How can a business make better decisions when information is incomplete, outcomes are biased, capital is limited and markets keep changing? HyperC takes on these problems at the intersection of economics and machine learning. We combine economic theory, market simulations and business telemetry to develop foundational models for generalized arbitrage: finding and acting on differences in value across goods, places, time and operating capabilities, after costs and risk.
Economic theory explains why an apparent opportunity can disappear once information, selection and execution enter the picture. Our research uses these questions to shape data contracts, simulator assumptions and evaluation. The published work connects to specific contributions by Nobel laureates:
James Heckman studied how selective samples distort inference. A business sees outcomes for deals it accepted, while the outcomes of rejected deals remain missing. Our working paper connects this selection problem to the need for historical opportunity menus, explicit missing outcomes and evaluation beyond the old policy.
George Akerlof, Michael Spence and Joseph Stiglitz established how unequal information changes markets. Our market analysis draws on adverse selection and costly information: hidden quality, private histories and the cost of investigation help explain why prices differ and why an apparently cheap asset may be a bad purchase.
Paul Milgrom and Robert Wilson developed auction theory that accounts for different information and valuations. Wilson’s common-value models explain why winning can mean overpaying. This connects to our analysis of the winner’s curse and motivates tests of how bidding rules, competition and uncertainty affect a deal’s economics.
These ideas motivate the problems we test. HyperC’s contribution is the attempt to turn economic structure into models that can be evaluated against decisions and realized outcomes.
The Menu-Sales-Description contract connects available actions, realized cash flows and operating rules. It makes costs, timing, constraints and missing outcomes part of the learning problem.
Controlled simulations expose outcomes that real histories conceal. We use them to test selection bias, changing regimes and the cost of wrong decisions. Grounded simulations are hypotheses to check against telemetry.
P34/PARML is an early implementation of our broader foundation-model agenda: learn reusable economic decision structure, then align it to an operator’s market and constraints. How far that knowledge transfers remains an empirical question.
The API, Business Builder, managed workspaces and market tools connect the model to real workflows. Shadow tests, capped pilots and realized outcomes determine whether a proposed advantage survives execution.
Generalized arbitrage reaches beyond buying the same asset on two exchanges. It includes inventory, freight, capacity and other businesses where better information or lower operating costs can change a deal’s value. Every application still has to earn its case after costs, delays, losses and capital constraints.
The executed slower-market-waves notebook in the Computable Markets working paper (August 2026) tests how selection bias can turn a strong observed-data backtest into losses on a wider opportunity menu.
A tuned XGBoost classifier scored 0.9266 ROC-AUC on the business-observed holdout and showed +$61.7k in naive observed-data evaluation — then lost $417.4k applied to the full 20-week future opportunity menu it had never been forced to refuse.
P34 selected 115 trades, predicted +$2,404, realized +$2,250. The calibrated XGBoost baseline selected 177 trades and realized −$8,789 on $18,821 deployed. Live system invocation on a synthetic market — not a live-market trial.
Partial observability, selection bias, regime change, optimistic backtest traps. In the easy stationary/growing benchmark P34 preserved upside: +$228.6k vs the baseline's +$227.1k.
Disciplined refusal is the central mechanism: no-trade is a rewarded output, and false-positive control is explicit in the model class.
Company-reported figures. Note: not audited by a human licensed auditor. Detailed per-market performance packs (evaluated counts, rejection rates, predicted-vs-realized economics, win/loss distributions) are published per reference deployment as part of the evidence engine.
Explore when an AI-operated strategy looks like a business, whether its operating edge survives costs and uncertainty, and how much capital it needs to sustain itself. Compare four linked views with configurable formulas and assumptions.
Illustrative theory explorer. Economic sustainability, operating character and legal-role screens are shown separately; outputs are not measured P34 performance.
APP Open the interactive model → GIT hyperc-ai/business-investment-atlas — source, model checks & docsWhy a cost advantage has to survive the whole operation, why repeatability needs evidence, how starting capital changes what is feasible, and why economic sustainability and legal classification answer different questions.
NOTE Read the introductionThe operator-relative theory of computable markets, the Menu-Sales-Description (MSD) data contract, the Computable-Market Thesis, application templates for 10+ verticals, and a falsification protocol. Includes the revised notebook results above.
PDF Download the working paper (1.7 MB)The P34/PARML technical report: benchmark construction (partial observability, selection bias, regime change), baseline tuning, results and limitations — with notebooks. Presented at computablemarkets.com.
GIT hyperc-ai/p34-technical-report — notebooks & resultsProvisional patent filed on the P34 architecture: synthetically-trained universe selection, Pareto-front alignment, and the confidence interface. Multiple defensive publications published.
How to try to break P34: holdout menu construction, baseline calibration rules, refusal accounting, and what a falsifying result would look like. Ask your own AI to audit it. Published as section 11 of the working paper; a standalone evidence pack follows.
PDF Read the protocol in the working paperA robot that can pursue profit creates a much larger set of questions. How should it manage risk? What should it refuse to do? How much autonomy should it have? How should it account for consequences that never show up in its P&L? How should many such systems cooperate — or compete? HyperC treats these as engineering and economic research questions to be studied alongside profitability, not after autonomous business is deployed at scale.
Can the system create positive economic outcomes from biased, incomplete, changing real-world data?
Can it avoid strategies that look profitable until they destroy the business?
Can owners specify objectives, constraints, approvals and stop conditions that stay meaningful as autonomy grows?
Who captures the efficiency these systems create — and how should they behave toward customers, suppliers, workers and competitors?
What happens when a profitable action imposes costs that never enter the robot's objective?
Can autonomous businesses share data, coordinate capacity and create more value together than apart?
When many robots learn and compete at once, do margins compress? Does capital concentrate? What new structures emerge?
Who can inspect, constrain, pause, audit or change the objective of a system that participates in the economy continuously?
These remain open research questions. Our goal is to make them measurable through explicit objectives, inspectable simulations, controlled deployments and feedback from real economic outcomes.
P34 is for computable markets; it does not yet automatically prove that a market is computable.
If the deployment-time menu is far from training context, an explicit refusal path is not yet exposed.
The business must have had more options historically than it chose — otherwise there is insufficient policy-selection signal.
Latency ranges from under a minute to hours; very high-frequency applications are outside current scope.
Generalization across all markets is not fully known; pilots proceed through shadow testing and controlled deployment.
Unknown unknowns remain — P34 is an early release in a new model class.