Cybernetic Meadow

---

Our first commercial application is a fraud investigation copilot for public company filings.

A statistical classifier ranks each filing against the historical fraud distribution. A parallel symbolic stream explains and corroborates that ranking at the level of specific accounting anomalies. A language model handles extraction of structured facts from unstructured filings. The two analyses are reported separately because they catch different cases — and for every flagged firm, the system assembles an investigation dossier: the evidence found, the accounting criterion each anomaly violates, and the review tasks a professional should perform to confirm or dismiss it. The dossier is the product.

In our empirical evaluation, the software identifies firms subsequently prosecuted by the SEC for accounting violations. Within a review budget of fifteen percent of the ranked population, it surfaces approximately nine in ten of those whose violations involve misstated financials. More importantly, it identifies the precise accounting mechanisms behind each flag — the specific disclosures, ratios, and patterns underpinning potential fraud or legal violations.

The architecture is patent pending. The software is in beta. We are engaging with forensic accountants, law firms, and financial institutions for feedback on our workflows.

 The architecture is domain-agnostic. It is built for problems with a specific shape: sparse consequential cases hidden in a large population of ordinary ones, adversarial patterns that evolve to evade the last generation of detection, and stakes that require every conclusion to be compelling. Statistical learning alone is structurally underpowered on this class of problem — the long tail of the probability distribution these cases live in is precisely where the training distribution is thinnest.

This architecture rests on a broader empirical position we develop in our research: that current AI systems have systematically misplaced the boundary between what statistical methods can reliably do and what still requires explicit human knowledge. Measuring where that boundary actually sits, per domain, is what the research is for. Building software that respects it is what the company is for. Accounting fraud is our first application. Rare-disease diagnosis, regulatory compliance in novel technology environments, and security threat detection share the same problem shape and are the natural extensions.