Methodology

Temporal coverage — which years each figure type covers

Each chart's per-chart picker exposes only those (year, figure type) combinations that the upstream data actually carries. The table below summarises what's available for the donut, treemap, and peers chart from RBI's e-STATES Appendix-1 (revenue) and Appendix-2 (expenditure):

• Actuals: FY 1990-91 → FY 2023-24 (34 years). FY 2024-25 + 2025-26 still in the audit lag — the CAG has not yet published actuals for those years.
• Revised Estimate (RE): FY 2001-02 → FY 2024-25 (24 years). RE was not separately reported in the RBI workbook before 2001-02.
• Budget Estimate (BE): FY 2002-03 → FY 2025-26 (24 years). BE was not separately reported before 2002-03.

Telangana exclusion. Telangana was created on 2 June 2014. The peer-states chart excludes TS for any fiscal year before FY 2014-15 — pre-2014 TS rows in the underlying RBI workbook are residuals or back-projections, not directly comparable to today's TS.

Per-capita unavailable for FY < 2001-02. The per-capita toggle uses the projection_series census2001_2011_linear@v2 (described below). Years before FY 2001-02 have no Census 2001 anchor; the toggle silently does not apply, and an inline note is shown on the chart.

Population emitter v2. The previous projection_series census2011_linear@v1 used Census 2011 as a single anchor and projected linearly forward and backward. v2 adds the Census 2001 anchor so the FY 2001-2010 range uses interpolation rather than backward extrapolation. AP/TS pre-2014 totals are back-derived from the undivided-AP 2001 enumeration via the 2011 split ratio (≈ 0.4161), tracked as a documented approximation in RISKS.md.

Budget Estimates, Revised Estimates, and Actuals

Indian state budgets publish three figures for each year, each meaning something different.

• Budget Estimate (BE). The plan, presented to the Assembly before the year starts. This is what the government intends to spend or collect. It is rendered with a diagonal hatch on every chart and suffixed (planned) on standalone numbers.
• Revised Estimate (RE). A mid-year update — the plan adjusted as the year unfolds. Solid fill, no suffix.
• Actuals. What was actually spent or collected, audited and published one to two years after the year closes. Solid fill with a slightly bolder stroke.

Why this matters. The most likely citizen misreading of any government chart is "this is what was spent" when it actually shows the plan. The hatch pattern is the visual cue. If a chart is hatched, the government has not yet spent that money — it has only said it intends to.

Per-capita figures: how the projection works

Per-capita figures use the Census 2011 final population totals as a base, projected forward using a simple linear extrapolation through the ten years of the data series. This is the census2001_2011_linear@v2 method recorded on every per-capita figure.

Method. We start with the Census 2011 state populations, fit a linear trend to the inter-census growth rate, and extrapolate forward. There is no smoothing for migration, no district-level redistribution. The method is honest about its bluntness.

Uncertainty. Per-capita figures show a ±3% sensitivity band underneath the headline number. The band is not a confidence interval in the statistical sense; it is a deliberately conservative envelope that reflects two known sources of error: linear extrapolation drift over a 14-year horizon, and inter-state migration that the Census-base method cannot see.

Andhra Pradesh / Telangana caveat. The 2011 Census was taken before the 2014 bifurcation. The five-state series (TN, KA, KL, AP, TS) uses post-bifurcation populations from the Census 2011 redistricting tables; pre-2014 totals would not be directly comparable to today's AP and Telangana boundaries.

FY 2023 projected populations (the year used in the home-page peer comparison):

AP — Andhra Pradesh

5,61,67,496

KA — Karnataka

7,30,08,948

KL — Kerala

3,54,58,031

TN — Tamil Nadu

8,62,24,466

TS — Telangana

4,00,26,032

FRBM ceiling values

The TN Fiscal Responsibility and Budget Management Act, 2003, sets a revenue-deficit ceiling of 0% of GSDP — i.e., the state must run a revenue surplus. The fiscal-deficit and outstanding-debt ceilings are set by the 15th Finance Commission's TN-specific glide path, with COVID-era relaxations applied to FY 2020-21.

FY 2020-21 (COVID year): the value 4.5% of GSDP shown below is a documented conservative midpoint between the 15th FC baseline (4.0%) and the conditional power-sector-reform allowance (5.0%). RISKS.md tracks this as an open item until a finance-domain reviewer signs off.

Data quality rules

Every build runs the following six checks against the curated parquet. Any failure exits non-zero and blocks promotion (IMPLEMENTATION.md §3.5).

1. Source completeness — no row without a source document ID. If a single number cannot be traced, the build fails.
2. Source registry integrity — every source ID resolves to a real document on disk, and the SHA-256 of that document matches what we recorded at fetch time.
3. Sign sanity — receipts and expenditure are non-negative; deficits are typed as totals with a documented sign convention.
4. Sector coverage — every Budget-Estimate or Revised-Estimate expenditure row has a sector classification. No "miscellaneous" bucket on charts.
5. Parser provenance — every row records which parser version produced it, when, and has a unique fact ID.
6. Population / FRBM join coverage — every state-year used on the frontend has a matching row in the population and FRBM tables. No silent NaN denominators.

Contested classifications

Some expenditure heads admit more than one defensible sector classification. Where reasonable people disagree, we render the alternative classifications transparently rather than hiding the choice. Sector classification is judgment-laden, especially when an accounting flow (state-to-utility transfer, residual social-services bucket) sits between policy intent and where the rupees land. For each head where the answer is genuinely contested, we document both candidate sectors and the case for each — and we mark the head as under independent review until a paid TN-fiscal-policy reviewer signs off.

When reviewer feedback returns, mappings may bump to v2. The audit trail of changes lives in `data/mappings/CHANGELOG.md`.

W0 feasibility evidence

Before any cloud spending, we ran a feasibility gate documented in notebooks/00-feasibility.md. The decision verdict, recorded on 2026-05-13, was:

MVP path: RBI e-STATES Excel only. PDF parsing deferred to post-MVP.

asia-south1 mandate: weakened. Region choice to be revisited at W1 once a personal-India fetch run is scheduled.

LlamaParse account: not created. Will not be created unless a personal-India PDF fetch returns a real PDF and pdfplumber accuracy on it is < 95%.