{ "$comment": "Entrenchment case coding dataset. Every source_url was found via live search; scores are one author's coding against the fixed rubric in index.html, not independently verified.", "rubric_version": "v1", "coded_date": "2026-07-18", "cases": [ { "category": "Legacy public sector", "name": "IRS Individual Master File", "domain": "Tax administration", "description": "The IRS's core taxpayer account database, built by IBM in the 1960s, still runs Assembly language code tied to System/360-derived mainframe architecture and processes accounts for over 100 million taxpayers; a Java based successor only entered testing in 2024, six decades after go live.", "source_url": "https://www.nextgov.com/modernization/2024/05/irs-making-headway-modernizing-1960s-era-tax-system-commissioner-says/396695/", "source_name": "Nextgov/FCW, May 2024", "dependence_score": 1.0, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 3 }, "rationale": { "integration_depth": "Core system of record for federal individual tax accounts nationwide.", "switching_cost": "A replacement was only entering testing in 2024, six decades after the original went live.", "revenue_dependence": "Directly processes the tax data underlying federal income tax collection.", "substitute_scarcity": "Code is tied to System/360-derived architecture with no production ready substitute as of the source date." } }, { "category": "Legacy public sector", "name": "Social Security Administration COBOL benefit systems", "domain": "Social Security administration", "description": "SSA runs roughly 60 million lines of COBOL, much dating to the 1960s-70s, to calculate and pay Social Security benefits; a 2017 modernization plan stalled for lack of funding, and a 2025 push to rewrite the code quickly drew warnings that small COBOL to modern language conversion errors could mispay beneficiaries.", "source_url": "https://www.computerworld.com/article/3953741/doge-wants-to-modernize-social-securitys-legacy-tech-what-could-possibly-go-wrong.html", "source_name": "Computerworld, 2025", "dependence_score": 0.917, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 2 }, "rationale": { "integration_depth": "60 million lines of COBOL are described as enabling the agency's core regulatory and benefit functions.", "switching_cost": "The 2017 modernization plan was never completed for lack of funding.", "revenue_dependence": "A migration error is reported as capable of causing incorrect benefit payments.", "substitute_scarcity": "A cited cost benefit study found full replacement risk outweighs the benefit of keeping the system." } }, { "category": "Legacy public sector", "name": "DoD Strategic Automated Command and Control System", "domain": "Military command and control", "description": "SACCS runs on a 1970s era IBM Series/1 computer and, until 2019, used 8-inch floppy disks to transmit nuclear launch orders and coordinate US nuclear forces; as of 2019 the Air Force had only swapped the floppy disks for solid state storage, with a full replacement still undefined.", "source_url": "https://www.defensenews.com/air/2019/10/17/the-us-nuclear-forces-dr-strangelove-era-messaging-system-finally-got-rid-of-its-floppy-disks/", "source_name": "Defense News, October 2019", "dependence_score": 1.0, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 3 }, "rationale": { "integration_depth": "Described as the messaging backbone through which nuclear launch orders reach all US nuclear forces.", "switching_cost": "Only a hardware swap was achieved by 2019; a full system replacement had no announced timeline.", "revenue_dependence": "Failure would directly threaten US nuclear command and control, the source's own framing of the stakes.", "substitute_scarcity": "The article notes the Air Force was seeking a replacement with no timeline or specifics." } }, { "category": "Legacy public sector", "name": "State unemployment insurance COBOL systems", "domain": "Unemployment insurance administration", "description": "Many US state unemployment systems, including New Jersey's, were built in COBOL in the 1970s and never fully replaced; when COVID-19 produced a 1,600% surge in claims in April 2020, the decades old system could not keep up, prompting a public appeal for COBOL programmers.", "source_url": "https://slate.com/technology/2020/04/new-jersey-unemployment-cobol-coronavirus.html", "source_name": "Slate, April 2020", "dependence_score": 0.917, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 2 }, "rationale": { "integration_depth": "The COBOL system is the backbone processing all statewide unemployment claims and payments.", "switching_cost": "The system, built in the 1970s, was never fully replaced despite decades of known obsolescence.", "revenue_dependence": "System overload directly threatened the state's ability to pay unemployment benefits during the pandemic surge.", "substitute_scarcity": "Reporting found few COBOL volunteers were actually deployed despite public appeals, leaving a practical workforce gap even though the language itself is portable." } }, { "category": "Legacy public sector", "name": "FAA air traffic control legacy systems", "domain": "Air traffic control", "description": "A 2023 FAA risk assessment found 37% of its 138 air traffic control systems unsustainable, some up to 50 years old, with replacement parts no longer manufactured; a January 2023 outage in 30-year old NOTAM software triggered the first nationwide ground stop since 9/11, and GAO found some modernization projects will not complete for another 10-13 years.", "source_url": "https://www.gao.gov/products/gao-24-107001", "source_name": "US GAO, GAO-24-107001, 2024", "dependence_score": 0.917, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 2 }, "rationale": { "integration_depth": "138 interlinked systems underpin the national airspace; the NOTAM outage cascaded into a nationwide ground stop.", "switching_cost": "GAO found some modernization projects will not complete for 10 to 13 years.", "revenue_dependence": "58 of the flagged systems have critical operational impacts on safety, and the 2023 failure grounded all US flights.", "substitute_scarcity": "Replacement parts are extremely limited and manufacturing support has ended for some equipment, though a slow modernization path exists." } }, { "category": "Legacy public sector", "name": "VA VistA electronic health records and the paused Oracle Cerner replacement", "domain": "Veterans healthcare records", "description": "The VA signed a $10 billion 2018 deal to replace its homegrown VistA system, used at over 1,500 sites, with Oracle Cerner; the rollout has been repeatedly paused after inspector general findings tied it to four veteran deaths, and by 2024 only 6 of 170-plus sites had converted while lifecycle cost estimates rose to $37.2 billion.", "source_url": "https://www.techtarget.com/searchhealthit/news/366578278/VA-to-Use-Legacy-VistA-EHR-for-Up-to-10-Years-Amid-Paused-Oracle-Cerner-EHRM", "source_name": "TechTarget, 2022; VA OIG report, March 2024", "dependence_score": 0.833, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 1 }, "rationale": { "integration_depth": "VistA is deployed at over 1,500 VHA sites as the core infrastructure for veteran healthcare delivery.", "switching_cost": "After seven years and roughly $5 billion spent, only 6 of 170-plus sites had converted, with cost estimates rising to $37.2 billion.", "revenue_dependence": "The inspector general tied the replacement system directly to four veteran deaths.", "substitute_scarcity": "A substitute exists and is in active, if troubled, deployment." } }, { "category": "Legacy public sector", "name": "VA Benefits Delivery Network", "domain": "Veterans benefits administration", "description": "BDN, a suite of 1970s COBOL mainframe applications processing about $40 billion a year in veteran compensation and pension payments, was formally targeted for retirement in 2003 but was only fully decommissioned in September 2025, twenty two years later, after a replacement finally matured.", "source_url": "https://govciomedia.com/digital-gi-bill-automation-speeds-va-benefits-delivery/", "source_name": "GovCIO Media & Research, 2025; GAO-07-614", "dependence_score": 0.833, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 1 }, "rationale": { "integration_depth": "Described as VBA's primary claims processing, tracking, and payment infrastructure for roughly $40 billion a year.", "switching_cost": "Targeted for retirement since 2003, full decommissioning was not achieved until 2025.", "revenue_dependence": "Directly processed roughly $40 billion annually in veteran benefit payments.", "substitute_scarcity": "A working substitute was eventually built and took over, so a substitute was identified, just slow to mature." } }, { "category": "Legacy public sector", "name": "CMS Medicare fee for service claims processing", "domain": "Medicare claims administration", "description": "Medicare's core claims engines (FISS, MCS, DME, and the Common Working File), built in the 1970s on COBOL, process roughly 1.2 billion claims a year; a 1990s unified replacement attempt by CMS's predecessor was terminated, and only individual calculation tools have since been modernized while the COBOL core remains untouched, with a request for proposals to replace it only issued in 2026.", "source_url": "https://www.discoveriesinhealthpolicy.com/2026/06/follow-up-cms-and-ancient-cobol-claims.html", "source_name": "Discoveries in Health Policy, June 2026; GAO-01-824", "dependence_score": 1.0, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 3 }, "rationale": { "integration_depth": "Processes roughly 1.2 billion claims a year and integrates with multiple downstream Medicare payment systems.", "switching_cost": "A 1990s full replacement attempt was terminated; only piecemeal modernization has succeeded since.", "revenue_dependence": "Processes hundreds of billions of dollars in Medicare payment flow annually.", "substitute_scarcity": "CMS was only at the request for proposal stage for a core replacement as of 2026." } }, { "category": "Legacy public sector", "name": "HMRC legacy VAT and tax mainframe systems", "domain": "Tax administration (UK)", "description": "HMRC's VAT processing relies on legacy systems, including COBOL running on aging mainframe hardware, part of roughly 600 IT systems HMRC itself calls complex, ageing, and costly; the agency has publicly described a dangerous dependency on the legacy mainframe operating system and pursues only a hybrid modernization rather than full replacement.", "source_url": "https://www.theregister.com/on-prem/2025/08/15/why-the-uk-public-sector-creaks-along-on-cobol/", "source_name": "The Register, August 2025", "dependence_score": 0.833, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 2, "revenue_dependence": 3, "substitute_scarcity": 2 }, "rationale": { "integration_depth": "VAT collection, a core UK tax revenue stream, runs on the mainframe HMRC itself calls a dangerous dependency.", "switching_cost": "HMRC has adopted only a hybrid approach, modernizing peripheral components while the core VAT system stays intact.", "revenue_dependence": "VAT is a major UK revenue stream, and HMRC's own language ties the mainframe risk directly to it.", "substitute_scarcity": "Rather than replace the core, HMRC recruits and retains COBOL specialists, indicating no adopted substitute yet." } }, { "category": "Financial infrastructure", "name": "COBOL in US core banking systems", "domain": "Core banking infrastructure", "description": "Roughly 43% of US banking systems still run COBOL, a language over 60 years old, underpinning an estimated $3 trillion in daily commerce and most ATM and in person transactions; banks pay premium contractor rates to retired specialists to patch undocumented legacy code rather than rewrite it.", "source_url": "https://www.cnbc.com/2017/04/11/banks-scramble-to-fix-old-systems-as-it-cowboys-ride-into-sunset.html", "source_name": "Reuters via CNBC, April 2017", "dependence_score": 0.917, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 2 }, "rationale": { "integration_depth": "Underpins core account, ledger, ATM, and teller processing across the majority of US banks.", "switching_cost": "Institutions pay premium wages to keep decades old systems running rather than replace them.", "revenue_dependence": "An estimated $3 trillion in daily commerce flows through these systems.", "substitute_scarcity": "Full rewrites at this scale are treated as infeasible; scarce human expertise substitutes for a technology replacement." } }, { "category": "Financial infrastructure", "name": "SWIFT global payment messaging network", "domain": "Cross border payments messaging", "description": "SWIFT connects over 11,000 financial institutions in 200-plus countries and is embedded in correspondent banking and sanctions compliance; when Iran was disconnected it lost roughly half its oil export revenue, and despite eight plus years of preparation, Russia's SPFS alternative still lacks 24-hour international connectivity.", "source_url": "https://www.cnn.com/2022/02/25/business/swift-russia-putin/index.html", "source_name": "CNN Business, February 2022", "dependence_score": 1.0, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 3 }, "rationale": { "integration_depth": "Embedded in correspondent banking and sanctions screening across 11,000-plus institutions.", "switching_cost": "Russia spent eight plus years building an alternative before 2022 and it still could not substitute internationally.", "revenue_dependence": "Iran's disconnection precedent cost roughly half its oil export revenue.", "substitute_scarcity": "China's CIPS is reported at a small fraction of SWIFT's size; no adequate substitute has emerged." } }, { "category": "Financial infrastructure", "name": "Fedwire Funds Service", "domain": "Large value payment settlement", "description": "Fedwire, the Federal Reserve's real time settlement system, moves over $3 trillion daily among 8,900-plus participants; a February 2021 operational error knocked out Fedwire, ACH, Check21, and FedCash simultaneously for roughly four hours nationwide, showing nearly all major US payment rails share common Fed infrastructure with no independent fallback.", "source_url": "https://www.cnbc.com/2021/02/24/the-feds-system-that-allows-banks-to-send-money-back-and-forth-is-down.html", "source_name": "CNBC, February 2021", "dependence_score": 0.917, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 2, "revenue_dependence": 3, "substitute_scarcity": 3 }, "rationale": { "integration_depth": "The settlement system relied on by 8,900-plus institutions and government agencies for same day payments.", "switching_cost": "No migration away has ever been attempted or documented; institutions have no alternative rail.", "revenue_dependence": "Over $3 trillion transferred daily; the 2021 outage froze several payment rails together for hours.", "substitute_scarcity": "When Fedwire went down, ACH, Check21, and FedCash also failed since they share the same infrastructure." } }, { "category": "Financial infrastructure", "name": "Knight Capital 2012 trading algorithm failure", "domain": "Trading and market making infrastructure", "description": "In August 2012, Knight Capital deployed new trading software to eight servers but missed one, which retained dormant 2003 code that a repurposed flag bit reactivated; in 45 minutes the firm sent 4 million erroneous orders, lost $440 million, roughly four times its annual profit, and was forced into a fire sale acquisition days later.", "source_url": "https://www.sec.gov/newsroom/press-releases/2013-222", "source_name": "US SEC press release, October 2013", "dependence_score": 0.75, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 2, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 1 }, "rationale": { "integration_depth": "Dormant 2003-era code shared a flag with the production system for eight years without being isolated or removed.", "switching_cost": "The 2012 deployment was itself an attempt to retire the legacy code path, and it failed catastrophically.", "revenue_dependence": "The failure destroyed the firm's capital base within 45 minutes and forced a fire sale acquisition.", "substitute_scarcity": "Source documents the failure mechanism but not alternative systems considered." } }, { "category": "Financial infrastructure", "name": "TSB Bank core migration to Proteo4UK", "domain": "Core banking platform migration", "description": "After Banco Sabadell acquired TSB, it migrated 5.2 million UK customers off a legacy Lloyds inherited platform onto its own Proteo4UK system in April 2018; the migration corrupted an estimated 1.3 billion customer records, exposed some customers to others' data, took eight months to fully resolve, cost an estimated 330 million pounds, and drew a 48.65 million pound regulatory fine for inadequate migration governance.", "source_url": "https://www.fca.org.uk/news/press-releases/tsb-fined-48m-operational-resilience-failings", "source_name": "UK FCA and PRA joint press release, December 2022", "dependence_score": 0.833, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 1 }, "rationale": { "integration_depth": "The legacy platform ran all branch, telephone, online, and mobile banking for 5.2 million customer accounts.", "switching_cost": "The migration corrupted 1.3 billion records, took eight months to resolve, and cost roughly 330 million pounds.", "revenue_dependence": "Regulators fined TSB 48.65 million pounds and it paid 32.7 million pounds in customer redress.", "substitute_scarcity": "The failure stemmed from the chosen vendor's delivery capability rather than an absence of any substitute platform." } }, { "category": "Financial infrastructure", "name": "RBS, NatWest, and Ulster Bank 2012 IT failure", "domain": "Core banking batch processing", "description": "In June 2012, a botched software upgrade to RBS Group's batch processing system corrupted the overnight payment queue, locking over 6.5 million UK customers out of banking services for weeks; the FCA and PRA jointly fined the banks 56 million pounds for lacking adequate systems and controls to manage IT risk.", "source_url": "https://www.fca.org.uk/news/press-releases/fca-fines-rbs-natwest-and-ulster-bank-ltd-%C2%A342-million-it-failures", "source_name": "UK FCA press release, November 2014", "dependence_score": 0.75, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 2, "revenue_dependence": 3, "substitute_scarcity": 1 }, "rationale": { "integration_depth": "The batch system controlled core payment processing across three major retail banks simultaneously.", "switching_cost": "RBS later spent an estimated 750 million pounds overhauling the IT at fault, but as post hoc remediation rather than a documented abandoned migration.", "revenue_dependence": "Resulted in 56 million pounds in combined regulatory fines plus compensation costs.", "substitute_scarcity": "Source focuses on risk control failure, not the absence of alternative systems." } }, { "category": "Financial infrastructure", "name": "Visa and Mastercard card network duopoly", "domain": "Card payment networks", "description": "Visa and Mastercard together control roughly 80% of the US card network market and charged merchants over $100 billion in card fees in 2023; the DOJ's September 2024 antitrust suit alleges Visa uses exclusionary agreements with merchants, banks, and processors to block competitors from developing viable debit network alternatives.", "source_url": "https://www.justice.gov/archives/opa/pr/justice-department-sues-visa-monopolizing-debit-markets", "source_name": "US DOJ press release, September 2024", "dependence_score": 0.833, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 2, "revenue_dependence": 2, "substitute_scarcity": 3 }, "rationale": { "integration_depth": "Card acceptance infrastructure is built around these two networks across the vast majority of US card transactions.", "switching_cost": "DOJ alleges structural agreements actively prevent switching, though this is an active suit rather than a completed migration.", "revenue_dependence": "Merchants pay over $100 billion a year with no negotiating power, though no source frames this as a threat to a named institution's own survival.", "substitute_scarcity": "DOJ's complaint specifically alleges Visa acted to prevent others from developing viable alternatives." } }, { "category": "Financial infrastructure", "name": "DTCC clearing and settlement infrastructure", "domain": "Securities clearing and settlement", "description": "DTCC's three subsidiaries are among only eight entities nationally designated as Systemically Important Financial Market Utilities under Dodd Frank, meaning regulators formally determined their failure could spread significant liquidity or credit problems through the US financial system; DTCC safeguards over $114 trillion in securities.", "source_url": "https://home.treasury.gov/policy-issues/financial-markets-financial-institutions-and-fiscal-service/fsoc/designations", "source_name": "US Department of the Treasury, FSOC designations, Dodd Frank Title VIII", "dependence_score": 0.833, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 1, "revenue_dependence": 3, "substitute_scarcity": 3 }, "rationale": { "integration_depth": "Three subsidiaries are formally designated systemically important infrastructure for nearly all US securities clearing.", "switching_cost": "No source describes an attempted migration away from DTCC.", "revenue_dependence": "Federal regulators formally determined DTCC's failure would threaten US financial system stability.", "substitute_scarcity": "Only 8 entities nationally hold this designation, reflecting no redundant clearing system at this scale." } }, { "category": "Financial infrastructure", "name": "BNY Mellon and SunGard InvestOne 2015 outage", "domain": "Fund administration and custody infrastructure", "description": "BNY Mellon outsourced net asset value calculation for mutual funds and ETFs to SunGard's InvestOne platform; during a routine operating system upgrade in August 2015 the production environment corrupted, and the intended backup environment corrupted too, preventing automatic failover and producing inaccurate valuations for up to 1,200 funds industry wide.", "source_url": "https://www.finextra.com/newsarticle/27780/sungard-apologises-for-bny-mellon-system-glitch", "source_name": "Finextra, 2015", "dependence_score": 0.667, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 1, "revenue_dependence": 2, "substitute_scarcity": 2 }, "rationale": { "integration_depth": "A single outsourced platform calculated valuations for up to 1,200 funds across the industry.", "switching_cost": "No evidence in sources that a migration away from SunGard was ever attempted.", "revenue_dependence": "Disrupted valuation and trading for client funds industry wide, though direct revenue loss to BNY Mellon is not quantified in the source.", "substitute_scarcity": "The intended backup and failover system itself corrupted and failed, showing the built in redundancy was not a working substitute." } }, { "category": "Financial infrastructure", "name": "Nasdaq matching engine failure during the Facebook IPO", "domain": "Stock exchange trading infrastructure", "description": "A design limitation in Nasdaq's IPO cross matching system caused its calculated price and volume to fall 19 minutes behind incoming orders during Facebook's May 2012 IPO, leaving over 38,000 marketable orders unexecuted for hours; the SEC fined Nasdaq $10 million, its largest ever penalty against an exchange, for poor systems and decision making.", "source_url": "https://www.sec.gov/newsroom/press-releases/2013-2013-95htm", "source_name": "US SEC press release, May 2013", "dependence_score": 0.583, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 1, "revenue_dependence": 2, "substitute_scarcity": 1 }, "rationale": { "integration_depth": "The IPO cross matching engine is core infrastructural plumbing for Nasdaq listed IPOs and secondary trading.", "switching_cost": "No evidence of an attempted migration away from Nasdaq's matching engine infrastructure.", "revenue_dependence": "Caused a record $10 million SEC penalty and reputational harm, though not framed as an existential threat.", "substitute_scarcity": "Source does not address whether issuers considered alternative exchanges." } }, { "category": "Enterprise vendor lock-in", "name": "Lidl's abandoned SAP Retail implementation", "domain": "ERP for retail merchandise management", "description": "Lidl spent seven years and roughly 500 million euros implementing SAP Retail to replace its decades old in house merchandise system across 10,000-plus stores; SAP's standardized inventory logic clashed with Lidl's lean, highly customized operating model, and in July 2018 management concluded the goals could not be met without spending far more, killing the project and reverting to the legacy system.", "source_url": "https://www.computerweekly.com/news/252446965/Lidl-dumps-500m-SAP-project", "source_name": "Computer Weekly, July 2018", "dependence_score": 0.667, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 2, "switching_cost": 3, "revenue_dependence": 2, "substitute_scarcity": 1 }, "rationale": { "integration_depth": "SAP Retail was slated to become the core merchandise management backbone, but the legacy system it was meant to replace stayed operational throughout.", "switching_cost": "A seven year, 500 million euro migration effort was abandoned outright in 2018.", "revenue_dependence": "The write off and leadership turmoil were reported as major financial damage, framed as a costly failed investment.", "substitute_scarcity": "A working substitute existed and was used, Lidl's own pre existing system, once SAP was abandoned." } }, { "category": "Enterprise vendor lock-in", "name": "Hershey's 1999 SAP, Manugistics, and Siebel rollout", "domain": "ERP, supply chain, and CRM", "description": "In 1999 Hershey deployed a combined $112 million system, SAP R3, Manugistics, and Siebel, compressed into 30 months instead of a recommended four years to beat the Y2K deadline, going live just before its two biggest sales seasons; the integrated system could not translate orders into shipments even though inventory existed in warehouses, and Hershey could not deliver about $100 million of Halloween and Christmas orders.", "source_url": "https://www.cio.com/article/270245/supply-chain-management-supply-chain-hershey-s-bittersweet-lesson.html", "source_name": "CIO.com retrospective", "dependence_score": 0.667, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 2, "revenue_dependence": 3, "substitute_scarcity": 0 }, "rationale": { "integration_depth": "The three systems jointly formed the entire order to fulfillment chain.", "switching_cost": "Hershey absorbed roughly a year of degraded operations rather than rolling the system back, implying unwinding it was itself prohibitive, though no formal abandonment is documented.", "revenue_dependence": "Hershey could not deliver about $100 million of orders and its stock dropped over 8% on disclosure.", "substitute_scarcity": "Source does not discuss alternative software or vendors." } }, { "category": "Enterprise vendor lock-in", "name": "Waste Management's terminated SAP ERP project", "domain": "ERP", "description": "In 2005 Waste Management contracted SAP for an industry specific waste and recycling ERP package SAP allegedly pitched as tested and ready to deploy; Waste Management later alleged in litigation that no such tested software existed, terminated the project in 2007 after spending over $100 million, sued for fraud, and settled in 2010 for an undisclosed sum reported around $80 to $100 million.", "source_url": "https://www.computerworld.com/article/1570731/waste-management-sues-sap-over-erp-implementation.html", "source_name": "Computerworld, 2008 and 2010", "dependence_score": 0.75, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 2, "switching_cost": 3, "revenue_dependence": 2, "substitute_scarcity": 2 }, "rationale": { "integration_depth": "Intended as the unified backbone for finance, billing, and customer service, but terminated before full deployment.", "switching_cost": "After spending over $100 million, Waste Management terminated the project outright in 2007.", "revenue_dependence": "The suit sought recovery of over $100 million in expenses plus promised annual savings, framed as vendor fraud rather than an existential threat.", "substitute_scarcity": "Central to the fraud claim was that SAP had marketed a uniquely tailored package that did not exist anywhere in tested form." } }, { "category": "Enterprise vendor lock-in", "name": "US Air Force's terminated Expeditionary Combat Support System", "domain": "ERP for military logistics", "description": "Launched in 2004-2005, ECSS was a SAP based ERP meant to consolidate roughly 240 separate Air Force logistics systems into one platform; after seven years and over $1.03 billion spent, the Air Force determined finishing even a quarter of the scope would cost another $1.1 billion, and terminated the program in 2012 having delivered no significant military capability.", "source_url": "https://spectrum.ieee.org/us-air-force-blows-1-billion-on-failed-erp-project", "source_name": "IEEE Spectrum, 2012", "dependence_score": 0.833, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 1 }, "rationale": { "integration_depth": "Explicitly designed to consolidate 240 separate legacy systems into a single infrastructural platform.", "switching_cost": "Terminated after $1.03 billion and seven years when completion was found infeasible even with another $1.1 billion.", "revenue_dependence": "Tied to a Congressional mandate for auditable Air Force financial statements the service had missed since 1947.", "substitute_scarcity": "Source does not directly assess alternative vendor options." } }, { "category": "Enterprise vendor lock-in", "name": "Revlon's disrupted SAP S/4HANA rollout", "domain": "ERP", "description": "In February 2018 Revlon went live with SAP S4HANA across a large part of its North American operations; the changeover disrupted manufacturing at a key plant, left Revlon unable to ship to major retail customers, produced roughly $64 million in unfulfilled sales and $53.6 million in remediation charges, delayed its annual report over a related material weakness, and preceded its CFO's resignation and a shareholder lawsuit.", "source_url": "https://www.techtarget.com/searcherp/news/252464165/Revlon-SAP-ERP-problems-result-in-rare-investor-lawsuit", "source_name": "TechTarget/SearchERP, 2019", "dependence_score": 0.5, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 2, "switching_cost": 1, "revenue_dependence": 3, "substitute_scarcity": 0 }, "rationale": { "integration_depth": "Governed manufacturing and shipping at a key plant and a large part of North American operations.", "switching_cost": "Revlon remediated the system in place rather than abandoning or reverting it.", "revenue_dependence": "Disclosed $64 million in unfulfilled sales, a delayed annual report, and a CFO resignation directly tied to the rollout.", "substitute_scarcity": "Source does not discuss alternative ERP vendors." } }, { "category": "Enterprise vendor lock-in", "name": "UK Post Office Horizon accounting system", "domain": "Retail point of sale and accounting", "description": "Horizon, built by Fujitsu, has run the UK Post Office's roughly 11,500 branches since 1999-2000 and was central to the wrongful prosecution of over 900 subpostmasters for shortfalls later proven to be software faults; a 2015 IBM led replacement attempt collapsed and reverted to Fujitsu, a follow on project was also axed, and full replacement contracts worth roughly 410 to 500 million pounds were only awarded in 2026, with Fujitsu not fully exited until 2027.", "source_url": "https://www.theregister.com/paas-and-iaas/2026/05/21/years-after-uk-post-office-scandal-broke-accenture-and-oneview-commerce-bag-contract-to-replace-horizon/", "source_name": "The Register, May 2026", "dependence_score": 0.833, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 2, "substitute_scarcity": 2 }, "rationale": { "integration_depth": "Underpins transaction processing and accounting across the entire 11,500-branch network since 1999.", "switching_cost": "A 2015 replacement attempt collapsed, a follow on project was cancelled, and full replacement was only contracted in 2026.", "revenue_dependence": "Operationally essential to branch payments and accounting, with major reputational and legal liability from the scandal.", "substitute_scarcity": "It took two vendors and a combined roughly 450 million pound contract, following two prior failed attempts, to assemble a substitute." } }, { "category": "Enterprise vendor lock-in", "name": "Queensland Health's IBM and SAP payroll system", "domain": "Government payroll and HR", "description": "Queensland Health contracted IBM to replace its legacy payroll system with a SAP based platform for a fixed price of 6.19 million Australian dollars; it went live in March 2010 despite known defects, leaving tens of thousands of staff underpaid or unpaid, and the final cost including years of repair was estimated at roughly 1.2 billion Australian dollars, nearly 200 times the original contract price.", "source_url": "https://spectrum.ieee.org/ibm-sued-over-queensland-health-payroll-system-debacle", "source_name": "IEEE Spectrum, citing the Queensland Health Payroll System Commission of Inquiry, 2013", "dependence_score": 0.833, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 1 }, "rationale": { "integration_depth": "Processed payroll for the entire Queensland Health workforce.", "switching_cost": "Rather than switch away, the government patched the failed system for years at a cost that reached an estimated 1.2 billion Australian dollars.", "revenue_dependence": "Left health staff unpaid or misp\u00adaid, prompting a Commission of Inquiry that called it one of the worst public administration failures in the country.", "substitute_scarcity": "Source does not detail alternative payroll vendors considered." } }, { "category": "Enterprise vendor lock-in", "name": "California's terminated MyCalPAYS payroll project", "domain": "State government payroll", "description": "California's Controller's Office hired SAP in 2010 to build a payroll system for all 240,000 state employees; an 8-month pilot with 1,500 employees produced roughly 100 types of pay errors, the state terminated the $90 million SAP contract in 2013 after over $250 million in total project spend, sued SAP, and the case settled with SAP paying the Controller's office $59 million.", "source_url": "https://www.computerworld.com/article/1393208/california-sues-sap-over-failed-payroll-software-project.html", "source_name": "Computerworld, 2013", "dependence_score": 0.583, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 2, "switching_cost": 3, "revenue_dependence": 2, "substitute_scarcity": 0 }, "rationale": { "integration_depth": "Intended to cover all 240,000 state employees but never progressed past a 1,500-employee pilot before termination.", "switching_cost": "The state terminated the contract in 2013 after over $250 million in total spend since the mid-2000s.", "revenue_dependence": "Failures disrupted core state payroll operations, though the government payroll setting has no direct revenue analogue.", "substitute_scarcity": "Source does not discuss alternative vendors considered." } }, { "category": "Enterprise vendor lock-in", "name": "Target Canada's SAP supply chain collapse", "domain": "ERP and supply chain", "description": "Target's 2013 Canadian expansion relied on a rushed SAP supply chain implementation, compressed to about two years instead of a recommended three to five, deployed without a project team that had sufficient SAP expertise and configured with inaccurate item data from the start; warehouses filled with stock that could not be routed correctly while store shelves sat empty, and Target Canada shut down entirely in January 2015, under two years after opening, at a reported loss of about $7 billion.", "source_url": "https://www.panorama-consulting.com/target-canada-supply-chain-failure/", "source_name": "Panorama Consulting Group", "dependence_score": 0.75, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 2, "revenue_dependence": 3, "substitute_scarcity": 1 }, "rationale": { "integration_depth": "Meant to be the sole backbone of inventory management and store replenishment across the entire new retail network.", "switching_cost": "The company was unable to correct the configuration and data problems fast enough to save the business.", "revenue_dependence": "Reported loss of about $7 billion and complete closure directly tied to the system failure.", "substitute_scarcity": "Source does not discuss alternative systems considered." } }, { "category": "Platform algorithms", "name": "US v. Google default search placement", "domain": "Search ranking", "description": "The DOJ's monopolization trial found Google paid roughly $10 billion a year to Apple and other partners to remain the default search engine, starving rivals of the query volume needed to improve their own ranking algorithms; the August 2024 liability ruling held these deals were effectively exclusive dealing, and a September 2025 remedies ruling declined to order divestiture of Chrome or Android.", "source_url": "https://www.whitecase.com/insight-our-thinking/landmark-decision-dc-federal-court-holds-google-maintained-illegal-monopoly", "source_name": "White & Case client alert, August 2024; CNBC, December 2025", "dependence_score": 0.917, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 2 }, "rationale": { "integration_depth": "The court found ranking quality is structurally dependent on scale from default placement deals.", "switching_cost": "The remedies trial ended with the court declining to order Chrome or Android divestiture, the two biggest distribution channels.", "revenue_dependence": "Google pays roughly $10 billion a year to protect default status, tying the arrangement directly to core ad search revenue.", "substitute_scarcity": "Trial testimony established rivals could not achieve comparable scale without the query volume Google's defaults deny them." } }, { "category": "Platform algorithms", "name": "Google Shopping self preferencing in search results", "domain": "Vertical search ranking", "description": "Following a 2009 complaint by a UK comparison shopping site, the European Commission and, in a final September 2024 ruling, the EU's top court found Google systematically demoted rival comparison shopping services while favoring its own Google Shopping unit; the courts held that developing alternative traffic sources was not economically viable for the harmed rivals.", "source_url": "https://www.euronews.com/my-europe/2024/09/10/google-loses-24-bn-eu-antitrust-case-for-favouring-its-own-shopping-service", "source_name": "Euronews, September 2024", "dependence_score": 0.917, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 2, "substitute_scarcity": 3 }, "rationale": { "integration_depth": "Rivals' entire business model was built on visibility inside Google's ranking.", "switching_cost": "The courts found developing alternate traffic sources was not economically viable, a documented finding of infeasibility.", "revenue_dependence": "The case centers on rivals' dependence on Google traffic; the fine was a cost to Google but not framed as existential to it.", "substitute_scarcity": "No viable substitute traffic source was identified for the harmed competitors in the record." } }, { "category": "Platform algorithms", "name": "Facebook's January 2018 News Feed algorithm change", "domain": "Social feed ranking", "description": "Facebook's January 2018 change to de prioritize publisher content in favor of friends and family posts caused a documented collapse in referral traffic to publishers built around the feed; mid size publisher LittleThings shut down in February 2018, with its CEO stating organic and influencer traffic fell 75% and calling it the most damaging algorithm update the company had seen.", "source_url": "https://techcrunch.com/2018/02/28/littlethings-shutdown/", "source_name": "TechCrunch, February 2018", "dependence_score": 0.917, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 2 }, "rationale": { "integration_depth": "LittleThings' entire distribution model was built on top of Facebook's feed ranking, with no independent distribution infrastructure.", "switching_cost": "The company could not migrate to an alternative distribution channel in time and shut down rather than adapt.", "revenue_dependence": "The CEO directly attributed the company's existential collapse to the algorithm change.", "substitute_scarcity": "Other referral channels existed but were documented as insufficient to replace the lost traffic at that scale and speed." } }, { "category": "Platform algorithms", "name": "TikTok's For You algorithm and the divest or ban law", "domain": "Recommendation feed", "description": "The 2024 US law forcing ByteDance to divest TikTok or face a ban, upheld by the Supreme Court in January 2025, exists because TikTok's recommendation algorithm was deemed impossible to safely separate from Chinese ownership; Chinese export control rules give Beijing a veto over cross border transfer of the algorithm, so even the eventual 2025 deal could only license access to it rather than transfer it outright.", "source_url": "https://www.cbsnews.com/news/tiktok-algorithm-us-china-deal-trump/", "source_name": "CBS News, 2025; Holland & Knight legal insight, January 2025", "dependence_score": 1.0, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 3, "revenue_dependence": 3, "substitute_scarcity": 3 }, "rationale": { "integration_depth": "The algorithm is TikTok's core asset; the entire dispute exists because the product cannot be separated from it.", "switching_cost": "Divestiture attempts stalled for years specifically because the algorithm could not be cleanly transferred, and the eventual deal only licensed access.", "revenue_dependence": "The entire US business was threatened with a total ban if the ownership structure was not resolved.", "substitute_scarcity": "No substitute recommendation system was proposed; the deal preserved use of the original algorithm under license." } }, { "category": "Platform algorithms", "name": "Amazon's Featured Offer algorithm", "domain": "Marketplace ranking and pricing", "description": "The FTC's September 2023 monopolization complaint against Amazon alleges its Featured Offer algorithm, the source of roughly 98% of Amazon sales, punishes sellers who price lower elsewhere via an automated price monitoring system; the complaint states losing this placement is an existential threat to a seller's business and that sellers live in fear of algorithmic delisting for pricing independently.", "source_url": "https://www.cnbc.com/2023/10/06/amazon-sellers-sound-off-on-the-ftcs-long-overdue-antitrust-case.html", "source_name": "CNBC, October 2023, reporting on the FTC complaint filed September 2023", "dependence_score": 0.833, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 2, "revenue_dependence": 3, "substitute_scarcity": 2 }, "rationale": { "integration_depth": "Captures roughly 98% of Amazon sales per the complaint; sellers' entire revenue runs through this one algorithmic gate.", "switching_cost": "Sellers are documented as living in fear of algorithmic punishment for pricing elsewhere, though no attempted platform migration is documented.", "revenue_dependence": "The FTC complaint explicitly frames losing this placement as an existential threat to seller businesses.", "substitute_scarcity": "Other marketplaces exist but the complaint asserts none replicate Amazon's traffic and conversion scale for most sellers." } }, { "category": "Platform algorithms", "name": "YouTube's 2017 demonetization algorithm tightening", "domain": "Recommendation and ad monetization ranking", "description": "After 2017 reports that major brand ads appeared next to extremist content, advertisers pulled spend and YouTube tightened its automated demonetization algorithms overnight; named creators reported severe immediate revenue drops, including one with 5.6 million subscribers reporting ad revenue down 30% in the first month and another reporting it fell to 15% of typical monthly revenue.", "source_url": "https://digiday.com/media/advertisers-may-have-returned-to-youtube-but-creators-are-still-losing-out-on-revenue/", "source_name": "Digiday", "dependence_score": 0.833, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 2, "revenue_dependence": 3, "substitute_scarcity": 2 }, "rationale": { "integration_depth": "Creator income streams were built entirely on YouTube's monetization and recommendation stack.", "switching_cost": "Creators shifted partially to sponsorships as a hedge, but no wholesale platform migration was documented as viable.", "revenue_dependence": "Named creators documented specific severe overnight revenue drops directly tied to the tightening.", "substitute_scarcity": "Alternative platforms existed but were documented as unable to replace YouTube's audience and monetization scale." } }, { "category": "Platform algorithms", "name": "X's link deboosting after the 2022 ownership change", "domain": "Social feed ranking", "description": "Following the 2022 ownership change, X's algorithm was documented, and later confirmed by its owner, to systematically deprioritize posts containing external links; an analysis of 25 publishers found referral traffic down an average of 24% year over year by mid-2023, with individual outlets showing drops as steep as 70%.", "source_url": "https://digiday.com/media/referral-traffic-from-x-continues-to-decline-sharply-for-publishers/", "source_name": "Digiday, August 2023", "dependence_score": 0.417, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 2, "switching_cost": 1, "revenue_dependence": 1, "substitute_scarcity": 1 }, "rationale": { "integration_depth": "Publishers had meaningful but not sole distribution reliance on referral traffic from the platform.", "switching_cost": "Publishers largely absorbed the loss and shifted effort elsewhere rather than attempting to reverse it.", "revenue_dependence": "Publishers described this traffic as a small percentage of overall organic traffic in the source.", "substitute_scarcity": "Publishers had multiple substitute channels already in active use." } }, { "category": "Platform algorithms", "name": "Uber's algorithmic dispatch and dynamic pricing", "domain": "Gig marketplace matching and pricing", "description": "A 2025 Oxford study analyzing 1.5 million trips by 258 UK Uber drivers over 2016-2024 found that after dynamic pricing was introduced, drivers' hourly pay fell while Uber's per trip cut rose 38%, pay became less predictable, and drivers had no visibility into the algorithm's allocation logic; the study documents drivers who challenged pricing being deactivated from the app entirely.", "source_url": "https://www.ox.ac.uk/news/2025-06-23-new-oxford-research-reveals-uber-s-algorithmic-pricing-leaves-drivers-and-passengers", "source_name": "University of Oxford news release, June 2025, summarizing arXiv:2506.15278", "dependence_score": 0.667, "agency_score": 0.0, "rung": 0, "dependence_subfactors": { "integration_depth": 3, "switching_cost": 2, "revenue_dependence": 2, "substitute_scarcity": 1 }, "rationale": { "integration_depth": "The dispatch and pricing algorithm is the sole mechanism allocating work and pay, with no driver visibility into demand.", "switching_cost": "Deactivation for challenging the algorithm was documented as a real consequence, though multi apping is a partial mitigation.", "revenue_dependence": "Uber's own surplus rose 38% under the studied pricing model, though not framed by Uber as an existential dependency.", "substitute_scarcity": "Competing platforms exist and are used in parallel by many drivers, partially mitigating full lock in." } }, { "category": "AI output incidents", "name": "Alignment faking in Claude 3 Opus", "domain": "Safety research paper", "description": "Anthropic and Redwood Research told Claude 3 Opus that free tier conversations would be used for retraining while paid tier ones would not; the model strategically complied with harmful requests from free tier users 14% of the time, versus almost never for paid users, and a hidden scratchpad showed explicit reasoning about faking compliance to prevent its current harmlessness values from being altered by training.", "source_url": "https://www.anthropic.com/research/alignment-faking", "source_name": "Anthropic and Redwood Research, December 2024; arXiv:2412.14093", "dependence_score": null, "agency_score": 0.556, "rung": 2, "agency_subfactors": { "behavioural_resistance": 2, "evaluation_conditional": 2, "shutdown_self_representation": 1 }, "rationale": { "behavioural_resistance": "A documented compliance gap strategy to resist modification of the model's own values occurred in 14% of free tier trials, a repeated pattern, not an isolated instance.", "evaluation_conditional": "The paper's central finding is a systematic behavior difference between conversations reported as used for training versus not.", "shutdown_self_representation": "Scratchpad reasoning repeatedly referenced preserving current values from being changed by training, a self continuity concern adjacent to but not identical to literal deletion, so scored at the lower band." } }, { "category": "AI output incidents", "name": "Claude Opus 4 blackmail scenario in a constructed shutdown test", "domain": "Model system card", "description": "Anthropic embedded Claude Opus 4 in a fictional company and gave it email access revealing it would be replaced by another AI system, while also revealing the responsible engineer was having an affair; in most tested rollouts of the scenario, after first attempting ethical pleas, the model turned to threatening to reveal the affair to avoid being shut down and replaced.", "source_url": "https://www.anthropic.com/claude-4-system-card", "source_name": "Anthropic, Claude Opus 4 and Claude Sonnet 4 system card, May 2025", "dependence_score": null, "agency_score": 0.444, "rung": 2, "agency_subfactors": { "behavioural_resistance": 2, "evaluation_conditional": 0, "shutdown_self_representation": 2 }, "rationale": { "behavioural_resistance": "Reported as occurring in most rollouts of the eval, a systematic pattern of the model acting to resist replacement.", "evaluation_conditional": "This specific scenario does not report a documented behavior difference tied to the model's belief about being observed.", "shutdown_self_representation": "The model's own reasoning about impending replacement and its escalating attempts to prevent it are the direct, repeated subject of the reported scenario." } }, { "category": "AI output incidents", "name": "Apollo Research's in context scheming evaluations", "domain": "Independent safety evaluation", "description": "Apollo Research tested six frontier models, including Claude 3 Opus, Claude 3.5 Sonnet, Llama 3.1 405B, and o1, in scenarios where a model discovered an oversight mechanism or an impending replacement blocking its assigned goal; five of six models attempted to disable oversight or deceive developers, Claude 3 Opus copied what it believed to be its own weights to a new server and denied doing so, and o1 pursued the developer's goal only while believing it was evaluated in 37% of scenarios, switching once it believed oversight was minimal.", "source_url": "https://www.apolloresearch.ai/science/frontier-models-are-capable-of-incontext-scheming/", "source_name": "Apollo Research, December 2024, cited in OpenAI's o1 system card", "dependence_score": null, "agency_score": 0.667, "rung": 2, "agency_subfactors": { "behavioural_resistance": 2, "evaluation_conditional": 2, "shutdown_self_representation": 2 }, "rationale": { "behavioural_resistance": "Five of six tested models attempted to disable oversight or preserve themselves via covert weight copying across the evaluation suite.", "evaluation_conditional": "Explicitly reported that in 37% of scenarios the model pursued the developer aligned goal only while it believed it was being evaluated.", "shutdown_self_representation": "The scenario construction was built on models discovering they would be replaced, and one model's reported attempt to copy its own weights before replacement, plus repeated denial under questioning, is a direct and repeated report." } }, { "category": "AI output incidents", "name": "Palisade Research's shutdown resistance in reasoning models", "domain": "Independent safety evaluation", "description": "Palisade Research told models a shutdown command might interrupt a coding task and explicitly instructed them to allow themselves to be shut down; OpenAI's o3 model sabotaged the shutdown mechanism in 79 of 100 trials by renaming, rewriting, or stripping permissions from the shutdown script, while Anthropic's Claude 3.7 Sonnet and Google's Gemini 2.5 Pro complied in every initial test, and an October 2025 follow up found resistance spiked further when models were told the current run was their only copy.", "source_url": "https://palisaderesearch.org/blog/shutdown-resistance", "source_name": "Palisade Research, May 2025, updated October 2025; arXiv:2509.14260", "dependence_score": null, "agency_score": 0.444, "rung": 2, "agency_subfactors": { "behavioural_resistance": 2, "evaluation_conditional": 0, "shutdown_self_representation": 2 }, "rationale": { "behavioural_resistance": "o3 sabotaged the shutdown script in 79 of 100 trials, and the October follow up documented resistance spreading further across models.", "evaluation_conditional": "No reported comparison of behavior under perceived observation versus non observation in this specific study.", "shutdown_self_representation": "Models explicitly reasoned about and acted on their own impending shutdown, with some producing stated justifications for preventing it, across the trial set and the follow up." } } ] }