rents, and Housing Market Pressure: A Hedonic and Spatial Econometric Analysis

Hedonic rent equations on 8,303 Quebec rental listings and 3,456 Airbnb listings (2026) show each Airbnb within 500 m associated with 0.4% higher rent under city fixed effects, but only 0.08% and insignificant under spatial inference once postal-code neighbourhood fixed effects are added.

Hedonic Pricing of Short-Term Rental Properties: A Multi-Method Analysis of Airbnb Listing Prices in Quebec

Hedonic, quantile, spatial and machine-learning models of Quebec Airbnb nightly prices from a scrape of 5,000 listing cards, after a two-regime mixture removes the 28% of displayed prices that are multi-night stay totals; a city-fixed-effects model explains 51% of log-price variance.

Hedonic Housing Price Models for the United States: A Multi-Method Comparison of Parametric, Quantile, and Machine Learning Approaches

OLS, quantile regression and gradient boosting are compared on 788,842 Zillow listings across all 50 states; XGBoost reaches R² = 0.833 under random validation but 0.425 when ten states are held out, and OLS is as accurate out of state (0.539) once corrupt records are removed.

Decoding Real Estate Descriptions: Semantic Embeddings and Hedonic Pricing of Residential Properties in Quebec

Sentence embeddings of 17,087 Quebec single-family listings are projected onto 20 named reference descriptions and added to a hedonic model; adjusted R² rises from 0.452 to 0.511, but net of listing language and 415 area fixed effects the semantic block adds only 2.0 percentage points.

Does the Language of Artificial Intelligence Move Markets? A Semantic Analysis of Russell 3000 SEC Filings and the Cross-Section of Stock Returns, 2010–2026

Scores AI language in 527,548 SEC filings of about 2,500 Russell 3000 firms with sentence embeddings and links it to forward returns over 27,468 firm-years: a modest persistent premium (+5.5 pp per year, t = 2.41); the post-ChatGPT jump seen in the S&P 500 is a large-cap re-rating, not a market-wide regime change.

Heterogeneous Amenity Capitalization in the Quebec Residential Market: A Spatial Quantile Regression Approach

Quantile and two-stage spatial quantile regressions on 42,481 Centris listings (January 2026) show that, for houses, the highway capitalisation differential falls from 0.205 at the 10th percentile to 0.034 at the 90th, and a spatial lag attenuates median differentials by a third to a half.

From Corporate Disclosures to a Russell 3000 Strategic Project Database: Identification, Measurement, Validation, and Research Applications

Documents SPID, a project-level database built from 527,548 SEC filings (10-K, 10-Q, 8-K, 2010–2026) of 2,564 Russell 3000 issuers: a schema-constrained LLM extractor yields 308,065 evidence-linked project mentions resolved into 121,099 candidate projects, with a validation framework stated but not yet executed.