Methods: SIR publication-snapshot aggregates
Version: 2026-10-16-pilot-v1

Scope and sources
This dataset transcribes statewide aggregate tables from the Bihar Chief Electoral Officer's 30 September 2025 release (page 3) and the West Bengal Chief Electoral Officer's 28 February 2026 release (pages 1–2). It describes dated publication snapshots, not the current electoral roll. Official URLs, publication dates, retrieval timestamps and SHA-256 hashes are recorded in dataset-metadata.json and each aggregate row links to its source page.

Acquisition pilot
A purposive sample of 17 publicly accessible official PDFs, comprising 26 pages, covered final-publication releases, Bengal bulletins and an officially redacted Bengali sample form. This sample was chosen to test native text, image scans and legacy fonts; it is not statistically representative of electoral-roll pages. Downloads were sequential with a 1.5-second delay, timeout and size limits. Successful files were reused only after hash verification. All 17 files were reused in the resume check. No CAPTCHA bypass, paid API or cloud infrastructure was used.

Extraction and verification
Poppler pdftotext 26.03.0 extracted native text first. Nine pages had no extractable text. The three-page Hindi release had legacy-font text that could not be treated as readable Hindi. Installed Tesseract 5.5.3 models were benchmarked on four purposive pages at 200 DPI using hin+eng, ben+eng or eng. Reviewers must not treat non-empty text as proof of usable extraction.

The published aggregate values were manually compared with rendered source tables. Derived values were calculated from the reported counts. Five of six reconciliation checks passed; the sixth is the documented Bihar discrepancy below. Verification means faithful transcription of official releases, not independent confirmation of every underlying electoral decision.

OCR limitations
The Hindi page's nine checked total-column cells yielded only five exact matches. All six checked English table total tokens were present in the OCR output. These small spot checks do not establish general page accuracy. Bengali speed was measured on an official redacted sample form; Bengali word or character accuracy was not scored. OCR alone is not a publication gate. The published figures were visually verified rather than accepted from OCR.

Stage distinctions and discrepancies
Draft omissions, post-draft deletions, additions, pending adjudication and net change are distinct metrics. Do not sum across stages, label additions as restorations, or treat a signed net reduction as gross deletions. Restoration counts are explicitly null because these releases do not report them separately; show “Not separately reported”, never zero.

Bihar reports baseline 78,969,844 and draft omissions 6,564,075, whose difference is 72,405,769. The printed draft electorate is 72,405,756: a 13-elector discrepancy. Both source values are retained; no correction or causal explanation is inferred. The draft-to-final bridge reconciles.

Bengal's 28 February 2026 table includes 6,006,675 pending adjudication cases, separate from Form 7 deletions. Subsequent supplementary lists are outside this dataset. Its signed baseline-to-publication change therefore must not be described as final permanent deletions or as a current October 2026 total.

Publication and reproducibility
Each row carries an individual publication decision, numeric eligibility and mandatory caveats. Publish only reviewed aggregate rows, methods and public metadata. Keep sources/, text/ and renders/ local; no personal voter records are included in the public dataset. Do not imply an open licence for source PDFs or claim search indexing.

Reproduce locally: run pilot.py, ocr_benchmark.py, build_aggregates.py, build_report.py, then finalize_release.py in the owner folder. Aggregate construction is reviewed transcription rather than unattended table inference. Finalize release metadata after any dataset change because it records the exact dataset hash. Benchmark and scale estimates are available in the local handoff; estimates apply only to similar release documents and exclude human review costs.