Three sources, one question, different answers

What this register can and cannot itemise For central schemes the Union Budget supplies an independent list of names, so a scheme missing from myScheme can be named, and 36 of them are named on this page. DBT Bharat publishes a per-state count and no state scheme list, so nothing central can itemise the shortfall of 2,477 schemes across the 28 states where DBT counts more.

The way to name them is to read each state’s own budget, which has to be done state by state: there is no common format and no guarantee a state publishes a readable one. Karnataka is done and is below. Andhra Pradesh is collected. Gujarat was attempted and its budget cannot be read by machine, which is a finding about Gujarat rather than a gap here. The survey of what each state publishes, including the failures, is in the repository.

This is a floor, and the errors in it are counted rather than estimated A budget book lists colleges, commissionerates and building heads beside schemes, and calling one of those a hidden scheme would be a false accusation. Every row is scored on signals published with the arithmetic.

Precision at the published bar is 97.3%, and that is a count rather than an estimate: every row scoring at or above the coverage threshold carries a hand label, so the errors in this table are 2 named ones. Validated against 324 hand labels in all. Recall is 32%, so this is a floor and never a total. Gruha Lakshmi, Karnataka’s largest welfare scheme, is missing from this table: 580 of the 969 rows carry no purpose line, and a real scheme with a plain name and no stated purpose cannot clear a high bar on the evidence the books print.

One row per scheme, not per departmental share, and a floor again Andhra Pradesh funds a scheme separately out of each social-category department, so NTR Bharosa Pension is six budget lines. They are added rather than listed six times, because the departments are distinct and the shares are of one provision. That is the opposite of the rule used across the six books, where the publications report overlapping slices and the largest is taken.

Precision at the published bar is 95.7%, and that is a count rather than an estimate: every row scoring at or above the coverage threshold carries a hand label, so the errors in this table are 4 named ones. Validated against 302 hand labels in all. Recall is 36%, so this is a floor and never a total. No Andhra Pradesh book prints a purpose line and 150 rows print no head of account either, so the classifier reads a name and nothing else. The largest thing it rejects is Annadata Sukhibhava at 6,600 crore, on a score of 2.

This table is the smallest and least complete on the page Kerala’s Annual Plan is a plan document, so most of its rows are institutions and establishment heads rather than schemes: an academy, a university, a directorate. Every row is scored on signals published with the arithmetic.

Precision at the published bar is 97.4%, and that is a count rather than an estimate: every row scoring at or above the coverage threshold carries a hand label, so the errors in this table are 1 named ones. Validated against 584 hand labels in all. Recall is 12%, so this is a floor and never a total. The National Old Age Pension, PMAY-Gramin, PMMVY and DDUGKY are all absent from this table, each one point below the bar. Kerala’s schemes are named Deendayal Antyodaya Yojana or PM KUSUM, and 28 of 50 sampled carry no English benefit or beneficiary word at all, so a classifier reading English for evidence of a benefit finds none in a name that is simply a name.

Tamil Nadu prints object heads, which is why this is the best measured state These are the full detailed estimates rather than a welfare-only cut, so five rows in six are establishment, works or accounting heads. What separates them is the object head: a row whose every object head is a benefit transfer is a scheme 89.5% of the time, and one whose every object head is an accounting head never is.

Precision at the published bar is 96.3%, and that is a count rather than an estimate: every row scoring at or above the coverage threshold carries a hand label, so the errors in this table are 12 named ones. Validated against 829 hand labels in all. Recall is 41%, so this is a floor and never a total. Magalir Urimai Thogai, the state’s largest cash transfer at ₹14,414 crore, is missing from this table for ₹9,803 crore of its value. Its two sub-plan heads carry a benefit object head alone and are here; its general head books Advertising and Publicity under the same sub-head as the transfer, which costs it exactly the margin. The Breakfast Scheme fails the same way. That is not a classifier failing to read, it is a state booking a benefit and its marketing in one place with no line saying which is which.

How these were separated from budget heads, and how often that is wrong Statement 4B mixes welfare schemes with infrastructure and accounting heads like “Road Works”, “Rolling Stock”, “Manufacturing Suspense”, none of which a citizen applies to. A classifier scores each line on independent signals: named in DBT Bharat’s list (+3), Centrally Sponsored (+2), benefit words in the name (+2), has an outcome framework (+1), asset or accounting words (−3), capital-heavy demand (−2). Every line’s arithmetic is published so the verdict can be rechecked.

Validated against myScheme membership as ground truth: at the F1-optimal threshold of 2 precision is 36%; this table runs at the stricter threshold of 4, where precision is 60%. Naming a scheme as missing is an accusation, so it runs at the high-precision end. Recall is a floor and not a measurement: a line called a scheme that myScheme lacks may be the classifier being right and the portal being incomplete, which is the thing this page is about. Residual errors are visible above: “Space Technology” should not be here.