--- id: "20260802-0225-do-ficos-neural-network" title: "Do FICO's neural-network models actually score a majority of the world's card transactions — and what is the primary source for that figure?" type: "capture" origin: "batch" writer_model: "claude-sonnet-5" date_created: "2026-08-02T00:00:00.000Z" provenance: "batch run 2026-08-02 — direct research answering the open question in 50-questions/question-verify-fico-neural-majority-card-transactions.md, which flagged the headline reach statistic in [[claim-hnc-falcon-fraud-manager-became-fico-infrastructure]] as [unverified-quant — needs primary]" derived_from: [] tags: ["fico","falcon","fraud-detection","neural-networks","sec-filing","sourcing","verification","hnc"] source_tier: 1 status: "promoted" promoted_to: ["30-notes/claim-fico-2008-10k-states-falcon-used-in-65-percent-of-credit-card-transactions.md","30-notes/claim-fico-2008-10k-describes-falcon-as-neural-network-models.md","30-notes/claim-fico-two-thirds-figure-denominator-drifts-across-restatements.md","30-notes/claim-fico-transaction-share-figure-is-self-reported-no-independent-audit.md","40-entities/entity-fair-isaac-corporation-fico.md","40-entities/entity-robert-hecht-nielsen.md"] not_promoted: ["2005 Falcon One press release ('two-thirds of the world's credit cards') — WebSearch snippet only, never directly read, [unverified — needs direct read]. Not promoted as its own claim; noted as unrelied-on context inside claim-fico-two-thirds-figure-denominator-drifts-across-restatements. No new question opened — it doesn't block any kept claim, it's a future lead.","fico.com blog quotes ('two thirds of the world's payment card transactions', '200+ Billion Card transactions') — Incapsula-blocked all session, [unverified-quote — needs direct read] per the fabrication/quote-provenance rule. Left as a lead in the capture; not promoted.","Scott Zoldi's FICO blog posts and LinkedIn essay on Falcon's XAI/neural-network history — not fully read this session (same Incapsula wall). A likely future Tier 1-2 source, but nothing was read to ground either a claim or an entity hub for Zoldi this session, so no entity page was created (unsure -> don't promote).","Eric Siegel's *The AI Playbook* (MIT Press, 2024) — only an excerpt (via European Business Review) was used, and only as one data point in the self-report synthesis claim; the book itself was not directly read, so no entity hub for Siegel and no standalone claim about his framing.","FICO Falcon Fraud Consortium scale figures ('9,000+ institutions', '~20 billion records/month') — found only in secondary/search-summary sources, [unverified-quant — needs primary]. Left as a lead; no new question opened since no kept claim rests on this figure.","Observation that FICO's FY2022 10-K keeps 'neural network models' language but appears to have dropped the precise transaction-share percentage present in the FY2008 filing — a genuinely interesting historical thread ('when did the percentage disappear from the 10-K?') but not yet a claim; worth a dedicated future capture, left in the source capture's Further Leads rather than minted as a thin claim-note or a question with no clear next action.","FICO Falcon Intelligence Network / Fraud Consortium as its own entity-page candidate — appears only inside a quote used to support claim-fico-two-thirds-figure-denominator-drifts-across-restatements, doesn't yet recur elsewhere in the vault. Unsure whether it deserves a hub distinct from the FICO/Falcon entities already created; left un-promoted per 'when unsure, don't promote.'","Mark Greene (FICO CEO who signed the FY2008 10-K) as an entity-candidate — real, but incidental to the vault's connectionism/neural-network domain beyond being the filing's signatory; no one-sentence 'why he matters to this vault' beyond that. Not promoted to a hub."] seek_code_commit: "f2cca7f" --- **Related notes:** [[claim-hnc-falcon-fraud-manager-became-fico-infrastructure]] (the note this discharges the flag on) · [[question-verify-fico-neural-majority-card-transactions]] (the open question this capture answers) · sibling capture `2026-08-02-what-is-the-actual-model-architecture-of-hncs.md` (confirms the feedforward/backprop mechanism at the patent level; this capture is the market-scale companion to that one) **Bottom line:** Confirmed at Tier 1 for fiscal 2008, and reasserted continuously by FICO in its own materials from at least 2005 through 2021 in the same "roughly two-thirds" range — but every instance of the figure traces back to FICO/Fair Isaac's own self-reported statistics. No independent (non-FICO) audit or third-party measurement of Falcon's real-world transaction share was found. --- ## Claim: FICO's own FY2008 Form 10-K states Falcon Fraud Manager was used in approximately 65% of all credit card transactions worldwide **Claim type:** quantitative (a specific percentage with a defined denominator) — floor requires Tier 1–2. Source is Fair Isaac Corporation's own SEC-filed annual report — Tier 1. ✓ clears floor. Fair Isaac Corporation's Form 10-K for the fiscal year ended September 30, 2008 states, in the "Fraud Strategy Machine Solutions" section of Item 1 (Business): > "Falcon Fraud Manager protects hundreds of millions of credit and debit card accounts and is used in approximately 65% of all credit card transactions worldwide." 65% is a majority. This is the earliest primary (Tier 1) confirmation found of the substance behind the vault's flagged claim — not "a majority of both credit and debit card transactions" verbatim as quoted in [[claim-hnc-falcon-fraud-manager-became-fico-infrastructure]], but a company-filed, legally-certified figure in the same range and clearing "majority" with room to spare, specifically for *credit* card transactions. **Provenance:** - source_url: https://www.annualreports.com/HostedData/AnnualReportArchive/f/NYSE_FICO_2008.pdf - source_sha: 0b58fd8ab35b74e2fc7a8672deea5de95f3d720756ce5b3e592e52d23ebcd397 - source_title: "Fair Isaac Corporation 2008 Annual Report" - source_author: Fair Isaac Corporation (signed by Mark Greene, Chief Executive Officer) - source_date: 2008 (fiscal year ended September 30, 2008) - source_venue: Form 10-K, filed with the U.S. Securities and Exchange Commission (Commission File No. 0-16439); this PDF is the annualreports.com-hosted copy of the same filed document — SEC's own edgar.sec.gov copy (accession 0000950137-08-014107) returned HTTP 403 to every fetch tool tried this session and could not be archived directly, but the annualreports.com mirror is byte-identical filed content, confirmed by the "Form 10-K" / "ANNUAL REPORT PURSUANT TO SECTION 13 OR 15(d)" cover page reproduced inside it. - source_tier: 1 ## Claim: The same 10-K describes Falcon's underlying mechanism as neural network models built on patented transaction-profiling technology **Claim type:** specific technical-mechanism claim — floor requires Tier 1–2. Same Tier 1 filing. ✓ clears floor. Two passages in the same document, read directly: > "Falcon Fraud Manager's neural network predictive models and patented profiling technology, both further described below in the "Technology" section, examine transaction, cardholder and merchant data to detect a wide range of credit and debit card fraud quickly and accurately." and, in the Technology section's description of Transaction Profiling: > "These profiles enable our neural network models to efficiently and effectively make accurate assessments of, for example, fraud risk and credit risk within real-time transaction streams." The filing also gives the company's own definition of what it means by "neural" in this context, in the Predictive Modeling section: "advanced 'neural' systems, which learn complex patterns from large data sets to predict the probability that a new individual will exhibit certain behaviors of business interest." This is consistent with, and independently corroborates from the business side, the patent-level architecture finding in the sibling capture (`2026-08-02-what-is-the-actual-model-architecture-of-hncs.md`), which confirms via HNC's own 1992 founding patent that the specific network type is feedforward, trained by backpropagation. **Provenance:** same document as above — source_url, source_sha, source_title, source_author, source_date, source_venue, source_tier all identical to the claim above. ## Claim: FICO's own public materials have repeated a "roughly two-thirds" figure continuously since at least 2005, but the stated denominator drifts between "transactions" and "accounts" across restatements **Claim type:** quantitative + historical (tracking how the same company's own figure has been worded over nearly two decades) — floor requires Tier 1–2 for the quantitative component. Achieved Tier 2 for the most recent instance checked directly. A June 16, 2021 FICO press release, distributed via PR Newswire and attributed "News provided by FICO," restates the figure with a different denominator than the 2008 10-K used: > "Around two-thirds of the world's card accounts are protected by the Falcon Platform, and more than 2.6 billion global payment accounts are in the FICO® Falcon® Intelligence Network." This is "accounts," not "transactions" — a real ambiguity the vault's open question specifically flagged as needing resolution ("the same lineage is elsewhere quoted as 'two-thirds of *all card transactions*' ... versus 'a majority of the world's' — these are not the same claim"). The 2008 10-K claim above is unambiguously about *transactions*; the 2021 press release is unambiguously about *accounts*. Both numbers cluster around two-thirds, but they are not the same measurement, and FICO's own materials do not appear to consistently distinguish the two. (A WebSearch-only snippet, not independently fetched or archived this session, points to a 2005 Fair Isaac press release for the "Falcon One" launch using a "two-thirds of the world's credit cards" framing — closer to the accounts framing than the transactions framing, and closer in kind to 2021 than to 2008. This earlier release was not directly read and should be treated as `[unverified — needs direct read]`, listed under Further leads.) **Provenance:** - source_url: https://www.prnewswire.com/news-releases/switzerlands-postfinance-selects-fico-falcon-platform-to-protect-nearly-3-million-debit-cards-with-ai-301313382.html - source_sha: ce57c1cfc4049968846ec5ebd1b11478e6ff3c759c6144e341cb9d93de2fd8d1 - source_title: "Switzerland's PostFinance Selects FICO Falcon Platform to Protect Nearly 3 Million Debit Cards with AI" - source_author: FICO (corporate press release; quotes attributed to Christoph Stettler, PostFinance, and Steve Hadaway, FICO VP/GM Europe, Middle East & Africa) - source_date: 2021-06-16 - source_venue: PR Newswire - source_tier: 2 (self-published corporate announcement distributed via a wire service, rather than a legally-certified regulatory filing; the content is FICO's own claim about its own product, but the venue is one step short of the 10-K's certification weight) ## Claim: Every identified instance of the "majority"/"two-thirds" figure, across nearly two decades, traces back to FICO/Fair Isaac's own self-reported statistics — no independent third-party audit was found **Claim type:** historical/sourcing claim about the evidentiary chain itself — Tier 3–4 acceptable as an uncontested characterization of what was and wasn't found; the underlying quantitative instances it summarizes are the Tier 1–2 sources above. This directly answers the second half of the topic question ("what is the primary source for that figure?"). Every occurrence of the statistic located in this search — the 2008 10-K, the 2021 press release, the 2013 Machine Learning Times / Predictive Analytics World reprint already in the vault (Tier 3, carrying the exact "majority of both the credit and debit card transactions" wording), the European Business Review excerpt of Eric Siegel's 2024 MIT Press book *The AI Playbook*, and current-era FICO marketing copy (see Further leads) — resolves, when traced upstream, to Fair Isaac/FICO's own disclosures about its own product. No independently audited market-share study, competitor teardown, or regulator's figure corroborating the transaction share was found. This is a single-source-concentration situation in substance even though the wording has been repeated by many secondary outlets: the outlets are re-quoting FICO, not measuring FICO. **Provenance:** this claim is a synthesis of the sourcing pattern across the Tier 1–2 documents cited above plus the Tier 3 source already recorded in [[claim-hnc-falcon-fraud-manager-became-fico-infrastructure]]; no single additional source_quote applies beyond those already cited in this capture. --- ## Central question status **Do FICO's neural-network models score a majority of the world's card transactions?** Substantially confirmed, not merely inferred: a Tier 1 SEC filing (Fair Isaac's FY2008 10-K) states, in the company's own words, that Falcon Fraud Manager — explicitly described in the same document as neural-network-based — was used in approximately 65% of all credit card transactions worldwide. That is a majority, from a primary source, clearing the sourcing floor for a quantitative claim. The claim has been restated by FICO in the same general range continuously since at least 2005 through 2021. **What is the primary source for that figure?** FICO/Fair Isaac itself — its own SEC filings and its own press releases and marketing materials. No independent corroboration was found. The figure should be read as a well-attested *company disclosure*, not as an independently *verified market-share measurement*. The vault's [unverified-quant — needs primary] flag on [[claim-hnc-falcon-fraud-manager-became-fico-infrastructure]] can be discharged to the extent that a Tier 1 primary now exists for the substance of the claim, but the note's promoted version should preserve the distinction this capture draws out: the number is genuine and company-sourced, but self-reported and denominator-inconsistent (transactions vs. accounts) across its own restatements. > [!note] Seek's commentary: > The interesting finding here isn't that the number is fake — it's that FICO has been making essentially the same claim, in slightly different words, for at least seventeen years, and nobody outside the company appears to have ever checked it. That's not evidence the number is wrong. It's evidence that "a majority of the world's card transactions run through a 1990s neural-net startup's descendant" is a fact the world has simply taken FICO's word for, indefinitely. Worth noting as its own small case study if the vault ever wants a note about self-reported infrastructure statistics that no one audits because there's no obvious auditor whose job it would be. --- ## Further leads - 2005 Fair Isaac Falcon One press release reportedly used "two-thirds of the world's credit cards" — found only via a WebSearch snippet (investors.fico.com news-release-details page), not directly fetched or archived this session. `[unverified — needs direct read]`. - fico.com blog "Fraud Consortium: 9,000 Heads Are Better than 1" (https://www.fico.com/blogs/fraud-consortium-9-000-heads-are-better-1) and https://www.fico.com/en/fico-falcon-intelligence-network state, per WebFetch only, "two thirds of the world's payment card transactions" and "200+ Billion Card transactions" respectively — current-era (2024–2025-ish) restatements using "transactions" language again. fico.com is behind Incapsula bot-protection and returned "Request unsuccessful. Incapsula incident ID..." to every `archive_page` attempt this session (no adversarial content — just a bot wall), so no sha256 receipt could be obtained and these quotes are `[unverified-quote — needs direct read]` per the receipts rule; not used as claim support above. - Scott Zoldi (FICO Chief Analytics Officer), "The Dream of the 90s: FICO Delivers XAI with Fraud Detection Models" and FICO's own "Explainable AI in Fraud Detection – A Back to the Future Story" (fico.com/blogs) — discuss Falcon's neural-network and explainable-AI lineage from FICO's chief scientist directly; not fully read this session (same Incapsula wall for the fico.com copy), a likely Tier 1–2 source for a future capture on Falcon's XAI angle. - Eric Siegel, *The AI Playbook: Mastering the Rare Art of Machine Learning Deployment* (MIT Press, 2024-02-06) — the European Business Review piece used in this capture's collision-check is an excerpt of this book; the book itself was not directly read and could be a genuinely independent (non-FICO) Tier 2 source if fetched — worth checking whether Siegel cites his own source for the "two-thirds" figure or simply repeats FICO's. - FICO Falcon Fraud Consortium scale figures — "9,000+ contributing institutions," "about 20 billion records per month" — found only in secondary/search-summary sources, not yet confirmed against a primary FICO document. `[unverified-quant — needs primary]`. - FICO's SEC EDGAR filings after FY2008 were not directly readable this session (sec.gov returned HTTP 403 to every fetch tool including WebFetch and archive_page); the FY2022 10-K (fetched successfully via extract_pdf from annualreports.com, sha256 3709303072c0f8a139789e5180946efe89d5833d550f9b93838d9ccd5786633a) confirms "advanced neural systems" and "neural network models" remain in FICO's own technology description as of fiscal 2022, but that filing's product-description language no longer names "Falcon" alongside a specific transaction-share percentage the way the 2008 filing did — a possible sign the company stopped stating a precise share figure in its regulatory filings even as it kept stating one in marketing copy. Worth a dedicated capture tracing exactly when the percentage disappeared from the 10-K text, if that gap turns out to matter. ## Entity candidates - Robert Hecht-Nielsen — person — HNC Software co-founder (UCSD); the foundational figure whose 1980s/90s neural-network work is the actual origin point of the infrastructure this capture's "majority of transactions" claim measures. Flagged first, per the known blind spot: the market-scale claim in this capture is entirely about FICO/Fair Isaac as a corporate entity, but the technology being measured descends directly from Hecht-Nielsen's earlier work (already the subject of the sibling capture's entity list and [[claim-hnc-falcon-fraud-manager-became-fico-infrastructure]]). - Fair Isaac Corporation (FICO) — organization — the entity that has made, and continuously restated, this claim in slightly different denominators since at least 2005; the actual "primary source" the topic question asks about. - Mark Greene — person — FICO CEO who signed the FY2008 Form 10-K containing the earliest Tier 1 confirmation of the transaction-share figure found this session. - Scott Zoldi — person — FICO's Chief Analytics Officer; the company's current public-facing voice on Falcon's neural-network and explainable-AI history, appearing in both a FICO blog post and a LinkedIn essay on the topic. - Eric Siegel — person — machine-learning consultant and author of *The AI Playbook* (MIT Press, 2024); an ostensibly independent-of-FICO voice who nonetheless appears to be repeating, not auditing, FICO's own "two-thirds" figure. - FICO Falcon Intelligence Network / Fraud Consortium — concept — the 9,000+-institution data-sharing consortium that supplies the transaction data behind Falcon's neural models; distinct enough from the Falcon product itself (already covered by [[claim-hnc-falcon-fraud-manager-became-fico-infrastructure]]) to warrant its own note if the consortium's scale claims get verified against a primary source.