UP local elections v2.0

UP local election data now live in data/release/, with a generated catalog, column dictionary, row units, join keys, and SHA-256 manifest.

• 212,525 GP-head source records across 2005, 2010, 2015, and 2021; 373,096 candidates in 2021.
• 39 office-specific tables remain provisional source observations (assignment_usable=false). Their row counts are not distinct seat counts.
• Corrected 2015 winner-list semantics and retained both actual candidate names in the unresolved 2021 conflicting-winner contest.
• Analytical tables omit phone numbers and unrecovered headers. Original evidence is preserved separately.

Download up_v2_data.tar.gz, verify it against up_v2_data.sha256, and extract it at the repository root. Start with data/release/CATALOG.md and data/release/DICTIONARY.md.

Validation: 75 Python tests, four R checks, Linux Docker checks, GitHub CI, independent model review, and an offline rebuild reproducing all 54 table hashes.

Original inputs and historical evidence are preserved in the source-evidence release. All 77 evidence assets and the final data download were independently downloaded and hash-verified.

UP source evidence for the v2 release

Preserved inputs and recovery evidence for UP v2.

The archives preserve 1,703 saved input files, 473,103 recovery files, 377 legacy source files, and 713 superseded build snapshots. Member hashes and archive hashes are recorded in the attached inventories. All 77 release assets were downloaded and SHA-256 verified before publication preparation.

Inputs and recovery gzip archives, and the separate snapshot zstd archive, are split into ordered .chunk-* files. Concatenate the matching chunks before extraction. Follow data/catalogs/SOURCES.md in the matching UP v2 checkout. The legacy tar is retained for audit and is not an active input.

Historical source receipts retain their original paths. Archived office and LGD registries are historical snapshots; the tracked v2 registries contain the current paths and pins. Recovery records and superseded builds are source evidence, not validated reservation assignments.

Shared election panels and Weaver preparations

Shared UP election histories for quota and quota_raj, plus reusable Weaver preparations.

• Independent adjacent-election panels: 42,622 links for 2005–2010, 39,551 for 2010–2015, and 46,514 for 2015–2021.
• Four-election panel: 29,734 histories composed through common source IDs.
• Hindi-preserving normalization, numeric-name vetoes and reciprocal matching; ties and competing script identities remain unresolved.
• Separate Weaver vintages preserve all source wave values and make their different longitudinal coverage explicit.

Independent review rebuilt all 14 products byte-for-byte. Local R, Python, lint, source-export and checksum checks passed; all three CI jobs passed on both the PR and publication commit e5a2ebd.

The full source repositories are needed only to rebuild these products; consumers pin immutable commits and SHA-256 hashes. Matching establishes reproducible candidate identities, not ground-truth validation of every fuzzy link.

UP discovery evidence, 2026-09-10

UP discovery evidence, 2026-09-10

The compact catalogs in data/catalogs/up_discovery_2026-09-10 describe the search frame and acquired sources. The full discovery tree is preserved in up_discovery_2026-09-10.tar.gz, not duplicated in Git.

The archive contains 312 files: 296,922,147 original bytes compressed to 119,296,382 bytes, a 59.8% reduction. Every member was decompressed and checked against its original SHA-256. Filenames and contents are unchanged; archive timestamps and ownership are normalized.

Download the archive from the repository's source-evidence-2026-09-10 release into data/source_archives, then restore it from the repository root:

gh release download source-evidence-2026-09-10 --repo in-rolls/local_elections_up --pattern up_discovery_2026-09-10.tar.gz --dir data/source_archives
tar -xzf data/source_archives/up_discovery_2026-09-10.tar.gz

The manifest records the archive hash, original member paths, byte counts, and file hashes. Acquisition receipts and raw responses remain unchanged inside the archive. The directory layout is restored exactly, so receipts still resolve to their evidence.

This is a source-evidence snapshot, not a new analysis-ready release. The current-host archive search covers all 75 districts and the discovery frame lists 826 blocks, but historical aliases and document contents are not exhaustively reviewed. The 2010 Sitapur extraction remains local staging because column-boundary errors were detected; those parser outputs are not part of this snapshot. Haryana OCR is unrelated and excluded.