Schema reference — S3 layout

All structured data is in Hetzner Object Storage at s3://tnelection2026/. Total: ~2.2 GB across 9 dataset folders.

results/

ECI 2026 candidate-level + round-wise. Already in S3 for all 234 ACs.

results/curated/year=2026/ac={N}/candidate_totals.parquet

Column	Type	Source
sn	int	ECI table row number
candidate	str	candidate full name
party	str	party name (ECI spelling)
evm_votes	int	EVM tally
postal_votes	int	postal/ETPBS tally
total_votes	int	evm + postal
pct_votes	float	percent of valid votes in AC
state	str	"TN"
ac_no	int	1..234
ac_name	str	e.g. "KOLATHUR"
year	int	2026
run_date	str	fetch date

results/curated/year=2026/ac={N}/round_votes.parquet

Column	Type
round_no	int
candidate	str
party	str
brought_forward	int (running total at start of round)
current_round	int (votes this round)
total	int (running total at end of round)
state, ac_no, ac_name, year, run_date	metadata

historical/

2021 + 2016 + 2011 TN AE results. Used as the swing baseline.

historical/curated/year=2021/kracekumar_detailed.parquet

4,232 rows — one per (AC, candidate) for the 2021 TN AE.

Column	Type
sn	int (O.S.N. in source)
Candidate	str
Party	str
evm_votes, postal_votes, total_votes	int
pct_votes	float
ac_name	str
ac_no	int
district_name	str
constituency_type	str ("General" / "SC" / "ST")
position	int (rank, 1 = winner)
Gender	str ("M" / "F")

demographics/

Census 2011 religion mix at district level.

demographics/curated/district_religion.parquet

32 rows — one per TN district as of Census 2011.

Column	Type
district_code	str (Census 6-digit MDDS code)
district_name	str
Total_persons	int
Hindu_persons, Hindu_pct	int, float
Muslim_persons, Muslim_pct	int, float
Christian_persons, Christian_pct	int, float
Sikh_*, Buddhist_*, Jain_*, Other_*, NotStated_*	int, float

demographics/curated/ac_demographics.parquet

234 rows — each AC joined to its district's religion mix.

Column	Source
ac_no, ac_name, district_name, constituency_type	2021 results
district_code, Hindu_pct, Muslim_pct, Christian_pct, Sikh_pct, ...	Census 2011
dist_norm	normalised district name (used for the join)

Gap

31 of 234 ACs have null religion fields. Those are in districts created after Census 2011 (Tirupathur, Mayiladuthurai, Ranipet, Chengalpattu, Kallakurichi, Tenkasi).

insights/

insights/curated/swing_2021_to_2026.parquet

3,393 rows — one per (AC × party) for parties contesting in 2021 or 2026.

Column	Type
ac_no	int
ac_name, district_name, constituency_type	from 2021
party_norm	str — normalised to
pct_2021	float (null if party didn't contest in 2021, e.g. TVK)
pct_2026	float (null if party didn't contest in 2026)
swing_pct	float = pct_2026 − pct_2021 (with nulls treated as 0)
votes_2021, votes_2026, total_2021, total_2026	int

Other folders

Folder	Purpose	Status
form20/	Booth-level PDFs + (partial) booth × candidate parquets	partial — 12 ACs of 234
voters/	Voter-roll PDFs + OCR'd voter lists	partial — 10 booths in Kolathur
candidates/	MyNeta candidate data	deprioritized
caste/	Reservation + Wikipedia prose per AC	234 ACs (sparse)
religion/	State-baseline religion per AC	234 ACs (state-level only)
alliance/	Party→alliance map per year	4 years (2026, 2021, 2016, 2011)
geo/	DataMeet shapefile + parquet WKT	234 polygons

Quick DuckDB query

INSTALL httpfs; LOAD httpfs;
SET s3_endpoint='hel1.your-objectstorage.com';
SET s3_access_key_id='...';
SET s3_secret_access_key='...';

-- Most-swung TVK ACs by district
SELECT ac_no, ac_name, district_name, swing_pct
FROM read_parquet('s3://tnelection2026/insights/curated/swing_2021_to_2026.parquet')
WHERE party_norm = 'TVK'
ORDER BY swing_pct DESC
LIMIT 10;