Sources and method
The atlas attempts to answer the following questions: what was promised, where, when and how much? We've read Tamil Nadu's budget speeches (2011–12 to 2025–26, English and Tamil), and extracted every spatially-specific announcement, threaded them into projects, and put them on the map with their receipts.
1Sources
Budget speeches from tnbudget.tn.gov.in and financedept.tn.gov.in — 36 documents across 15 budget years, including the two budgets each election year brings. Every document is listed on the data page.
District boundaries come from the TN GIS portal, the state's own geospatial layer, carrying LGD codes and official areas. Taluk and assembly-constituency geometry still comes from the datameet community shapefiles; place names from GeoNames (CC-BY); state finances from the RBI Handbook of Statistics on Indian States, shown on the factsheet.
Two defects in the TN GIS district layer are corrected rather than inherited: every geometry fails a validity check and is repaired on ingest, and eight of the 38 Tamil-name fields are misassigned (Vellore reads "Tenkasi", Kanniyakumari reads "Coimbatore"), so district names in Tamil come from our own table.
2Method
Extraction is LLM-assisted, validation is deterministic: every record must carry a verbatim quote that fuzzy-matches its source at ≥0.9, amounts are re-parsed by code (canonical unit: ₹ lakh, as printed alongside), and records failing any audit go to a review queue — not onto the map. Entity resolution (threading announcements into projects) combines deterministic name-matching with LLM adjudication; every merge is reversible and overrides are versioned.
A 5% random quote audit passes at 100%; the full pipeline, prompts, and every intermediate file are in the project repository.
3Precision tiers
Every location claim carries an explicit precision tier and confidence. site / settlement render as points; taluk / district / constituency / region render as shaded polygons; statewide schemes are labelled so. A district-level promise is never dropped as a pin on the district headquarters. District boundaries are year-aware: announcements from before the 2019–20 reorganisation attach to the 32-district map of their time.
Corridors — "between Avadi and Sriperumbudur", "the Chennai–Kanyakumari corridor" — are shown as shading across every district the route runs through, not as a line. We know the two endpoints the speech names; we do not know the alignment, and drawing one would invent geometry nobody surveyed.
4Caveats
Amounts are nominal rupees of their year — comparing 2011 crore to 2025 crore without deflation misleads, and a real-terms toggle is planned. Announced ≠ spent: this is the promise layer, not an expenditure audit. Around 14% of place-carrying announcements resolve below district precision (rivers, roads, generic phrases) and are queued rather than guessed. Extraction confidence below 0.7 keeps a record out of the map entirely.
5Licence & corrections
Source documents are Government of Tamil Nadu publications. The derived dataset is CC-BY 4.0; cite as "TN Budget Atlas, Dravidian Urbanism". Corrections welcome — every item has a stable ID; quote it.