Research & evidence

Built on evidence, not demos

HyperC sits between a research lab and a product company: working papers, a proposed data contract for the category, a defined falsification protocol — and a commercial product measured in customer-realized profit. This page keeps the evidence inspectable, with synthetic and live results clearly separated.

How to read this page. Synthetic stress test results come from controlled benchmark markets with known ground truth — they demonstrate mechanism, not live profitability. Live deployment results are company-reported production outcomes — note: not audited by a human licensed auditor unless stated. Historical results do not guarantee future outcomes.
Notebook results

The selection-bias trap, measured

From the executed slower-market-waves notebook in the Computable Markets working paper (August 2026). The point is not that P34 wins — it's why conventional models lose.

Synthetic stress test
0.9266 AUC → −$417.4k
The conventional baseline's great backtest, then reality

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.

Synthetic stress test
+$2,250 on $4,146
P34, executed REST-API run on the withheld week-80 menu

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.

Synthetic stress test
+$14.8k vs −$219.8k
Difficult benchmark: P34 vs tuned conventional baseline

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.

Synthetic stress test
99.4%
of evaluated orders rejected in the published study

Disciplined refusal is the central mechanism: no-trade is a rewarded output, and false-positive control is explicit in the model class.

Production

Live deployments

$30M+
sales generated for customers, >95% of trades unsupervised
Company-reported, cumulative since 2023
~$100M/yr
reseller operated with limited supervision
Thousands of signals; POs and shipping orders generated
3,000+
loans issued by the model in a live lending test
Micro-credit vertical, 2025

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 supported market as part of the evidence engine.

Publications

Papers, benchmarks & code

Working paper · Aug 2026

Computable Markets: Business Menus, Sales Event Tapes, and Cross-Market Profit-Directed Learning

The 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
Technical report · Jun 2026

P34: Learning When Not to Trade

The P34/PARML technical report: benchmark construction (partial observability, selection bias, regime change), baseline tuning, results and limitations — with notebooks.

GIT hyperc-ai/p34-technical-report — notebooks & results
Patent

Menu-structured trade selection with many-worlds model ensembles

Provisional patent filed on the P34 architecture: synthetically-trained universe selection, Pareto-front alignment, and the confidence interface. Multiple defensive publications published.

Protocol

Evaluation & falsification protocol

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.

PDF Evidence pack — drop file to publish
Limitations

What we tell investors and customers up front

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.