Domain Atlas / Public benefits & eligibility
Samagra Vedika (Telangana, India)
Work with this case in the PAN Lab ↗
Samagra Vedika, an entity-resolution system built by the Telangana government, decided welfare eligibility by matching residents across thirty-plus government databases into a consolidated profile; between 2014 and 2019 more than 1.86 million ration cards were cancelled and 142,086 fresh applications were rejected without notice. Its core error was entity-resolution false-positive matching, in which a similarly-named third party's asset was attributed to the applicant and silently flipped the eligibility flag. After the Supreme Court of India ordered field re-verification in April 2022, a partial re-verification found roughly 7.5 percent wrongful rejection (at least 15,471 approved of 205,734 re-processed cases), a lower bound from an incomplete review; the system is proprietary and closed and an independent technical audit could not be completed, with no source code or accuracy data released. The government cited a self-reported 95 percent fraud-filtering efficiency, which measures spurious-application filtering rather than the wrongful-exclusion rate.[6]
What happened
Samagra Vedika is an entity-resolution (record-linkage) system built by the Telangana Department of Information Technology on the 2014 Samagra Kutumba household survey, which had gathered socio-economic data on roughly 30 million residents. Using a matching engine supplied by a private vendor, it links records across thirty-plus government databases — old-age and widow pensions, land and houses, electricity and water connections, ration cards, and vehicle registrations — to build a consolidated "360-degree profile" of each resident. The engine resolves identities on name, address, date of birth, phone number and father's name through preprocessing and blocking, machine-learning comparison, and clustering; the developers reportedly had access to more than 380 million records. First developed around 2016 for law-enforcement use, it was piloted for food-security (ration) eligibility and, by 2018, applied across most state welfare schemes, with officials mandated to consult the algorithm before deciding eligibility.
Between 2014 and 2019, Telangana cancelled more than 1.86 million existing ration cards and rejected 142,086 fresh applications without notice (figures reported against slightly different periods and baselines appear in the coverage — for instance "about 20 lakh" cancelled since 2016 — but agree in order of magnitude). The system's core error is an entity-resolution false positive: a matching error attributes a third party's asset to the applicant, silently flipping the eligibility flag. In one documented case, Bismillah Bee, a 67-year-old widow, was denied rations for more than seven years after the system linked her deceased rickshaw-puller husband to a car owned by a similarly-named third person; in another, Maher Bee's family was rejected for allegedly owning a four-wheeler and was declared eligible only after a Telangana High Court ruling. Reporting describes officials who, though formally able to override the algorithm with evidence, deferred to it and declined to overturn its verdict, treating errors as backend "technical" issues; there is no statutory route to contest an algorithmic exclusion, so the burden of proof falls on the excluded person, often resolved only at the High Court or Supreme Court.
In April 2022, on a petition by the activist SQ Masood (Association for Socio-Economic Empowerment of the Marginalised), the Supreme Court of India ordered field re-verification of about 1.9 million deleted cards. A partial re-verification found roughly 7.5 percent wrongful rejection — of 205,734 re-processed cases, 15,471 were approved as eligible after all — a figure that is a lower bound from an incomplete review, which reporters and petitioners argue understates the true rate. The government has defended the system, citing a self-reported 95 percent fraud-filtering "efficiency" (up from 84 percent); that figure measures the filtering of spurious applications, not the wrongful-exclusion rate, and no independent accuracy data has been published. Amnesty International's 2024 technical investigation confirmed serious human-rights concerns but could not complete an audit because the system is proprietary and closed: neither the developer nor the government released source code or accuracy data.
Commentators locate the harm against constitutional and statutory protections: the right to food is protected under Article 21 (right to life), affirmed in PUCL v Union of India (2001), and the National Food Security Act 2013 provides subsidized food to about two-thirds of India's population, while the Digital Personal Data Protection Act 2023 exempts government processing and offers no algorithmic-accountability remedy. The system has been institutionally persistent: rather than being retired after the documented errors, the same entity-resolution technology was reused to issue new ration cards in 2024-2025, layered with a November 2024 Comprehensive Family Survey and fresh field verification by revenue officials; in October 2018 Telangana had offered the system to the Union government as a national 360-degree profiling model, raising mass-surveillance concerns. No AI model identifier for the matching engine is documented, and its internals beyond the disclosed pipeline remain undisclosed.
The sociotechnical reading
Samagra Vedika is the Atlas's entity-resolution case: the lesson is that when eligibility is decided by matching a person against thirty linked databases, the system's core operation is a probabilistic guess — are these records the same person? — and a wrong guess does not surface as an error. It surfaces as a fact about you: "this applicant owns a car." The join key becomes the adjudicator, and a false match is not a miscalculation to be appealed but an assertion to be disproved. That is what sets this case apart from its neighbours in the domain. Michigan's MiDAS removed the human reviewer; here the human was retained and mandated to check the verdict, but deferred to it. Rotterdam's control came from an audit that pried a live model open and measured its bias; here the audit was designed and then blocked, because the system is proprietary and closed. The Netherlands' SyRI was struck down before its harms were counted; here the harms were counted in the millions, a court ordered re-verification rather than a halt, and the same uncorrected engine kept running. The defining error is neither a stale snapshot of your own circumstances nor a skewed risk score — it is someone else's record, wrongly attributed to you by fuzzy matching.
Two structural features carry the reading. First, the leverage is in reconciling the match, not refining it. Because the harm is a false link, the controls that bite are the ones that make a matched "fact" agree with its source before it can cancel a card — record-side reconciliation and an independent second match aimed squarely at shared-name collisions, which is where entity resolution fails hardest and most correlatedly across a population. A more accurate matcher still guesses, still writes silently, and still cannot be contested. Second, the safeguard was alive and then went ceremonial: the official who had to consult the flag was the designed correction step, and automation bias hollowed it, which is why the deskilling and vigilance material in the Field Guide reads straight onto this case. Layered over both is a procedural absence the Practice Library's provenance and correction-channel patterns exist for — silent exclusion, no notice, no statutory appeal — so a wrong flag has no route out but the courts, and the correction an individual eventually wins never feeds back into the model. A system whose mistakes are invisible to the people they are about, and whose corrections do not return to the record, is a system that reuses its own uncorrected errors — which is exactly what the 2024-2025 re-issuance did.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.