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

hncfalconficofraud-detectionneural-networksbackpropagationfeedforwardpatentsdeep-learning-historyverification

Related notes: observation-falcon-helmholtz-inference-embedding-false-friend (the note this gates) · claim-hnc-falcon-fraud-manager-became-fico-infrastructure · claim-rhw-1986-demonstration-not-invention · entity-paul-werbos · entity-geoffrey-hinton

Bottom line: confirmed at Tier 1. HNC Software's own founding patent for the fraud-scoring system — filed September 8, 1992, the same year Falcon Fraud Manager launched, and whose own figures literally label a screen "FALCON Monitor" — explicitly describes the network as a "feed forward" neural network trained by "backpropagation gradient descent optimization," using a supervised training loop (labeled fraud/non-fraud examples, output compared against a desired/target output, weights adjusted to reduce error). This directly discharges the [unverified-mechanism] flag on observation-falcon-helmholtz-inference-embedding-false-friend.

All quotes below were extracted via extract_pdf from the patent PDF (TLS verified) and checked for verbatim grounding with quote_check before being recorded. The patent's OCR text is a two-column layout that pdftotext interleaves line-by-line (left-column and right-column text land on the same output line), so quotes here are kept to short fragments that survive fully intact on a single physical line; connective narration between fragments is paraphrase, not quotation, and is marked as such.

• source_url: https://patents.google.com/patent/US5819226A/en (PDF fetched at https://patentimages.storage.googleapis.com/84/19/be/a7501246e89234/US5819226.pdf)
• source_sha: d4d140f1c06181682339474604bf194ff2c3dcbfe741f6073561885b23cb607
• source_title: "Fraud Detection Using Predictive Modeling"
• source_author: Krishna M. Gopinathan, Louis S. Biafore, William M. Ferguson, Michael A. Lazarus, Anu K. Pathria, Allen Jost (named inventors); assignee HNC Software Inc., San Diego, CA
• source_date: 1992-09-08 (filed, Appl. No. 941,971); patent granted 1998-10-06
• source_venue: U.S. Patent No. 5,819,226 (USPTO; accessed via Google Patents, treated per sources.md house convention as a primary-document mirror for reproduced grant text)
• source_tier: 1

Claim: HNC's founding patent for the system explicitly names the network type as "feed forward"

Claim type: Specific technical-mechanism claim → floor requires Tier 1–2. Source is HNC's own patent filing describing its own invention — Tier 1. ✓ clears floor.

The patent's "Network Training" section states the architecture directly:

"technique known as a "feed forward" neural network."

This is the exact phrase the vault's flagged claim needed: the patent, in describing the preferred embodiment used to build the fraud-scoring model, names the network class as feedforward (input → hidden → output, no recurrent connections), not a generative, recurrent, or Hopfield-style architecture.

Claim: The patent specifies backpropagation gradient descent as the training method, applied in a supervised loop

Claim type: Specific technical-mechanism claim → floor requires Tier 1–2. Same Tier 1 patent source. ✓ clears floor.

Immediately following the feedforward identification, the patent names the preferred training algorithm:

"is called "backpropagation gradient descent"

The patent separately describes the training procedure in three numbered steps, framed as standard supervised learning — labeled examples in, a model output compared against a known/desired output, and an error-driven weight update:

"1) Repeatedly presenting examples of a particular input/"

"2) Comparing the model output and desired output to"

"3) Modifying model weights to reduce the error."

(Full connective sentence, paraphrased from the surrounding non-quoted text: this cycle repeats until further iteration fails to decrease the error, at which point the network is described as "trained." Training examples are drawn from a labeled fraud database — accounts already known to have had fraudulent activity versus a sampled set of non-fraud accounts — which is the supervised-learning setup: the network is trained against known fraud/non-fraud labels, not learning an unsupervised generative model of transaction structure.) Together with the feedforward claim above, this confirms the specific "supervised backpropagation feedforward" characterization that observation-falcon-helmholtz-inference-embedding-false-friend had flagged as unverified.

Claim: The patent is HNC's own technical description of the Falcon product specifically, not an unrelated generic filing

Claim type: Historical/identifying claim connecting a specific patent document to a specific named commercial product. Uncontested once the primary document is read directly — Tier 1 (the patent's own drawings). ✓ clears floor.

The patent's own Figure 2 (a sample system-monitor screen forming part of the invention's output interface) is captioned, in the patent's own text:

"FALCON Monitor"

Combined with the filing date (September 8, 1992 — the year HNC's Falcon Fraud Manager is independently dated to in claim-hnc-falcon-fraud-manager-became-fico-infrastructure) and the assignee (HNC Software Inc., San Diego), this ties the patent directly to the Falcon product rather than to some other, unrelated HNC filing. This is the piece that upgrades the architecture claim from "a plausible HNC patent" to "the patent describing Falcon itself."

Claim: The architecture is multilayer — one or more hidden layers between input and output, not a single-layer perceptron

Claim type: Specific technical-mechanism claim → floor requires Tier 1–2. Same Tier 1 patent source. ✓ clears floor.

Describing Figure 10 (a diagram of the network's hidden processing elements), the patent states processing elements sort into three categories — input, output, and hidden — and that the hidden layer is not necessarily singular:

"Typically there are several such"

"layers of hidden elements."

(Paraphrased connective content, not quoted: inputs are supplied to a layer of input processing elements, whose outputs pass to one or more layers of hidden elements, which in turn pass outputs to a layer of output elements that produce the final fraud-score value — the standard multilayer feedforward topology, consistent with the "backpropagation" framing above, since single-layer perceptrons cannot be trained by backpropagation in the way described.)

Further leads

• Ghosh, S. & Reilly, D.L., "Credit card fraud detection with a neural-network," Proc. 27th Hawaii International Conference on System Sciences (HICSS), 1994, vol. 3, pp. 621–630 (DOI 10.1109/HICSS.1994.323314) — an HNC-authored academic paper on the same fraud-scoring line of work. A WebSearch summary (not independently fetched or read — a search-summary layer, per the fabrication rule) described this paper's network as a "three-layer, feed-forward Radial Basis Function (RBF)" model, which would be a materially different training regime (RBF centers are typically not trained by backprop) than the backprop-MLP described in the 1992 patent. This is [unverified-mechanism — needs direct read; possible contradiction, not confirmed] and should not be treated as fact until the actual IEEE paper is fetched and read directly.
• U.S. application Ser. No. 07/814,179, "Neural Network Having Expert System Functionality," Curt A. Levey, filed Dec. 30, 1991 — cross-referenced inside US5,819,226 as a related HNC application; not read in this pass, may describe an earlier or different architecture predating the Falcon-branded patent.
• Same patent family, not separately read: WO1994006103A1, CA2144068A1, EP0669032A1 ("Fraud detection using predictive modeling(ing)") — foreign/international filings of essentially the same content; low-priority corroboration only.
• FICO's own retrospective, "Explainable AI in Fraud Detection – A Back to the Future Story" (fico.com/blogs) — FICO's own account of Falcon's neural-network history; not fetched this pass, a likely Tier 1–2 lead for the still-open scale-claim gap tracked at claim-hnc-falcon-fraud-manager-became-fico-infrastructure.
• The patent's own background citations for "a more detailed description of neural networks": D. E. Rumelhart et al., "Learning Representations by Back-Propagating Errors," Nature v. 323, pp. 533–36 (1986), and R. Hecht-Nielsen, "Theory of the Backpropagation Neural Network," in Neural Networks for Perception, pp. 65–93 (1992) — both cited by name inside the patent as the source of the backprop technique it applies commercially.

Entity candidates

• Robert Hecht-Nielsen — person — HNC co-founder; his own 1992 book chapter "Theory of the Backpropagation Neural Network" is cited by name as a background reference inside his own company's founding Falcon patent. This is the foundational figure the architecture claim in this capture actually rests on (per the vault's backprop-lineage cluster: claim-rhw-1986-demonstration-not-invention, entity-paul-werbos) — flagged first, per the known blind spot, ahead of the patent's own named inventors below.
• Rumelhart, Hinton & Williams (1986), "Learning Representations by Back-Propagating Errors," Nature — concept/paper — the other named foundational reference inside the same patent; already tracked in the vault via claim-rhw-1986-demonstration-not-invention and entity-geoffrey-hinton. This patent is a concrete, dated, commercial downstream application of exactly that lineage — worth a link from that cluster to this capture's promoted note.
• Krishna M. Gopinathan — person — first-named inventor of US Patent 5,819,226, the founding Falcon-era HNC patent.
• HNC Software Inc. — organization — assignee of the patent; referenced repeatedly across the vault's Falcon-cluster notes but has no dedicated entity page of its own yet.
• Curt A. Levey — person — inventor of the cross-referenced earlier related HNC patent application (filed Dec. 1991); possible earlier-architecture figure, unverified in this pass.