White paper · v2 · 2026-09
Cascade risk 2–3: methodology
HTML summary on this page. PDF is gated: name, email, company, consent. No class Brier fiction.
Executive summary
Most risk tools stop at hop 1 (news → price). Axelrod's Oracle builds calibrated 2nd–3rd order scenarios on a knowledge graph. It is not a buy/sell call. Operator: IP Burlakov A. A.
1. Why hop 1 is not enough
Markets often price the first print. The book still eats hop 2–3: corridors, sectors, suppliers, regulatory edges. LLMs summarize what is already known. They do not replace a structured cascade.
2. Architecture (what we actually run)
Data layer + knowledge graph + self-hosted Gemma (no third-party LLM API for production scenarios) + calibration after the window. We do not claim SOC 2 or a public Kubernetes SLA on this page.
3. Five steps
Monitor → graph paths (2–3 hops) → A/B/C draft → human review on high severity → score when the window closes. Full walk-through: /produkt/metodologija.
4. How to read the scores
Overall Brier on a mixed book can sit near a coin flip — that is disclosed. Short windows can make ROC-AUC unreadable when all outcomes are one class. The named product anchor is Model A risk AUC 0.6339 (Dual Model cards on /statistika). We do not invent per-class Brier tables here. We do not print “losses prevented”.
5–6. Use, don’t worship
Funds: committee language for hop 2–3. Corporates: supplier-region paths. Banks: analytical input, not a filing. Honest retro cases (lags) live on /keysy.
10. Disclaimer
Not investment advice. Past scores do not guarantee future scores. Capital can be lost. See /disclaimer and /risk-warning.
Request the PDF
Same pipeline as other platform leads. We will not invent a download that does not exist yet.