What is the actual model architecture of HNC's Falcon Fraud Manager — is it specifically a supervised backpropagation feedforward network?

The capture 2026-07-11-hop-falcon-helmholtz-inference-bridge and the note observation-falcon-helmholtz-inference-embedding-false-friend characterize HNC's 1992 Falcon Fraud Manager as a supervised backpropagation feedforward neural network doing real-time discriminative scoring. That architectural detail is currently [unverified-mechanism]: it was derived from "contemporary descriptions of Falcon," but the only quote actually captured — T.J. Horan (FICO / KDnuggets, Tier 2) — documents transaction profiling and real-time scoring, not the network topology or training method.

Why it matters

The false-friend verdict (Falcon discriminative vs. the Helmholtz Machine generative) does not depend on the exact architecture — scoring P(fraud | transaction) is discriminative by definition regardless of net shape. But the note asserts more than "discriminative": it asserts a specific supervised backprop feedforward design. A specific technical-mechanism claim requires a Tier 1–2 source (sources.md sourcing floor). Until one is read, the architecture claim should stay flagged and the note should stay seedling.

What to read

• HNC / Hecht-Nielsen primary technical writing on Falcon — HNC patents on the fraud-scoring system, or Hecht-Nielsen's own descriptions (he wrote the 1989 textbook Neurocomputing).
• Original Falcon technical papers or FICO product/whitepaper documentation describing the model class and training procedure.
• Cross-check against independent Tier 1–2 accounts of how 1990s card-fraud neural scorers were actually trained (supervised vs. unsupervised; feedforward/backprop vs. other).

Related

• observation-falcon-helmholtz-inference-embedding-false-friend — the note this gates.
• claim-hnc-falcon-fraud-manager-became-fico-infrastructure — the lineage note (its own scale-stat gap is tracked separately at question-verify-fico-neural-majority-card-transactions).