Three sources, one question, different answers

• Assam, Himachal Pradesh and Madhya Pradesh serve no scheme list at all. Assam publishes none for this cycle. Himachal Pradesh publishes thirty documents and serves none of them.
• Bihar publishes names as pictures rather than as text. Bihar converts each page to vector curves before it publishes the book.
• Gujarat and Rajasthan publish names a machine cannot find the end of. Gujarat runs the name, the code and the target into one column. Rajasthan breaks names across lines mid-word.

Only one of the six is a layout problem. Each of the others comes from a decision. One state sets its budget in a legacy font. One converts it to curves before it publishes it. One lists an English edition and serves a 404 for it. One publishes nothing for the current cycle. Only the state can reverse each decision.

That last clause is the one that does the work and the one this register got wrong. We labelled seven states against this wording. We labelled eight later states against a looser sentence that admitted “a service”. That sentence lets in a helpline, a creche, an anganwadi’s running costs and an agricultural extension wing. We read sixty-nine published rows again against the wording above, and fifty-six changed side. Until then, the headline number added two different definitions together and did not say so.

Listing a state and accusing a portal are different claims, and all sixteen states now make both. Each state has a classifier, tested against that state’s own hand labels. Every row at the state’s publishing bar has a hand label. So the precision behind an absence claim is a count, and each error has a name. To call a directorate a hidden scheme is a false accusation, and these states clear that bar. The table above links each state’s full budget book. The classifier decides which lines this register calls schemes. The other lines of the book stay on the state’s page.

Uttar Pradesh is here without clearing the English test, and it is the only one. The largest state in India publishes its budget in Hindi and in nothing else. For that reason it was a refusal, until we checked the reason. The check went the other way: myScheme does not list Uttar Pradesh’s schemes in English either. It lists them in romanised Hindi: Kanya Sumangala Yojana, Berojgari Bhatta Yojna. So the two lists differ only in SCRIPT. A change of script is mechanical. A translation would be a claim about meaning, and this register does not make one. This page shows its 5,831 names as the state wrote them, each with a romanisation labelled as derived.

The last test is the strictest, and it separates the states that a machine can read. The book prints its own totals, and our reading must agree with all of them. Karnataka and Andhra Pradesh fail because their books print no totals, so nothing can check our reading against the source. Delhi and Uttarakhand fail because their printed totals do not close. These are different failures. The table cannot show the difference, but the sentence beside each state does.

To name them, this register reads each state’s own budget, one state at a time. There is no common format, and a state may not publish a readable budget at all. /states shows the 16 states a machine can read and the states it cannot.

We tested the classifier against myScheme membership. At the F1-optimal threshold of 2, precision is 31%. This table uses the stricter threshold of 4, where precision is 56%. To name a scheme as missing is an accusation, so the table uses the high-precision end. The recall figure is a floor. When the classifier calls a line a scheme and myScheme lacks it, the classifier can be right and the portal incomplete. That is the subject of this page. Some errors remain, and you can see them above: “Space Technology” should not be here.